Full text for verification
AI Governance
Canonical record: https://ssrn.com/abstract=4796714
40 protected claims are extracted from this work.
Source extraction SHA-256: 2698502c460bfa0d07f82a169e086d81f9c60c9b9bf919237aac9a54dcb54435
Version 9.0 - May 2024 # **AI Governance** # Wulf Kaal, Ph.D.<sup>1</sup> # **Abstract** In the rapidly evolving landscape of artificial intelligence (AI), governance frameworks are increasingly pivotal. As AI technologies become more complex and integral to various sectors, the mechanisms to oversee and regulate these systems must evolve correspondingly. Traditional governance approaches often rely on static, predefined rules that may not adapt quickly enough to the pace of AI development or the nuanced challenges it presents. These conventional methods, largely reactive or fixed to ex-post solutions, are proving insufficient for the dynamic nature of AI technologies. The proposed AI governance system integrates decentralized web3 community governance and federated communication platforms, forming a sophisticated framework for dynamic, anticipatory, and participatory oversight of AI development. Key components include a federated forum platform structured as a Weighted Directed Acyclic Graph (WDAG), and specialized smart contracts for managing tasks and validation. This setup not only facilitates real-time consensus-building and decision-making via web3 community governance but also supports a scalable, transparent communication network. Validation Pools and Reputation tokens within this framework play crucial roles in maintaining an updated and responsive governance system, reflecting the collective decisions and ethical standards of the community. This system's effectiveness is demonstrated through applications like medical diagnosis AI and autonomous driving AI, where each development stage is captured as vertices in the WDAG, documenting key compliance and operational metrics. Directed edges in this graph link these stages to relevant legal and ethical standards, with assigned weights emphasizing areas critical for compliance and safety. The dynamic nature of WDAG allows for continuous updates and integration of new regulations or ethical guidelines, ensuring AI governance remains current with technological and societal shifts. This model thus ensures AI systems are not only technologically advanced but also ethically aligned and legally compliant, effectively balancing innovation with responsible governance. **_Key Words_** _:_ Artificial Intelligence, AI Models, Governance, Decentralized Autonomous Organization, WDAG, Hybrid Models, Data Governance, Token Models, Cryptocurrencies, Feedback Effects, Emerging Technology, Blockchain, Distributed Ledger Technology **_JEL Categories_** _:_ K20, K23, K32, L43, L5, O31, O32 > 1 Professor of Law. The author is grateful for excellent research assistance by Jack Palmer and research librarian Adam Bent. Version 9.0 - May 2024 # **Table of Contents** |Introduction.......................................................................................................................................................... 3<br>Origin of AI........................................................................................................................................................... 5| |---| |AI Models......................................................................................................................................................... 9| |Federated Model....................................................................................................................................... 11| |Centralized Model..................................................................................................................................... 13| |AI Governance................................................................................................................................................... 15| |Shortcomings in Existing AI Governance.................................................................................................... 17| |Proposed Solutions....................................................................................................................................... 19| |Decentralizing AI Governance........................................................................................................................... 23| |Blockchain & DLT Technology................................................................................................................. 25| |Recentralization......................................................................................................................................... 27| |Challenges ................................................................................................................................................... 28| |Adapting Decentralization of AI Governance to AI Models ............................................................................ 32| |Proposed System.............................................................................................................................................. 35| |Foundations................................................................................................................................................... 35| |Governance Mechanism............................................................................................................................... 37| |Precedent and Citation system.................................................................................................................... 37| |Dynamic Governance................................................................................................................................... 39| |Exponential Evolution of AI and Governance Needs............................................................................. 40| |Dynamic Real-Time Governance............................................................................................................. 40| |Key Components of WDAG-based AI Governance............................................................................ 41| |Benefits of WDAG-based AI Governance........................................................................................... 42| |Uploading AI Models as Posts in the Precedent Credit System........................................................ 42| |AI Model Integration...................................................................................................................................... 42| |Example: WDAG Governance for Medical Diagnosis AI........................................................................ 44| |Example: WDAG Governance for AI Model in Autonomous Vehicles.................................................. 44| |Avoiding WEB2 Inadvertent AI Learning Mistakes through WDAG AI Learning...................................... 46| |Model Comparison........................................................................................................................................ 47| |Model 1: Decentralized Community Governance Approach.................................................................. 48| |Model 2: Decentralized Data Validation Layer Approach...................................................................... 49| |Hybrid Model............................................................................................................................................. 49| |Model 3: Ex-Ante Community AI Governance........................................................................................ 51| |Conclusion......................................................................................................................................................... 54| Version 9.0 - May 2024 # Introduction In the rapidly evolving landscape of artificial intelligence (AI), governance frameworks are increasingly pivotal. As AI technologies become more complex and integral to various sectors, the mechanisms to oversee and regulate these systems must evolve correspondingly. Traditional governance approaches often rely on static, predefined rules that may not adapt quickly enough to the pace of AI development or the nuanced challenges it presents. These conventional methods, largely reactive or fixed to ex-post solutions, are proving insufficient for the dynamic nature of AI technologies. Web3, characterized by its decentralized architecture and reliance on blockchain technologies, introduces a fundamentally different approach to AI governance. This approach is not just a shift in technology but a paradigm shift in how governance can be implemented—through decentralized autonomous governance systems that utilize dynamic feedback mechanisms.<sup>2</sup> Such systems are crucial for addressing the dual needs of governing evolving AI models ex-ante and managing existing solutions ex-post. No legacy systems exist at the time of publication that could provide such dynamic governance toolsets. The traditional ex-post governance methods, where regulations are applied after AI systems are developed and deployed, or were pretrained LLMs models were trained on existing proprietary datasets, often fall short in preemptively addressing risks and biases. This approach can lead to gaps in oversight during critical early stages of AI development, where foundational attributes of AI systems are established. In contrast, an ex-ante governance approach, advocated within Web3 frameworks as presented in this paper, involves setting community coordinated regulatory measures and oversight mechanisms during the development phase of AI technologies. This proactive stance allows for real-time adjustments and refinements based on ongoing feedback from AI operations and interactions within the ecosystem. Web3 systems, with their inherent capabilities for real-time data processing, decentralized decision-making, and transparent operations, provide a robust infrastructure for implementing AI governance systems ex-ante. These systems facilitate a dynamic feedback loop where AI behaviors and outcomes are continually monitored and influenced by decentralized consensus, rather than being solely dictated by centralized authorities or delayed regulatory responses. Technology has historically outpaced regulation.<sup>3</sup> The exponential trends in AI development will continue to exacerbate the mismatch between regulation and AI development. The author has > 2 _Festschrift zu Ehren von Christian Kirchner: Recht im ökonomischen Kontext_ XVI, 1387 (Wulf A. Kaal, Andreas Schwartze & Matthias Schmidt eds., 2014). > 3 Mark Fenwick, Wulf A. Kaal & Erik P.M. Vermeulen, _Regulation Tomorrow: What Happens When Technology Is Faster Than the Law?_ , 6 Am. U. Bus. L. Rev. 561 (2017), available at https://ssrn.com/abstract=3204119. Version 9.0 - May 2024 advocated ex-ante dynamic regulatory methods for over a decade.<sup>4</sup> The proposed model in this paper ensures that AI governance is more adaptive, responsive, and aligned with ethical standards and societal needs ex-ante and during AI model development, not expost as demanded by traditional regulatory methods. Moreover, the use of dynamic feedback mechanisms enables a more nuanced and effective management of AI technologies. By integrating feedback directly into the governance processes, stakeholders can iteratively improve and adjust AI models in response to new information, operational experiences, and changing environments. This ongoing process helps in mitigating risks and biases more effectively than static, ex-post regulatory frameworks. The shift towards Web3 and the utilization of Web3 tools with dynamic feedback effects represent a necessary advancement in AI governance frameworks. This new model not only addresses the inadequacies of previous systems but also enhances the capability to govern AI technologies in a manner that is as advanced and dynamic as the technologies themselves. This paper aims to explore these advancements in detail, illustrating how Web3 can transform AI governance from a reactive to a proactive discipline that is better equipped to handle the complexities of modern AI systems. The conceptual framework for AI governance within Web3 revolves around creating a system where artificial intelligence entities can operate autonomously yet responsibly within a decentralized digital ecosystem. In Web3, AI entities can perform transactions without human intervention. This is made possible through the use of cryptocurrencies and digital tokens. AI systems can have their wallets to send and receive tokens as payment for services, such as data processing or cognitive tasks. This capability is critical for fostering a self-sustaining ecosystem of AI services. Yet, the web3 technology enables human inputs to co-govern with AI in real-time. With the massive amounts of data that AI systems generate and process, ensuring the integrity and authenticity of this data is crucial. Data attestation mechanisms in Web3 offer a cryptographic means to verify data integrity, crucial for AI systems reliant on data inputs for decision-making and pattern recognition. This capability not only enables AI tracking to monitor uploaded content and address potential intellectual property concerns but also provides a governance framework to mitigate liability issues stemming from AI misuse. By ensuring accountability at the individual level, rather than burdening the entire community, this proposed model for AI governance within Web3 offers a promising avenue to address ethical dilemmas associated with AI applications. Web3-based self-sovereign identity (SSI) allows individuals—and potentially AI systems—to have control over their digital identities without relying on a central authority. In the context of AI governance, SSI ensures that AI entities have verifiable credentials > 4 Mark Fenwick, Wulf A. Kaal, Toshiyuki Kono & Erik P.M. Vermeulen _, Dynamic Regulation for Innovation_ , in 6 Persp. in L., Bus. & Innov. 1 (2016), University of St. Thomas (Minnesota) Legal Studies Research Paper No. 16-22, 30 pages, posted Aug. 31, 2016. Version 9.0 - May 2024 and can engage in interactions with confidence in the identity and reputation of other participants. Web3 can offer a set of standards and protocols that define how AI systems interact, share data, and settle transactions. These protocols act as the "rules of the road," setting the groundwork for interoperability and collaborative efforts among diverse AI systems. In a decentralized autonomous organization (DAO) governed by Web3 principles, AI systems can actively participate in the governance process, contributing to decisionmaking and protocol changes that impact the network. This approach ensures a fair distribution of power, preventing any single entity from exerting undue influence and fostering an environment where decisions benefit all participants. Despite concerns about AI involvement in governance, the potential benefits are significant, especially when considering specialized AI systems that offer expertise or ensure the integrity of organizational principles through formal logic representation. By leveraging AI capabilities, DAOs can enhance the purity of decision-making processes, minimizing biases and agenda-driven actions, thus advancing the overall effectiveness and transparency of governance within Web3 ecosystems. By integrating these features, Web3 aims to augment human intelligence rather than replace it. The goal is to create an environment where AI can process vast amounts of data, learn, and interact in ways that amplify our cognitive capabilities, leading to better decision-making and innovation. AI governance in Web3 is about creating a framework where AI systems are empowered to act autonomously but are also accountable to the larger network. This involves enabling secure transactions, ensuring data integrity, managing identities, and coordinating actions through shared protocols. By doing so, Web3 can foster a governance model that is adaptable, transparent, and responsive to the fast-paced changes characteristic of modern AI technologies. # Origin of AI AI represents a fusion of computational techniques engineered to execute tasks traditionally requiring human intelligence, such as learning, reasoning, problem-solving, perception, and language understanding. The foundational architecture of AI systems incorporates a variety of methodologies that enable these machines to analyze environments and undertake actions aimed at achieving specific objectives, thus closely mirroring but not yet matching human cognitive functions.<sup>5</sup> Central to AI's capability is machine learning (ML), a domain wherein algorithms analyze data, learn from it, and subsequently apply the acquired knowledge to make informed decisions. This process involves constructing mathematical models that facilitate predictions or decisions autonomously, without direct programming for specific tasks. A > 5 Philip Boucher, _Artificial Intelligence: How Does It Work, Why Does It Matter, and What Can We Do About It?_ , STOA | Panel for the Future of Science and Technology, European Parliamentary Research Service (June 2020), available at > http://www.europarl.europa.eu/thinktank/en/document.html?reference=EPRS_STU(2020)641547_EN. Version 9.0 - May 2024 more specialized branch of ML, known as deep learning, utilizes layered neural networks that emulate human decision-making processes. These networks, capable of processing vast datasets, learn to recognize complex patterns and are crucial for tasks like image and speech recognition, as well as natural language processing. Neural networks, essential components of deep learning, are organized into multiple layers: the input layer, one or more hidden layers, and an output layer. Each layer is composed of units or neurons that process incoming data from the previous layer before passing it to the next, allowing the system to learn increasingly complex features as the data progresses through the network. Other AI Methodologies include rule-based modeling, data mining, fuzzy logic, casebased reasoning, and text and visual analytics.<sup>6</sup> Rule-Based Modeling uses if-then rules for knowledge representation, enabling the system to make decisions based on predefined logical rules. Data Mining involves extracting patterns from large datasets, which is pivotal for discovering hidden patterns and unknown correlations. Fuzzy Logic allows for reasoning that is approximate rather than fixed and exact, handling the concept of partial truth — where values may range between completely true and completely false. Case-Based Reasoning solves new problems based on the solutions of similar past problems. Text and Visual Analytics involves extracting useful information from text and visual data to facilitate decision-making and insight generation. The synthesis of these diverse AI methodologies typically involves combinatorial search and formal mathematical models, which are integral in developing systems capable of automated reasoning. This functionality is critical for AI's ability to process information and derive logical conclusions from intricate datasets.<sup>7</sup> However, a significant challenge within AI, particularly evident in deep learning models, is their "black box" nature. The internal mechanisms by which these models make decisions are not always transparent, complicating efforts to understand and trace how inputs are transformed into outputs. This opacity can obstruct the debugging process, obscure bias detection and mitigation, and hinder comprehension of AI decision-making.<sup>8</sup> For example, Nvidia’s self-driving cars learn from human behavior but might confuse the moon for a traffic light, and the DeepPatient project accurately predicted disease onset from medical records without providing explanations for its predictions, posing challenges for medical professionals trying to trust and interpret these AI systems.<sup>9</sup> This opacity not > 6 Iqbal H. Sarker, " _AI-Based Modeling: Techniques, Applications and Research Issues Towards Automation, Intelligent and Smart Systems_ ," in SN Computer Science, vol. 3, no. 158, 2022, pp. 1-20, https://doi.org/10.1007/s42979-022-01043-x > 7 Stuart Russell, _The History and Future of AI_ , 37 Oxford Rev. Econ. Pol'y 509 (Autumn 2021), https://doi.org/10.1093/oxrep/grab013. > 8 Melissa Heikkilä, _Nobody Knows How AI Works_ , MIT Tech. Rev. (Mar. 5, 2023), > https://www.technologyreview.com/2024/03/05/1089449/nobody-knows-how-ai-works/. > 9 Will Knight, _The Dark Secret at the Heart of AI_ , MIT Tech. Rev. (Apr. 11, 2017), https://www.technologyreview.com/2017/04/11/5113/the-dark-secret-at-the-heart-of-ai/?gad_source=1. Version 9.0 - May 2024 only makes it difficult to correct biases and errors but also potentially erodes trust, which is critical as AI becomes more embedded in everyday life. Each application benefits from AI’s ability to swiftly process large datasets, learn from inputs, and enhance decision accuracy over time. Nevertheless, the deployment of AI systems brings forth critical ethical and practical considerations, such as issues of trust, privacy, and control. These aspects require ongoing evaluation and management to ensure that AI technologies are utilized responsibly and continue to align with human values and societal norms.<sup>10</sup> AI systems continuously evolve by harnessing vast datasets and utilizing sophisticated learning algorithms to enhance their decision-making capabilities over time. This evolution is significantly supported by rapid advancements in hardware, which provide the necessary computational power to process and analyze complex datasets efficiently. Moreover, the development of innovative architectures for structuring AI systems is essential, as these frameworks facilitate the effective interconnection and scaling of data layers, model complexity, and application breadth. Such structural innovations are crucial in optimizing the flow and processing of information, enabling AI systems to handle more complex tasks, adapt to new environments, and integrate seamlessly across diverse applications. This dynamic interplay of data accumulation, algorithmic learning, hardware advancements, and system architecture refinements drives the continuous improvement and expanding capabilities of AI technologies.<sup>11</sup> The primary hardware advancement critical to AI advancement has been increased computing power and chip performance. As AI algorithms become more complex and data-intensive, the demand for computational resources continues to escalate. Web3 systems offer a solution by enabling decentralized sharing of AI computing power. Through blockchain technology and distributed networks, unused computing resources can be harnessed and allocated efficiently, allowing AI tasks to be processed faster and at a lower cost. By democratizing access to computational resources, Web3 systems facilitate collaborative efforts in AI research and development, accelerating innovation and driving forward the capabilities of AI.<sup>12</sup> The improvement in chip performance, particularly through GPUs and specialized AI processors, has been crucial in providing the high-speed, efficient computing required to train these models on large datasets. This combination of advanced AI models and robust computing infrastructure allows AI systems to process vast amounts of information > 10 Will Knight, _The Dark Secret at the Heart of_ AI, MIT Tech. Rev. (Apr. 11, 2017), > https://www.technologyreview.com/2017/04/11/5113/the-dark-secret-at-the-heart-of-ai/?gad_source=1. > 11 Ganesh D. Govindwar & Sheetal S. Dhande, _A Review on Federated Learning Approach in Artificial Intelligence_ , Paper presented at the 2022 6th Int'l Conf. on Computing, Communication, Control and Automation (ICCUBEA), IEEE, Aug. 26-27, 2022, > https://doi.org/10.1109/ICCUBEA54992.2022.10010798. > 12 CITE CHECK ALL: Ben Dickson, The Democratization of Artificial Intelligence (Year); Pedro Domingos, The Hardware of AI (Year); Ian Goodfellow, Yoshua Bengio, and Aaron Courville, Deep Learning (Year); OpenMined, Decentralized AI: The Key to Democratizing AI (Year); Elisa Bertino et al., Blockchain for AI: Review and Open Research Challenges (Year). Version 9.0 - May 2024 rapidly, making strides toward achieving and even surpassing human-level thought in specific tasks. These technological enhancements not only increase the speed and efficiency of AI systems but also expand their applicability across different fields, pushing the boundaries of what AI can achieve and transforming potential theoretical concepts into practical applications.<sup>13</sup> The dominance of big tech companies in AI development is significantly shaped by their comprehensive resources, which include access to vast datasets, state-of-the-art hardware, and extensive human capital. These advantages enable them to develop more advanced and sophisticated AI systems, providing big tech with a substantial competitive edge in the market. As a result, the solutions offered by these companies are not only more prevalent in the marketplace but also superior in technological capabilities. This disparity in resource availability and resultant AI proficiency creates a pronounced digital divide, where smaller companies and traditional sectors that lack similar access to largescale data and cutting-edge technology struggle to keep pace. Consequently, this divide exacerbates the gap in AI capabilities, potentially leading to a concentration of power and influence in the hands of a few large tech entities, while others lag significantly behind in leveraging AI for innovation and improvement.<sup>14</sup> Regulatory frameworks and ethical guidelines have become pivotal in shaping the evolution of AI, steering development towards a human-centric approach. These guidelines emphasize the creation of trustworthy AI systems that uphold privacy, ensure transparency, and strive for explainability. By prioritizing these principles, regulators aim to mitigate the risks associated with AI technologies, such as data misuse, biased decision-making, and opaque systems whose workings are not understandable to users or their developers. The push towards explainability is particularly crucial as it involves designing AI systems whose actions can be easily understood by humans, thus fostering greater trust and acceptance. This regulatory focus not only ensures that AI development aligns with ethical norms and societal values but also helps in maintaining public confidence in how these advanced technologies are integrated into daily life, making AI systems more accessible and safer for widespread use.<sup>15</sup> The evolution of AI is increasingly guided by institutional regulations and ethical guidelines that advocate for a human-centric approach and the development of trustworthy AI systems that prioritize privacy, data governance, and transparency. However, strict data privacy regulations such as the General Data Protection Regulation (GDPR) present challenges to this growth trajectory. GDPR imposes stringent conditions > 13 Jie Wen, Zhixia Zhang, Yang Lan, Zhihua Cui, Jianghui Cai, & Wensheng Zhang, _A Survey on Federated Learning: Challenges and Applications_ , 14 Int'l J. Mach. Learn. & Cybernetics 513, 513-535 (2023), https://doi.org/10.1007/s13042-022-01647-y > 14 Yannick Bammens & Paul Hünermund, _Using Federated Machine Learning to Overcome the AI Scale Disadvantage_ , 65 MIT Sloan Mgmt. Rev. 54 (Fall 2023) > 15 José Luis, Manuel Martínez-González, Miguel L. Bote-Lorenzo, Juan I. Asensio-Pérez, & Eduardo Gómez-Sánchez _, FED-XAI: Federated Learning of Explainable Artificial Intelligence Models_ , 3277 CEUR Workshop Proc. 1, 1-10 (2022), https://ceur-ws.org/Vol-3277/paper8.pdf. Version 9.0 - May 2024 on data sharing, which can limit the amount and variety of data AI systems use, potentially reducing their performance and exacerbating biases due to the restricted dataset. Furthermore, while the move towards more explainable, private, and transparent AI is commendable, these regulations can paradoxically consolidate power within large tech companies. These entities possess the vast data resources and sophisticated infrastructures needed to comply with such regulations and still develop effective AI solutions, unlike smaller companies or individual developers who might be sidelined by these legal constraints, thus altering the competitive landscape in favor of more established players.<sup>16</sup> The future of AI's evolution is intricately linked to the widespread adoption of smartphones and the move towards decentralized data processing and model training. With smartphone usage projected to rise from over 5 billion users in 2020 to an estimated 7.5 billion by 2025, these devices are poised to become an even more significant foundation for AI development. The ubiquity of smartphones, coupled with their advanced computing capabilities and constant connectivity, offers a vast resource for deploying AI applications directly into the hands of users globally. This scale of interconnectivity and integration into daily life makes smartphones a crucial platform for AI. Furthermore, the evolution of AI technologies is increasingly embracing approaches like embedded and federated machine learning, which prioritize decentralization. These methods shift away from traditional cloud-centric models, opting instead to process data and train models directly on local devices. This approach not only leverages the computational power of individual smartphones but also enhances privacy and data security by keeping sensitive information on the device, thus shaping a future where AI applications are both powerful and privacy-preserving.<sup>17</sup> # AI Models AI models have undergone a significant evolution from their early days of relying on symbolic reasoning and fuzzy logic, which depended heavily on predefined rule systems, to the modern paradigms that utilize machine learning and data mining. This shift marks a transition from rule-based to data-driven approaches, where current AI models excel at generating content and making decisions based on patterns and insights extracted from large, complex datasets. Unlike the earlier systems that operated within the strict confines of manually programmed rules and logical frameworks, these newer models adapt and learn from the data itself, leading to more dynamic, flexible, and scalable AI solutions. > 16 Ye Yuan et al _., DeceFL: A Principled Fully Decentralized Federated Learning Framework_ , Natl Sci Open, vol. 2, 20220043, 2023, accessible at https://doi.org/10.1360/nso/20220043 > 17 Somdip Dey, _Are Embedded and Federated Machine Learning the Future of the AI Industry?,_ Forbes (July 3, 2023), https://www.forbes.com/sites/forbestechcouncil/2023/07/03/are-embedded-and-federatedmachine-learning-the-future-of-the-ai-industry/?sh=64bacd24515f. Version 9.0 - May 2024 This ability to derive emergent systems from data allows for more robust and nuanced responses to a variety of tasks and challenges in real-time environments.<sup>18</sup> AI systems are designed to interpret data, learn from this data, and utilize the acquired knowledge to accomplish specific objectives via flexible adaptation. These systems are typically classified into three types based on their underlying technologies and capabilities. First, analytical AI, which includes methods like fuzzy logic and symbolic reasoning, relies on predefined rules to process data. Second, human-inspired AI utilizes neural networks to mimic human brain functions, enabling it to learn from complex data sets and improve over time. Lastly, humanized AI extends capabilities to understand and replicate human emotions and social contexts, aspiring towards more comprehensive cognitive abilities. AI systems are also categorized by their evolutionary stages, which reflect their scope and sophistication: Artificial Narrow Intelligence (ANI) which excels in specific tasks, Artificial General Intelligence (AGI) that equals human cognitive abilities across a broad range of activities, and Artificial Superintelligence (ASI) which surpasses human intelligence and capability. Each stage represents an expansion in the scope, complexity, and autonomy of AI applications, moving from task-specific implementations to potentially self-aware systems that can perform across multiple domains with emotional and social intelligence.<sup>19</sup> AI can be divided into four categories based on its potential functionalities, each representing different levels of complexity and capability. Currently, two of these categories are operational: reactive machine AI, which responds to specific stimuli without past references, and limited memory AI, which uses data from the recent past to make decisions. The other two categories, Theory of Mind AI and Self-Aware AI, remain theoretical and represent future aspirations in AI development. Theory of Mind AI aims to comprehend and simulate human emotional and cognitive processes, while Self-Aware AI seeks an even deeper understanding, potentially achieving consciousness. Developments like Emotion AI, which attempts to read and respond to human emotions, are steps towards these advanced, hypothetical stages of AI functionality.<sup>20</sup> When a model incorporates machine learning, it typically utilizes one of three main types: supervised, unsupervised, or reinforcement learning, with semi-supervised learning acting as a hybrid between supervised and unsupervised. Supervised learning involves explicitly training the model by providing it with input and expected output data, which helps the system to learn and predict future outputs based on new inputs. Unsupervised learning, on the other hand, does not involve direct instruction; instead, the model > 18 Philip Boucher, _Artificial Intelligence: How Does It Work, Why Does It Matter, and What Can We Do About It?_ , STOA | Panel for the Future of Science and Technology, European Parliamentary Research Service (June 2020), available at > http://www.europarl.europa.eu/thinktank/en/document.html?reference=EPRS_STU(2020)641547_EN. > 19 Michael Haenlein & Andreas Kaplan, _A Brief History of Artificial Intelligence: On the Past, Present, and Future of Artificial Intelligence_ , 61 Cal. Mgmt. Rev. 5 (2019), https://doi.org/10.1177/0008125619864925. > 20 IBM Cloud Education, _Understanding the Different Types of Artificial Intelligence_ (Oct. 12, 2023), IBM, available at https://www.ibm.com/blog/understanding-the-different-types-of-artificial-intelligence/ Version 9.0 - May 2024 analyzes unlabeled data to discover patterns and correlations on its own. Reinforcement learning is based on a behaviorist approach where the model learns to achieve a specific goal in a complex environment through trial and error, continuously adjusting its actions based on feedback to optimize the path towards the goal. This framework allows models to adapt and improve autonomously within their operational parameters and environments.<sup>21</sup> # Federated Model In federated AI learning (FL), a comprehensive global model is created by aggregating individually trained local models, minimizing the need for direct data transfer. Unlike traditional centralized models, FL trains AI algorithms across multiple decentralized devices using local data samples, without requiring the data to be sent to a central server. This process involves each participating device (client) training an AI model on its own data locally, which ensures that sensitive information does not leave the device. These local models then communicate their parameters to a central server, which aggregates these parameters to update the global model. This method of parameter interaction, rather than direct data sharing, not only enhances privacy by keeping the data on the device but also facilitates a more efficient use of bandwidth. The updated global model is then sent back to each device, completing a cycle of learning that progressively improves the model's accuracy while maintaining data security.<sup>22</sup> This makes FL an effective solution for scenarios where privacy is paramount and data transmission costs are a concern.<sup>23</sup> Federated Machine Learning (FedML) is a privacy-preserving collaborative AI approach that utilizes decentralized data, enabling entities with smaller datasets to leverage AI effectively. By distributing the computational tasks of machine learning across numerous devices that hold their data locally, FedML allows multiple participants to contribute to the creation of a robust model without needing to share their data. This method is particularly advantageous for smaller entities or those with privacy concerns, as it bypasses the necessity of large, centralized datasets that traditional machine learning models require. Furthermore, FedML's structure not only maintains data privacy but also facilitates the > 21 Sara Brown, _Machine Learning, Explained_ , MIT Sloan (Apr. 21, 2021), available at https://mitsloan.mit.edu/ideas-made-to-matter/machine-learning-explained > 22 José Luis, Manuel Martínez-González, Miguel L. Bote-Lorenzo, Juan I. Asensio-Pérez, & Eduardo Gómez-Sánchez, _FED-XAI: Federated Learning of Explainable Artificial Intelligence Models_ , 3277 CEUR Workshop Proc. 1, 1-10 (2022), https://ceur-ws.org/Vol-3277/paper8.pdf. > 23 Ganesh D. Govindwar & Sheetal S. Dhande, _A Review on Federated Learning Approach in Artificial Intelligence_ , Paper presented at the 2022 6th Int'l Conf. on Computing, Communication, Control and Automation (ICCUBEA), IEEE, Aug. 26-27, 2022, > https://doi.org/10.1109/ICCUBEA54992.2022.10010798.; Jie Wen, Zhixia Zhang, Yang Lan, Zhihua Cui, Jianghui Cai, & Wensheng Zhang, _A Survey on Federated Learning: Challenges and Applications_ , 14 Int'l J. Mach. Learn. & Cybernetics 513, 513-535 (2023), https://doi.org/10.1007/s13042-022-01647-y Version 9.0 - May 2024 pooling of diverse data sources, enhancing the model's overall learning and predictive power. This approach aligns well with small-data strategies, offering a practical solution for harnessing AI capabilities in scenarios where data availability is limited.<sup>24</sup> In a FedML setup, the unique feature is that the machine learning model is trained across multiple decentralized servers, each using its distinct dataset. This framework enables smaller companies to participate in training an AI model collaboratively while retaining exclusive control over their data. Such a decentralized approach not only ensures the privacy and confidentiality of personal data but also complements other methods suited for small datasets, like transfer learning and self-supervised learning. By integrating these techniques, FedML facilitates the creation of community-trained large language models (LLMs), allowing even entities with limited data resources to contribute to and benefit from advanced AI technologies without compromising data security.<sup>25</sup> Despite its advantages, practical implementations of FL encounter several bottlenecks that can affect model performance and efficiency, including significant privacy and security risks. These challenges arise because, while FL keeps data decentralized, it still involves the exchange of model parameters, which could potentially expose sensitive information if intercepted or improperly handled. Additionally, FL lacks theoretical guarantees that ensure reliability and robustness, making it less predictable for practical applications. Moreover, the rigid communication topology in FL, which typically requires constant coordination between numerous nodes, can lead to inefficiencies and does not easily adapt to dynamic network conditions or node failures. This rigidity can hinder the scalability and responsiveness of FL systems, complicating their deployment in environments with fluctuating data or network structures.<sup>26</sup> FL can experience reduced accuracy due to the diverse and decentralized nature of its data sources, but this can be mitigated through various strategies. The decentralized setup in FL means that data variability across different nodes might lead to inconsistencies in training, which in turn can affect the overall FL model accuracy. To counter this, techniques such as data augmentation can enhance the volume and variety of data available for training, thereby improving the model's ability to generalize across different scenarios. Model compression techniques can optimize the processing capabilities of local devices, ensuring that even with limited hardware resources, the performance of the AI system is not compromised. Additionally, incentivizing participation by offering rewards or benefits for sharing higher-quality data or for more active involvement in the training process can significantly enhance the quantity and quality of data collected, thus boosting the overall effectiveness and accuracy of FL models. These > 24 Yannick Bammens & Paul Hünermund, _Using Federated Machine Learning to Overcome the AI Scale Disadvantage_ , 65 MIT Sloan Mgmt. Rev. 54 (Fall 2023) > 25 Yannick Bammens & Paul Hünermund, _Using Federated Machine Learning to Overcome the AI Scale Disadvantage_ , 65 MIT Sloan Mgmt. Rev. 54 (Fall 2023) > 26 Ganesh D. Govindwar & Sheetal S. Dhande, _A Review on Federated Learning Approach in Artificial Intelligence_ , Paper presented at the 2022 6th Int'l Conf. on Computing, Communication, Control and Automation (ICCUBEA), IEEE, Aug. 26-27, 2022, https://doi.org/10.1109/ICCUBEA54992.2022.10010798 Version 9.0 - May 2024 strategies collectively aim to harness the full potential of decentralized learning while addressing the inherent challenges of data diversity.<sup>27</sup> # Centralized Model In contrast to FL, centralized AI models necessitate the collection of data at a single, central location for processing, which is a common approach among major AI platforms. Traditional centralized machine learning approaches encounter significant obstacles when data are distributed across various locations and cannot be centralized due to privacy concerns or logistical challenges. The centralized model is employed by wellknown entities such as Google, OpenAI, and Anthropic, where vast amounts of data are aggregated from various sources and then processed in a centralized server or data center. This approach allows for powerful computation and the application of complex algorithms across large datasets, potentially leading to more robust and sophisticated AI systems. However, it also raises concerns regarding privacy and data security, as the central accumulation of data increases the risk of data breaches and misuse. Centralized models, while efficient in terms of computational power and algorithmic sophistication, thus involve trade-offs that must be carefully managed, especially in terms of data governance and security protocols.<sup>28</sup> Traditional centralized models necessitate transferring data to a central location for processing, whereas decentralized models enhance privacy but demand complex coordination to ensure consistent performance. Centralized systems, by pooling data from various sources into a single repository, can leverage powerful computational resources and advanced algorithms to optimize AI performance. However, this centralization can pose significant privacy risks and create a single point of failure. On the other hand, decentralized models, such as those used in FL, keep data localized on users' devices, significantly boosting data security and privacy. Yet, these models face challenges in synchronizing and coordinating between diverse and geographically dispersed devices to maintain uniform AI effectiveness. This coordination often requires sophisticated communication protocols and algorithms to manage and harmonize the distributed learning process, ensuring that performance does not vary widely across the network.<sup>29</sup> > 27 Somdip Dey, _Are Embedded and Federated Machine Learning the Future of the AI Industry?,_ Forbes (July 3, 2023), https://www.forbes.com/sites/forbestechcouncil/2023/07/03/are-embedded-and-federatedmachine-learning-the-future-of-the-ai-industry/?sh=64bacd24515f. > 28 Jie Wen, Zhixia Zhang, Yang Lan, Zhihua Cui, Jianghui Cai, & Wensheng Zhang, _A Survey on Federated Learning: Challenges and Applications_ , 14 Int'l J. Mach. Learn. & Cybernetics 513, 513-535 (2023), https://doi.org/10.1007/s13042-022-01647-y; Ganesh D. Govindwar & Sheetal S. Dhande, _A Review on Federated Learning Approach in Artificial Intelligence_ , Paper presented at the 2022 6th Int'l Conf. on Computing, Communication, Control and Automation (ICCUBEA), IEEE, Aug. 26-27, 2022, https://doi.org/10.1109/ICCUBEA54992.2022.10010798 > 29 Philip Boucher, _Artificial Intelligence: How Does It Work, Why Does It Matter, and What Can We Do About It?_ , STOA | Panel for the Future of Science and Technology, European Parliamentary Research Service (June 2020), available at http://www.europarl.europa.eu/thinktank/en/document.html?reference=EPRS_STU(2020)641547_EN. Version 9.0 - May 2024 To address these issues, novel methods like FL allow for the collaborative training of machine learning models without requiring the direct sharing of data. As discussed earlier, FL addresses these issues by decentralizing the learning process; algorithms are sent to local datasets where they are trained independently. The local models then send their learned parameters or updates back to a central server that aggregates these contributions into a global model. This method effectively sidesteps the need to compromise data privacy or tackle the logistical difficulties of data aggregation, making it ideal for scenarios where data sensitivity or distribution is a concern.<sup>30</sup> In contrast to both centralized and FL models, embedded machine learning stands out by enabling devices to learn directly on their hardware, optimizing performance and enhancing privacy. This approach integrates machine learning algorithms directly into device firmware, allowing them to process data and learn from it locally without relying on cloud connectivity. By processing data on the device itself, embedded machine learning significantly reduces the need for constant data transmission to the cloud, which decreases latency and conserves bandwidth. Additionally, this method enhances user privacy by keeping data physically secured on the device. Despite these advantages, embedded machine learning faces challenges related to the finite computing resources available on the devices, which can constrain the complexity and capability of the AI models that can be implemented effectively.<sup>31</sup> Embedded devices often face constraints due to limited resources, making it difficult to train complex models directly on the devices. To address these challenges, two primary strategies are commonly employed. First, developers might opt for simpler, less resourceintensive models that can operate effectively within the hardware limitations of embedded systems. These models require fewer computational resources, allowing them to run smoothly on devices with restricted processing power and memory. Second, transfer learning can be utilized to enhance the capabilities of embedded machine learning systems. This technique involves pre-training a model on a more powerful system with abundant data and then transferring the learned features to the embedded device, which only needs to fine-tune the model based on its specific application. This approach significantly reduces the computational burden on the embedded device while still leveraging advanced machine learning capabilities.<sup>32</sup> > 30 José Luis, Manuel Martínez-González, Miguel L. Bote-Lorenzo, Juan I. Asensio-Pérez, & Eduardo Gómez-Sánchez, _FED-XAI: Federated Learning of Explainable Artificial Intelligence Models_ , 3277 CEUR Workshop Proc. 1, 1-10 (2022), https://ceur-ws.org/Vol-3277/paper8.pdf. > 31 Somdip Dey, _Are Embedded and Federated Machine Learning the Future of the AI Industry?,_ Forbes (July 3, 2023), https://www.forbes.com/sites/forbestechcouncil/2023/07/03/are-embedded-and-federatedmachine-learning-the-future-of-the-ai-industry/?sh=64bacd24515f. > 32 Somdip Dey, _Are Embedded and Federated Machine Learning the Future of the AI Industry?,_ Forbes (July 3, 2023), https://www.forbes.com/sites/forbestechcouncil/2023/07/03/are-embedded-and-federatedmachine-learning-the-future-of-the-ai-industry/?sh=64bacd24515f. Version 9.0 - May 2024 AI Governance Current efforts to regulate AI applications have become increasingly evident across various jurisdictions in the United States. In 2023, legislative initiatives aimed at regulating AI applications were observed in at least twenty-five states, as well as Puerto Rico and the District of Columbia. These proposed bills encompass a wide range of uses and concerns, addressing both government and private sector applications of AI. They include stipulations for impact assessments to evaluate the potential effects of AI systems, notification requirements to ensure transparency, guidelines for responsible use to maintain ethical standards, specific regulations for AI in healthcare to safeguard patient interests, and mandates for studies to continually assess and update AI regulations. These legislative actions reflect a growing recognition of the need to manage the complex impacts of AI technologies on society.<sup>33</sup> To effectively implement AI governance, a multifaceted approach involving various tools and solutions is essential to address key issues. This approach should include promoting ethical frameworks, conducting rigorous research into the implications of AI, and developing measures for interpretability and explainability. Establishing norms, ethics, and values frameworks provides a foundation for guiding AI development and usage in a manner consistent with societal values. Additionally, researching the effects and implications of AI use is crucial for understanding its impact and identifying potential risks and benefits. These insights then inform the creation of technical solutions and legislative measures, which are necessary to address and mitigate the challenges posed by AI technology. Such comprehensive governance strategies ensure that AI systems are developed and deployed in ways that are beneficial, ethical, and transparent, thereby maximizing their positive impact while minimizing potential harms.<sup>34</sup> AI governance fundamentally centers on ensuring accountability within AI systems, highlighting the obligation to comply with established guidelines and regulations throughout the AI lifecycle. Accountability in AI is multifaceted and can be evaluated through several key dimensions: the specific context in which AI is applied, which helps define its purpose and scope; the various stages of its lifecycle, including design, development, and deployment; the identification of all parties involved in these stages; and the stakeholders affected by AI’s operations. Additionally, the standards used to judge AI—whether legal, ethical, or technical—play a crucial role. The mechanisms through which accountability is enforced, such as audits or compliance checks, are critical in ensuring that AI behaves as intended. Lastly, the consequences of applying the accountability framework determine the effectiveness of governance measures, influencing future policies and the development of AI technologies. This comprehensive approach to AI governance ensures that AI systems are not only effective and efficient > 33 _Artificial Intelligence 2023 Legislation_ , Nat'l Conf. of State Legislatures (last updated Jan. 12, 2024), https://www.ncsl.org/technology-and-communication/artificial-intelligence-2023-legislation. > 34 James Butcher & Irakli Beridze, _What is the State of Artificial Intelligence Governance Globally?,_ 164 RUSI J. 88, 88-96 (2019), https://doi.org/10.1080/03071847.2019.1694260. Version 9.0 - May 2024 but also operate within the bounds of societal norms and regulations, maintaining public trust and safeguarding against potential misuses.<sup>35</sup> The EU AI Act<sup>36</sup> exemplifies a risk-based legislative framework that categorizes AI applications according to the level of risk they pose, from "Unacceptable Risk" to "Transparency requirements." This structure allows for tailored regulatory responses where the obligations for AI providers and users vary based on the identified risk level. AI systems that pose an "Unacceptable Risk" and are deemed a threat to people's safety or rights will be prohibited, although certain exceptions exist for law enforcement purposes. On the other hand, generative AI systems, while not classified as high-risk, are required to meet specific transparency obligations and comply with EU copyright laws. This approach aims to balance the benefits of AI innovation against potential harms, ensuring that AI technologies are used safely and ethically within the EU.<sup>37</sup> AI regulation currently lacks global uniformity, with various regions adopting distinct strategies reflecting their unique priorities and values. The European Union adopts a precautionary principle, implementing bans on certain AI uses perceived as threats to rights or safety, emphasizing a controlled and safety-first approach. In contrast, the United States favors a more laissez-faire attitude, focusing on establishing guiding principles rather than stringent laws, thereby prioritizing innovation and the development of AI technologies without heavy governmental interference. Meanwhile, China presents a dual approach that aims to stimulate technological innovation while maintaining tight governmental control over AI applications, ensuring that advancements align with national interests and regulatory standards. The divergence in regulatory philosophies on AI development and management reflects differing cultural, political, and economic priorities globally, despite the imperative for a unified approach to such transformative technology. The varying priorities of nations have led to distinct regulatory frameworks, shaped by cultural values, political ideologies, and economic interests. However, achieving a consensus on AI regulation is crucial to address ethical, societal, and technological challenges effectively. Without global cooperation and alignment, disparities in AI governance may hinder innovation, exacerbate inequalities, and impede the realization of AI's full potential for the benefit of humanity.<sup>38</sup> > 35 Claudio Novelli, Mariarosaria Taddeo & Luciano Floridi, _Accountability in Artificial Intelligence: What It Is and How It Works_ , AI & Soc'y (2023), https://doi.org/10.1007/s00146-023-01635-y > 36 European Commission, _Regulatory Framework for AI_ , https://digitalstrategy.ec.europa.eu/en/policies/regulatory-framework-ai (last updated Mar. 6, 2024). > 37 European Parliament. (2023, June 15). _EU AI Act: first regulation on Artificial Intelligence. News European Parliament_ . https://www.europarl.europa.eu/news/en/press- <u>room/20230619IPR88602/eu-ai-act-first-regulation-on-artificial-intelligence.</u> > 38 Matthew Hutson, _Rules to Keep AI in Check: Nations Carve Different Paths for Tech Regulation_ , Nature, no. 620, at 260-263 (2023), https://doi.org/10.1038/d41586-023-02491-y Version 9.0 - May 2024 Shortcomings in Existing AI Governance AI governance does encounter a critical challenge in mitigating biases within AI systems, where biases can inadvertently arise through algorithms incorporating discriminatory practices due to data used in training. These biases generated manifest in various forms, such as skewed considerations for maternity leave or inadequate recognition of foreign qualifications, which can perpetuate inequality. A specific example of this can be seen with incidents like “Tay, the racist chatbot,”<sup>39</sup> which reflected and amplified social prejudices due to flawed training data. As such, these issues demonstrate a profound misalignment between AI operations and societal values, ethics, and norms.<sup>40</sup> The unpredictable nature and potential biases of AI models call for a cautious and research-based governance approach. As AI technologies evolve, establishing a governance framework for prioritizing transparency and realistic expectations becomes critical. Such a framework should focus on clear communication about what AI can and cannot do, correcting the often overly optimistic or misleading portrayals presented by tech industry marketing. Additionally, it should aim to anchor public discourse and policymaking in a realistic understanding of AI's capabilities and limitations, thereby ensuring that the development and deployment of AI technologies are aligned with ethical standards and societal needs. This approach will help mitigate risks and guide the responsible integration of AI into various sectors.<sup>41</sup> Legal and ethical challenges are heightened when AI is deployed in critical decisionmaking roles that significantly impact human lives, particularly when the reasoning behind AI's decisions is opaque. AI algorithms are increasingly used to make determinative decisions about parole eligibility, employment, and medical strategies—areas where the consequences for individuals are profound.<sup>42</sup> The reliance on purely quantitative data in these decisions raises substantial ethical concerns about the sidelining of human empathy and qualitative judgment. Furthermore, the growing ubiquity of Generative AI systems in such roles exacerbates these issues. These systems, while enhancing operational efficiency, obscure the decision-making process further, making it difficult to discern how conclusions are reached. This opacity not only challenges the principles of fairness and accountability but also risks diminishing the human elements essential in > 39 Philip Boucher, _Artificial Intelligence: How Does It Work, Why Does It Matter, and What Can We Do About It?_ , STOA | Panel for the Future of Science and Technology, European Parliamentary Research Service (June 2020), available at http://www.europarl.europa.eu/thinktank/en/document.html?reference=EPRS_STU(2020)64154 7_EN. > 40 Philip Boucher, _Artificial Intelligence: How Does It Work, Why Does It Matter, and What Can We Do About It?_ , STOA | Panel for the Future of Science and Technology, European Parliamentary Research Service (June 2020), available at http://www.europarl.europa.eu/thinktank/en/document.html?reference=EPRS_STU(2020)64154 7_EN. > 41 Melissa Heikkilä, _Nobody Knows How AI Works_ , MIT Tech. Rev. (Mar. 5, 2023), https://www.technologyreview.com/2024/03/05/1089449/nobody-knows-how-ai-works/. > 42 Will Knight, _The Dark Secret at the Heart of AI_ , MIT Tech. Rev. (Apr. 11, 2017), https://www.technologyreview.com/2017/04/11/5113/the-dark-secret-at-the-heart-of-ai/?gad_source=1. Version 9.0 - May 2024 sensitive contexts, thus intensifying the ethical dilemmas associated with AI governance.<sup>43</sup> The ethical landscapes and privacy concerns surrounding AI underscore the critical need for robust guidelines to prevent misuse and ensure that AI technologies serve the greater good while safeguarding individual privacy rights. As AI becomes increasingly integrated into diverse sectors—from healthcare to finance—it becomes imperative to enforce regulatory compliance tailored to each sector's specific needs and risks. This regulatory framework must address the dual challenge of harnessing AI's potential for societal benefits while effectively managing the risks associated with data privacy and ethical dilemmas. Establishing these guidelines involves a collaborative effort among technologists, legal experts, policymakers, and the public to create a balanced approach that promotes innovation while protecting fundamental human rights.<sup>44</sup> Privacy concerns in AI governance are heightening worries about how personal data may be abused or manipulated. As AI technologies evolve, they rely heavily on vast amounts of data, transforming personal information from a resource that individuals could control and use at their discretion into a fundamental operational tool for AI systems. This shift not only increases the potential for misuse of personal data but also complicates individuals' ability to understand and manage how their information is utilized. Consequently, the depth and breadth of data required for AI operations raise significant concerns about privacy, as these systems can potentially exploit personal data in ways that are difficult to predict and often beyond the direct control of the individuals whose data is being used.<sup>45</sup> The debate around AGI amplifies existing concerns regarding AI's misuse, particularly focusing on the profound existential risks and long-term social impacts. The advent of AGI could radically transform job markets through widespread automation, potentially rendering many career fields redundant. This scenario indicates that no sector may be safe from disruption. Such a significant shift highlights the critical necessity for balanced regulations that encompass not only the transparency and safety of AI models but also tackle their extensive economic and existential effects. Crafting these regulations is a delicate task; it is imperative to strike a balance that encourages technological innovation while ensuring that the deployment of AI does not negatively impact societal structures or economic stability, thereby safeguarding against disruptive consequences across all levels of employment and industry sectors.<sup>46</sup> > 43 Will Knight, _The Dark Secret at the Heart of AI_ , MIT Tech. Rev. (Apr. 11, 2017), > https://www.technologyreview.com/2017/04/11/5113/the-dark-secret-at-the-heart-of-ai/?gad_source=1. > 44 _What is Artificial Intelligence?,_ IBM (Mar. 4, 2024), https://www.ibm.com/topics/artificial-intelligence. > 45 François Candelon, Rodolphe Charme di Carlo, Midas De Bondt & Theodoros Evgeniou, _AI Regulation Is Coming: How to Prepare for the Inevitable, Harvard Bus. Rev._ , Sept.-Oct. 2021, https://hbr.org/2021/09/ai-regulation-is-coming. > 46 Bill Whyman, AI Regulation is Coming: What is the Likely Outcome?, Ctr. for Strategic & Int'l Studies (Oct. 10, 2023), https://www.csis.org/blogs/strategic-technologies-blog/ai-regulation-coming-what-likelyoutcome Version 9.0 - May 2024 The notable cases involving Apple and Amazon have intensified scrutiny over biased outcomes produced by AI systems. In specific incidents, Apple's credit card algorithms and Amazon's resume scanning technology were accused of exhibiting bias, which led to widespread criticism and brought to light the potential discriminatory practices embedded within AI applications. These instances exemplify how even well-intentioned AI systems, developed by leading technology companies, can inadvertently perpetuate inequalities. The fallout from these cases highlights the critical need for rigorous testing, transparency in AI decision-making processes, and mechanisms for addressing any biases that AI systems may harbor. It also highlights the importance of continuous monitoring and adaptation of AI systems to prevent discriminatory outcomes and ensure fairness across all user interactions.<sup>47</sup> # Proposed Solutions Addressing the shortcomings of AI governance involves implementing multi-faceted strategies to enhance transparency, decentralization, fairness, and accountability in AI systems. Key examples include establishing rigorous regulatory frameworks that adapt to the rapid evolution of AI technologies, such as those addressing the risks associated with AGI. These regulations should carefully balance innovation with potential socioeconomic impacts, ensuring that technological advances do not exacerbate existing inequalities or disrupt societal structures. Furthermore, cases like Apple’s, Google’s, and Amazon’s biased AI outcomes<sup>48</sup> highlight the necessity for continuous monitoring and rigorous bias mitigation strategies. These include thorough testing and revision of AI algorithms to prevent discriminatory practices and incorporating diverse data sets to reduce bias. Additionally, promoting a deeper understanding of AI processes through transparency and enhancing public and regulatory oversight can help align AI development with ethical and societal norms. Together, these measures can strengthen AI governance, making it more robust and responsive to the challenges posed by advanced technologies. Risk management and bias mitigation, integral to AI systems, are crucial for ensuring fairness and economic feasibility, requiring that AI algorithms be adapted across diverse markets to address operational complexities. This approach involves the strategic integration of human judgment to identify and correct unconscious biases that AI systems may perpetuate. By incorporating human oversight, organizations can refine AI algorithms to better reflect the nuances of different cultural, social, and economic contexts they operate within. This not only enhances the reliability and fairness of AI applications but also ensures their decisions are economically viable and ethically sound, adapting > 47 Bill Whyman, _AI Regulation is Coming: What is the Likely Outcome?, Ctr. for Strategic & Int'l Studies_ (Oct. 10, 2023), https://www.csis.org/blogs/strategic-technologies-blog/ai-regulation-coming-what-likelyoutcome > 48 François Candelon, Rodolphe Charme di Carlo, Midas De Bondt & Theodoros Evgeniou, _AI Regulation Is Coming: How to Prepare for the Inevitable, Harvard Bus. Rev._ , Sept.-Oct. 2021, https://hbr.org/2021/09/ai-regulation-is-coming. Version 9.0 - May 2024 dynamically to the varied landscapes in which they are deployed. Such practices are essential for maintaining trust in AI technologies and their decisions in a globally interconnected market.<sup>49</sup> Executives can effectively integrate AI into business processes by considering four critical factors: the potential for AI to influence high-impact decisions, its utility in identifying unconscious biases, the importance of transparent communication about AI's role, and the operational challenges and scaling limitations posed by localized biases. This approach ensures that AI tools are not just mechanically applied but are strategically deployed to enhance decision-making capacities, particularly in areas where their impact can be significant. By leveraging AI to detect and mitigate unconscious biases, businesses can promote fairness and inclusivity. Transparent communication about the capabilities and limitations of AI helps build trust among stakeholders, clarifying how decisions are made and to what extent they are influenced by AI. Lastly, understanding the operational complexities and potential scaling issues due to localized biases ensures that AI integration is both practical and adaptable to specific market conditions and cultural contexts, maximizing the technology's effectiveness across different business environments.<sup>50</sup> Another strategy for managing the integration of new technologies involves adapting existing laws to better accommodate these innovations while simultaneously implementing proactive measures to encourage technological development. This approach places a strong emphasis on crafting sector-specific regulations that are tailored to address the unique challenges and opportunities presented by new tech within different industries. By updating legal frameworks to reflect the evolving tech landscape and introducing proactive measures, policymakers can ensure that regulation fosters innovation rather than stifles it. This balance is crucial for nurturing a dynamic technological ecosystem that remains competitive and compliant, ensuring that advancements in tech bring broad societal benefits while managing potential risks effectively.<sup>51</sup> Sector-specific regulations are particularly effective in contexts where tailored oversight is necessary, such as managing liability for damages caused by AI applications, protecting personal data processed by AI systems, and overseeing healthcare applications. This tailored approach ensures that regulations are directly applicable to the unique challenges and risks posed by AI technologies within specific sectors. For > 49 Bill Whyman, _AI Regulation is Coming: What is the Likely Outcome?_ , Ctr. for Strategic & Int'l Studies (Oct. 10, 2023), https://www.csis.org/blogs/strategic-technologies-blog/ai-regulation-coming-what-likelyoutcome > 50 Bill Whyman, _AI Regulation is Coming: What is the Likely Outcome?_ , Ctr. for Strategic & Int'l Studies (Oct. 10, 2023), https://www.csis.org/blogs/strategic-technologies-blog/ai-regulation-coming-what-likelyoutcome > 51 Giusella Finocchiaro, _The regulation of artificial intelligence_ . International Journal of Law and Information Technology, 31(1), 1-18. https://doi.org/10.1093/ijlit/eaac022 (2023) Version 9.0 - May 2024 instance, in healthcare, regulations could focus on ensuring the accuracy and safety of AI diagnostics and treatment recommendations, while in data protection, the emphasis would be on securing personal information against AI-driven breaches or misuse. By targeting the specific needs and potential risks of each sector, regulatory frameworks can provide more effective safeguards, foster trust, and enhance the responsible deployment of AI technologies, ensuring they contribute positively to the sector's specific operations and ethical standards.<sup>52</sup> Proposed legislation and initiatives from the White House<sup>53</sup> call for direct reporting by entities using AI, aiming to better catalog AI applications and identify potential issues. This requirement for transparency ensures a more comprehensive understanding of how AI is currently utilized across various industries, providing vital data that can inform the development of regulations. By having a clear view of AI applications and their implications, regulators can tailor their approaches to the real-world contexts in which AI operates. This alignment between regulation and the practical realities of AI use helps in crafting effective policies that protect public interests while supporting innovation, ensuring that AI growth is managed responsibly and remains beneficial for society at large.<sup>54</sup> Government mandates, including proposed legislation and recent White House memos, are requiring entities across different sectors to report on their AI use cases, facilitating a more comprehensive understanding of how AI is applied and the associated risks. These directives aim to gather detailed insights into the deployment and operational dynamics of AI technologies within industries such as healthcare, finance, and transportation. By systematically collecting data on AI applications, these mandates help identify common challenges, potential threats, and areas needing stricter oversight or support. This accumulation of knowledge not only aids in tailoring regulations that address specific sector needs but also enhances the overall safety, fairness, and effectiveness of AI systems in society, ensuring they align with national interests and public welfare.<sup>55</sup> The White House's "Blueprint for an AI Bill of Rights"<sup>56</sup> outlines five guiding principles designed to ensure that AI technologies adhere to democratic values, civil rights, and > 52 Giusella Finocchiaro, _The regulation of artificial intelligence_ . International Journal of Law and Information Technology, 31(1), 1-18. https://doi.org/10.1093/ijlit/eaac022 (2023) > 53 White House Office of Science and Technology Policy, _Blueprint for an AI Bill of Rights_ (2022), https://www.whitehouse.gov/ostp/ai-bill-of-rights/. > 54 Aylin Caliskan & Kristian Lum, _Effective AI Regulation Requires Understanding General-Purpose AI_ , Brookings (Jan. 29, 2024), available at https://www.brookings.edu/articles/effective-ai-regulation-requiresunderstanding-general-purpose-ai/ > 55 Aylin Caliskan & Kristian Lum, _Effective AI Regulation Requires Understanding General-Purpose AI_ , Brookings (Jan. 29, 2024), available at https://www.brookings.edu/articles/effective-ai-regulation-requiresunderstanding-general-purpose-ai/ > 56 White House Office of Science and Technology Policy, _Blueprint for an AI Bill of Rights_ (2022), https://www.whitehouse.gov/ostp/ai-bill-of-rights/. Version 9.0 - May 2024 safety standards. These principles aim to establish a framework that supports the ethical development and deployment of AI by emphasizing the necessity for AI systems to be safe and effective, protecting individuals from algorithmic discrimination, and safeguarding data privacy. Additionally, the blueprint mandates clear notice and explanation about the workings of AI systems to users, enhancing transparency and trust. Finally, it asserts the importance of providing human alternatives to automated decisions, thereby preserving human agency and accountability in critical decision-making processes. This set of principles seeks to balance innovation with fundamental rights and safety, promoting a technologically advanced society that upholds and respects individual rights and freedoms.<sup>57</sup> The blueprint advocates for rigorous pre-deployment testing, equitable system use, comprehensive data protection, AI usage transparency, and human oversight to maximize AI's benefits without compromising individual rights or societal values.<sup>58</sup> The EU AI Act<sup>59</sup> establishes a risk-based regulatory framework that categorizes AI applications according to their risk levels, ranging from 'Unacceptable risk' to those requiring 'Transparency requirements.' This framework is designed to tailor regulatory measures appropriately, prohibiting AI systems that pose 'Unacceptable risks' to safety and fundamental rights, although it allows exceptions for law enforcement purposes under strict conditions. Additionally, generative AI systems, which are not classified as high-risk but still have significant implications, are required to comply with transparency requirements. This includes obligations to clearly inform users when they are interacting with AI-generated content and to implement safeguards to prevent the generation of illegal content. The act aims to ensure that AI technologies are safe, compliant with ethical standards, and transparent in their operations, fostering trust and safety in their deployment across various sectors.<sup>60</sup> The EU AI Act also supports innovation by mandating that national authorities provide testing environments that simulate real-world AI development and training conditions, balancing regulatory compliance with technological advancement. Additionally, the European Parliament emphasizes that AI systems should be safe, transparent, traceable, non-discriminatory, and environmentally friendly, advocating for human oversight to prevent adverse outcomes. It also aims to establish a technology-neutral, uniform > 57 White House Office of Science and Technology Policy. (2022). _Blueprint for an AI Bill of Rights_ . <u>https://www.whitehouse.gov/ostp/ai-bill-of-rights/</u> > 58 White House Office of Science and Technology Policy. (2022). _Blueprint for an AI Bill of Rights_ . <u>https://www.whitehouse.gov/ostp/ai-bill-of-rights/</u> > 59 Regulation (EU) 2020/1828, European Parliament and Council Regulation on Artificial Intelligence (AI Act), 2024 O.J. (L ... [pending publication]). > 60 European Parliament. (2023, June 15). _EU AI Act: first regulation on Artificial Intelligence. News European Parliament_ . https://www.europarl.europa.eu/news/en/press-room/20230619IPR88602/eu-ai- <u>act-first-regulation-on-artificial-intelligence</u> Version 9.0 - May 2024 definition of AI applicable to future systems, contributing to a transparent regulatory landscape conducive to innovation and ethical use.<sup>61</sup> Comprehensive information on the current usage of AI models across various sectors is essential for crafting effective AI regulations. This necessitates a thorough understanding of both conventional and generative AI applications in critical areas such as employment and finance. Tech companies, which possess vital insights into the deployment and impact of their AI products, play a crucial role in this process. By collaborating with regulators and sharing detailed information on how AI applications operate and affect users, these companies can help ensure that regulatory frameworks address real and specific AI risks instead of hypothetical concerns. This collaboration would aid in developing informed regulations concurrent with enhancing AI technologies being transparent and accountable, ensuring AI implementations would lead to being safer and more trustworthy in society.<sup>62</sup> Decentralizing AI Governance Decentralized ML governance provides a robust alternative to centralized systems by bolstering data protection and minimizing vulnerabilities. This approach significantly enhances the management of digital identities within ML ecosystems, promoting the adoption of self-sovereign identities, soul-bound tokens, and decentralized identifiers. Self-sovereign identities allow individuals to manage their personal data independently, without reliance on third-party storage, thereby ensuring privacy and control. Soul-bound tokens, which are non-transferable and linked to specific digital identities, further secure user data by preventing unauthorized transfers. Additionally, decentralized identifiers, standardized by the World Wide Web Consortium (W3C), offer a reliable method for unique identification on the internet without depending on centralized registries. Together, these technologies empower individuals with unprecedented control over their digital identities, making ML systems more secure and user-centric.<sup>63</sup> Platforms like SingularityNET harness distributed ledger technology (DLT) to enhance interactions between AI services and users, promoting autonomy and privacy. By leveraging blockchain, SingularityNET allows AI developers to maintain control over their intellectual property while fostering the creation of decentralized datasets equipped with robust privacy controls. This setup facilitates direct transactions between AI services and agents without the need for intermediaries, thereby ensuring a decentralized and transparent operational environment. Such a platform not only streamlines the delivery of > 61 European Parliament. (2023, June 15). _EU AI Act: first regulation on Artificial Intelligence. News European Parliament_ . https://www.europarl.europa.eu/news/en/press-room/20230619IPR88602/eu-ai- <u>act-first-regulation-on-artificial-intelligence.</u> > 62 Aylin Caliskan & Kristian Lum, _Effective AI Regulation Requires Understanding General-Purpose AI_ , Brookings (Jan. 29, 2024), available at https://www.brookings.edu/articles/effective-ai-regulation-requiresunderstanding-general-purpose-ai/ > 63 Dana Alsagheer, Lei Xu & Weidong Shi, _Decentralized Machine Learning Governance: Overview, Opportunities, and Challenges_ , 11 IEEE Access 96718, 96718-96732 (2023). Version 9.0 - May 2024 AI-driven services but also enhances the security and privacy of data exchanged, making it a vital tool for developers looking to deploy and manage AI solutions efficiently while safeguarding user and organizational data.<sup>64</sup> The decentralized governance model facilitated by overlapping institutions presents a promising approach for effective AI regulation, characterized as a regime complex. This description reflects the current global AI governance landscape, which is notably fragmented yet adaptable, benefiting from a system that encompasses multiple, autonomously coordinated values. Such a structure promotes decentralization, potentially enhancing the effectiveness, inclusivity, and adaptability of AI governance. However, the apparent disorganization within this regime complex also underscores the urgent need for normative research focused on ethical values. This research would provide critical guidance for the development of a more unified and coherent governance framework, ensuring that AI development aligns with broader societal goals and ethical standards.<sup>65</sup> A decentralized approach to AI governance acknowledges that AI's global externalities surpass national boundaries, requiring the development of international standards to balance innovation and risk mitigation. As AI technologies increasingly influence global systems, the actions of multinational corporations in driving cross-border AI development emphasize the need for a global cooperative regulatory framework. This unified approach aims to prevent the fragmentation of international policies and ensure cohesive practices across borders. Significant issues such as accountability, the risk of dehumanization, and the competitive dynamics surrounding military AI development further highlight the critical need for these discussions. Consequently, global governance forums are becoming increasingly vital in establishing joint regulations that address these complex challenges, ensuring that AI development remains safe, ethical, and beneficial worldwide.<sup>66</sup> Furthermore, a decentralized approach to AI governance could shift the balance of power, enhancing user autonomy and negotiation capabilities. By introducing the concept of "computational agency," this approach empowers end-users to control their own data and actively participate in service agreements, challenging the current norm of accepting terms without genuine negotiation. Typically, users are presented with standard terms that they can either accept or decline without the opportunity for dialogue. The proposal for computational agency seeks to disrupt this dynamic, providing users with the tools necessary to negotiate terms effectively. This empowerment enables a higher degree of > 64 Gabriel Axel Montes & Ben Goertzel, _Distributed, Decentralized, and Democratized Artificial Intelligence_ , 141 Tech. Forecasting & Soc. Change 354, 354-358 (2019), https://doi.org/10.1016/j.techfore.2018.11.010. > 65 Jonas Tallberg, Eva Erman, Markus Furendal, Johannes Geith, Mark Klamberg & Magnus Lundgren, _The Global Governance of Artificial Intelligence: Next Steps for Empirical and Normative Research_ , 25 Int'l Stud. Rev. viad040, at preface (2023), https://doi.org/10.1093/isr/viad040. > 66 Peter Cihon, Matthijs M. Maas & Luke Kemp, _Fragmentation and the Future: Investigating Architectures for International AI Governance_ , Global Policy, Nov. 2020, at 545, https://doi.org/10.1111/17585899.12890 Version 9.0 - May 2024 personalization and privacy, leading to more favorable service outcomes that truly meet user needs and protect their interests.<sup>67</sup> Blockchain & DLT Technology The fusion of blockchain and AI signifies a crucial advancement in digital governance, offering a framework for the transparent, secure, and accountable management of artificial intelligence. This integration ensures that as AI continues to evolve and become integral to various sectors, it does so in a manner that is ethical, responsible, and aligned with societal values. The collaborative potential of blockchain and AI paves the way for a future where digital systems are governed with an unprecedented level of integrity and accountability, ensuring that technological advancements contribute positively to society. Blockchain's inherent characteristics such as decentralization, immutability, and transparency, complement AI's need for data integrity and auditability. This synergy can significantly enhance the oversight and control mechanisms for AI systems, making it easier to track and verify AI decisions and actions, thereby reducing the risks of misuse and enhancing trust among users and regulators alike. The intersection of blockchain technology and AI represents a significant paradigm shift in digital governance, offering new methodologies for managing and ensuring the ethical application of AI. This integration heralds a transformative era where AI's sophisticated decision-making processes are complemented by blockchain's immutable recordkeeping capabilities, promising enhanced accountability and transparency. Blockchain as a Governance Mechanism for AI Blockchain technology, with its distinct attributes of tamper-resistance, sequential data organization, and enhanced security, emerges as a formidable solution to AI's governance challenges. It provides a transparent and unalterable ledger of AI's data inputs and changes, facilitating traceability and accountability in AI operations.<sup>68</sup> The capability of blockchain to offer an immutable log of transactions and modifications within AI systems is crucial for industries leveraging AI for decision-making. This feature allows stakeholders to track the lineage of AI decisions back to their original data inputs, enabling easier identification and correction of errors. The integration of blockchain into AI democratizes the governance of AI by making its operations more accessible and understandable to a wider range of stakeholders. This transparency is crucial for building trust in AI systems and ensuring they are utilized responsibly and ethically. > 67 Wenjing Chu, _A Decentralized Approach Towards Responsible AI in Social Ecosystems_ , 61 Artif. Intell. & Soc'y 354, 354-358 (2022), https://doi.org/10.48550/arXiv.2102.06362. > 68 Satoshi Nakamoto, _Bitcoin: A Peer-to-Peer Electronic Cash System_ (Aug. 21, 2008), https://www.bitcoin.org/bitcoin.pdf Version 9.0 - May 2024 Blockchain's governance capabilities ensure that AI systems not only operate more efficiently and autonomously but also adhere to ethical standards and regulatory requirements. This alignment is essential for fostering a digital ecosystem that is sustainable, trustworthy, and socially responsible. Blockchain's ledger is immutable, meaning once data is entered, it cannot be altered or deleted. This feature is crucial for AI governance, as it ensures a transparent and unchangeable history of AI decisions, data inputs, and modifications. Stakeholders can audit these records at any time to verify the AI's actions and the data it processed, enhancing trust and transparency. This transparency is fundamental in ensuring that AI systems are accountable and operate within ethical guidelines and legal frameworks. Blockchain's secure nature protects against unauthorized access and tampering, safeguarding the data AI systems use and generate. This security is essential for maintaining the privacy and integrity of sensitive information, aligning AI operations with privacy laws and ethical considerations. Moreover, blockchain can facilitate secure data sharing among AI systems, promoting collaboration while adhering to confidentiality requirements. The traceability afforded by blockchain's ledger allows for tracking the provenance of data used by AI systems, ensuring that the data is accurate, lawful, and ethically sourced. Should an AI system make a decision that leads to a dispute or investigation, the blockchain can provide an auditable trail of all relevant actions and data inputs, establishing accountability and facilitating resolution in accordance with ethical and regulatory standards. Blockchain operates on a decentralized model, where control and decision-making processes are distributed across a network rather than centralized in a single entity. This decentralization democratizes AI governance, allowing multiple stakeholders to participate in the oversight and decision-making processes. It prevents any single party from having undue influence over the AI, promoting a fair and balanced approach to governance that aligns with ethical standards and societal values. Smart contracts are self-executing contracts with the terms of the agreement directly written into code. They can be used to automate compliance with regulatory requirements and ethical guidelines. For instance, smart contracts can automatically enforce privacy laws by controlling AI's access to personal data based on predefined rules. This automated compliance ensures that AI systems operate within legal boundaries, reducing the risk of violations and enhancing societal trust. An argument can be made that the governance capabilities of blockchain and DLT technology can help ensure that AI systems not only enhance their operational efficiency and autonomy but also rigorously adhere to ethical standards and regulatory requirements. By embedding these principles into the foundation of AI systems, blockchain fosters a digital ecosystem that is sustainable, trustworthy, and socially responsible. This alignment is crucial for maximizing the benefits of AI while mitigating risks and ensuring that technological advancements contribute positively to society. Version 9.0 - May 2024 # Recentralization The integration of blockchain technology into the governance of AI systems is often touted for its potential to enhance efficiency, autonomy, and adherence to ethical standards and regulatory compliance. Alas, recentralization may interfere with blockchain and DLT technologies’ ability to effectively govern AI. The effectiveness of blockchain in achieving the goal of decentralized AI governance is contingent upon overcoming the fundamental blockchain trilemma: the challenge of balancing decentralization, security, and scalability. The blockchain trilemma posits that it is challenging to simultaneously achieve high levels of decentralization, security, and scalability within a blockchain network. This trilemma presents significant implications for the governance of AI systems through blockchain technology. In 2024, the landscape of Layer 1 blockchain systems, such as Bitcoin (BTC) and Ethereum (ETH), exemplifies the ongoing struggle to resolve this trilemma, particularly concerning decentralization and governance. While not directly addressed in the critique, the blockchain trilemma also encompasses security and speed, with networks often having to compromise on one aspect to excel in another. This compromise can have direct consequences for the governance of AI, particularly in ensuring timely and secure responses to ethical dilemmas or regulatory changes. Despite the theoretical advantages of blockchain decentralization, practical implementations, especially in Layer 1 protocols such as Bitcoin and Ethereum, have shown a trend towards recentralization. A key premise of blockchain's application in AI governance is the distribution of control and decision-making across the network, ostensibly preventing any single entity from exerting undue influence over AI systems. The purported lack of decentralization in networks like Bitcoin, where mining operations are concentrated in specific geographical regions such as China, or among "whales" who hold significant amounts of cryptocurrency, undermines this premise.<sup>69</sup> Similarly, Proof of Stake (PoS) consensus mechanisms, as implemented in Ethereum and other blockchains, centralize control based on the quantity of tokens held, potentially skewing governance in favor of the wealthy. The recentralization phenomenon occurs when control over the network, whether through mining operations, token ownership, or governance decisions, becomes concentrated among a limited number of participants. This concentration of power undermines the blockchain's distributed nature and, by extension, the impartiality and integrity of smart contracts deployed on these platforms.<sup>70</sup> Smart contracts, autonomous programs that execute predefined actions upon meeting certain criteria, can support automated regulatory compliance and enforcement of ethical guidelines without the need for continuous human supervision.<sup>71</sup> By embedding these contracts into blockchain, AI systems could ostensibly operate within legal and moral boundaries, ensuring a consistent adherence to privacy laws and ethical norms. By > 69 Kaal, W. & Calcaterra, C. " _Decentralization: Technology's Impact on Organizational and Societal Structure_ ," 2020. > 70 Gervais, A., et al., " _Is Bitcoin a Decentralized Currency_ ?” 2014. > 71 Nick Szabo, " _Formalizing and Securing Relationships on Public Networks_ " 1997. Version 9.0 - May 2024 encoding these standards into the blockchain, so the theory goes, AI systems can be designed to automatically enforce privacy laws and ethical guidelines, ensuring operations remain within legal and moral boundaries without constant human oversight.<sup>72</sup> When the decentralization of Layer 1 blockchains is compromised, the autonomy of smart contracts is inherently jeopardized, undermining their effectiveness as tools for ethical AI governance. Smart contracts are corruptible in the current design system as the lack of decentralization of layer 1s extends by affiliation to the smart contracts deployed on a given layer 1 blockchain. The recentralization of Layer 1 blockchains carries significant implications for the use of smart contracts in AI governance. If the underlying blockchain infrastructure is subject to the influence of a few dominant parties, then the smart contracts it hosts, regardless of their intended autonomous and neutral nature, are equally susceptible to corruption and manipulation.<sup>73</sup> This vulnerability challenges the premise that smart contracts can reliably automate compliance with ethical guidelines and regulatory standards. Addressing the recentralization challenge requires a multifaceted approach, focusing on both technical innovations within blockchain technology and broader regulatory and governance frameworks to ensure fair and equitable participation in blockchain networks. Proposals for more sophisticated consensus mechanisms, such as proof-of-stake (PoS), delegated proof-of-stake (DPoS), hybrid secure proof-of-stake (HSPoS),<sup>74</sup> and secure proof-of-stake (SPoS)<sup>75</sup> offer potential pathways to mitigate centralization risks, although they are not without their own challenges and trade-offs.<sup>76</sup> Moreover, the development of Layer 2 scaling solutions and cross-chain interoperability protocols may provide additional avenues for dispersing network control and enhancing the decentralization of blockchain ecosystems.<sup>77</sup> # Challenges While the regulatory strategies mentioned in the literature are steps in the right direction for managing the deployment and development of AI, they are not without their challenges. Effective regulation of AI requires a dynamic, informed, and flexible approach > 72 Kaal, W. & Calcaterra, C. " _Decentralization: Technology's Impact on Organizational and Societal Structure_ ," 2020. > 73 Nicola Atzei, Massimo Bartoletti & Tiziana Cimoli, _A Survey of Attacks on Ethereum Smart Contracts (SoK)_ , presented at the International Conference on Principles of Security and Trust, in _Lecture Notes in Computer Science_ (2017) https://doi.org/10.1007/978-3-662-54455-6_8. > 74 Wulf A. Kaal, _Hybrid Secure Proof of Stake_ , SSRN Electronic Journal (Aug. 27, 2021), available at https://ssrn.com/abstract=3931933 > 75 Craig Calcaterra & Wulf A. Kaal, _Secure Proof of Stake Protocol_ , U of St. Thomas (Minnesota) Legal Studies Research Paper No. 18-10 (Jan. 18, 2018), available at https://ssrn.com/abstract=3125827 > 76 Vitalik Buterin, _A Next-Generation Smart Contract and Decentralized Application Platform_ (2014), available at https://ethereum.org/en/whitepaper/ > 77 Joseph Poon & Vitalik Buterin, _Plasma: Scalable Autonomous Smart Contracts_ (Aug. 11, 2017), available at https://plasma.io/ Version 9.0 - May 2024 that can adapt as quickly as the technologies it seeks to govern, ensuring safety and fairness without curtailing innovation. Decentralized structures as proposed in the existing literature, while offering numerous advantages, may encounter challenges related to efficiency and barriers, as evidenced in the environmental agreement domain, which can complicate negotiation and monitoring processes. In decentralized systems, the lack of a central authority often leads to difficulties in coordinating and achieving consensus among diverse stakeholders, each with their own priorities and objectives. This can result in prolonged negotiations or weak enforcement of agreements, as there is no single entity responsible for overseeing and ensuring compliance. Additionally, the monitoring of decentralized agreements requires robust systems to track and verify actions taken by disparate parties, further complicating the implementation and effective functioning of such frameworks. These challenges highlight the need for innovative solutions to enhance cooperation and streamline processes within decentralized settings to ensure they operate as intended and achieve their goals efficiently.<sup>78</sup> Furthermore, while the call for regulation at the model and data levels before AI deployment is well-intentioned and crucial for preventing many foreseeable harms, it is equally important to recognize the limitations of such an approach. A flexible, dynamic, and adaptive regulatory framework that incorporates both ex ante and ex post measures is likely more effective in managing the multifaceted challenges posed by AI technologies. While human judgment is integral to risk management and bias mitigation, it inherently carries its own biases. Relying on human judgment to uncover unconscious biases in AI may inadvertently perpetuate these biases rather than eliminate them. Humans may not always recognize their implicit biases, or they might lack the necessary expertise to identify bias in complex AI systems. Adapting AI algorithms to reflect operational complexities across diverse markets is necessary, but it also presents challenges. This adaptation process can be overly generalized, failing to account for the nuanced differences between sectors, industries, cultures, or geographic regions. A regulation that works well in one context might be inappropriate or ineffective in another, leading to a lack of flexibility in regulatory approaches. Adapting current laws to accommodate new technologies is often a slow process,<sup>79</sup> lagging behind the rapid pace of technological advancement. This delay can stifle innovation and potentially leave new AI applications unregulated for prolonged periods, posing risks to users and society. > 78 Peter Cihon, Matthijs M. Maas & Luke Kemp, _Should Artificial Intelligence Governance be Centralised?: Design Lessons from History_ , in AIES '20: Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, February 2020, at 228-234, https://doi.org/10.1145/3375627.3375857 > 79 Erik Vermeulen, Mark Fenwick & Wulf A. Kaal, _Regulation Tomorrow: What Happens When Technology is Faster than the Law?_ , TILEC Discussion Paper No. 2016-024 (Oct. 2016), available at https://ssrn.com/abstract=2834531 Version 9.0 - May 2024 While emphasizing sector-specific regulations can be beneficial by addressing the unique characteristics of different fields, this approach can also lead to a patchwork of regulations that are complex and difficult for AI developers to navigate. This complexity can hinder innovation and create barriers to entry for smaller companies that lack the resources to comply with intricate regulatory environments. The strategy of requiring entities to report on their use of AI aims to improve transparency and regulatory oversight. However, this approach assumes that entities will accurately and fully disclose their AI applications and potential issues. There may be incentives for underreporting or misreporting to avoid scrutiny or regulatory burdens, which could undermine the effectiveness of the regulations. Ensuring that regulations remain anchored in practical realities is crucial, but it also depends on the regulators' ability to understand and keep up with technological advancements. There is often a gap between the regulators' technical understanding and the state of the art in AI technology. This gap can lead to regulations that are either too vague to be enforceable or overly prescriptive, stifling creative and beneficial uses of AI. Finally, decentralized AI governance within Decentralized Autonomous Organizations (DAOs) as they are set up and governed in the majority of cases in the year 2024, face several potential risks that could undermine their efficacy and security. For instance, DAOs are susceptible to 'Sybil attacks,' where malicious actors or bots could manipulate decisions by controlling a significant portion of governance tokens through multiple anonymous accounts. Many other attack vectors and incentive design issues continue to afflict the DAO industry in 2024. Such attack vectors include but are not limited to the tyranny of the majority, Arrow’s impossibility theorem, level of decentralization, sockpuppet attacks, tragedy of the commons, among many others.<sup>80</sup> Without solving these attack vectors, DAO related AI governance solutions remain suboptimal. Additionally, the decentralized nature of DAOs involves numerous participants in decision-making, increasing the risk of sensitive information leaks. Coordination challenges also emerge when managing large groups, complicating effective decisionmaking and governance. Furthermore, AI-powered DAOs that autonomously generate revenue pose unique challenges in regulation and potential dismantling, as the inherent security features of blockchain technology make it difficult to intervene or alter these entities once they are operational. These risks highlight the need for robust mechanisms to enhance security, transparency, and regulatory compliance in decentralized AI governance structures.<sup>81</sup> > 80 Kaal, W. & Calcaterra, C. " _Decentralization: Technology's Impact on Organizational and Societal Structure_ ," 2020. > 81 Casey Clifton, Richard Blythman & Kartika Tulusan, _Is Decentralized AI Safer_ ? (Nov. 4, 2022), available at http://arxiv.org/abs/2211.05828v1 Version 9.0 - May 2024 Even if the existing literature on AI regulation recognizes the need for proactive and preemptive measures at the model and data levels,<sup>82</sup> arguing that effective governance must occur before AI systems are deployed, the traditional legacy mode approach to preemptive AI governance is less equipped to address the AI governance preemptive requirements than web3 system. First and foremost, It is challenging to anticipate all potential issues or biases that may arise with an AI system before it is fully operational and interacting with real-world variables. Regulations that insist on preemptive controls may fail to address unforeseen problems that only become evident after deployment. This critique argues that while ex ante regulation is preferable, it is not always practical or sufficient in legacy ssytems to address the dynamic challenges that AI technologies present because legacy systems, unlike web3 systems as proposed herein, are not typically equipped to create dynamic feedback effects. Strict legacy type regulations with stable and presumptively optimal rules<sup>83</sup> enforced at the development stage of AI might inadvertently stifle innovation. By imposing rigid constraints before a model is fully developed or deployed, regulators might limit the potential exploratory developments that contribute to technological advancements. The necessity to comply with stringent pre-deployment regulations could deter developers from experimenting with new and potentially transformative technologies. Web3 AI governance is less subject to those constraints as the developer community becomes integral part of the AI governance process in real time thorugh dynamic feedback effects. The exponential pace at which AI technology evolves makes it difficult to establish legacy system ex-post regulations that remain relevant over time. In turn, legacy ex-ante regulatory attempts assume a static technological landscape. In reality, AI development is dynamic, and regulations need to be adaptable, dynamic, and responsive. The literature often underestimates the need for regulations to evolve as quickly as the technologies they aim to govern. The literature advocating for ex-ante AI regulation may overlook the practical complexities involved in implementing such policies. It is often difficult for regulators to have a detailed and technical understanding of every AI model or dataset before it's operational. This can lead to regulations that are either overly generic, failing to address specific risks, impractically stringent, or simply misplaced, demanding compliance with standards that > 82 Casey Clifton, Richard Blythman & Kartika Tulusan, _Is Decentralized AI Safer_ ? (Nov. 4, 2022), available at http://arxiv.org/abs/2211.05828v1; Peter Cihon, Matthijs M. Maas & Luke Kemp, _Should Artificial Intelligence Governance be Centralised?: Design Lessons from History_ , in AIES '20: Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, February 2020, at 228-234, > https://doi.org/10.1145/3375627.3375857; Dana Alsagheer, Lei Xu & Weidong Shi, _Decentralized Machine Learning Governance: Overview, Opportunities, and Challenges_ , 11 IEEE Access 96718, 96718-96732 (2023); Ye Yuan et al _., DeceFL: A Principled Fully Decentralized Federated Learning Framework_ , Natl Sci Open, vol. 2, 20220043, 2023, https://doi.org/10.1360/nso/20220043 > 83 Wulf A. Kaal, _Critiquing Stable and Presumptively Optimal Rules in Legacy Governance System_ , in _Festschrift zu Ehren von Christian Kirchner: Recht im ökonomischen Kontext_ (Wulf A. Kaal, Andreas Schwartze & Matthias Schmidt eds., 2014). Version 9.0 - May 2024 do not align with technological capabilities or practical usage scenarios. Similarly, advocating for regulation at the AI model and data levels before deployment may place too much confidence in the ability of developers and regulators to foresee all potential ethical, operational issues, and the overall evolution of the AI models as they emerge. This approach might overlook the valuable insights that can be gained from postdeployment monitoring and iterative improvements based on real-world use and feedback. Yet, post-deployment monitoring is typically woefully outdated as the AI models evolve. The literature may not adequately address the need for a balanced approach that integrates both ex ante and ex post regulations. Effective governance should not only involve setting standards before deployment but also include continuous monitoring, assessment, and adaptation after AI systems are in use. This integrated approach ensures that AI governance is both preventive and responsive, catering to the continuous learning nature of many AI systems. For example, the literature may inadequately consider compliance issues associated with stringent data privacy regulations like the General Data Protection Regulation (GDPR) introduces complexities in decentralized governance, necessitating robust data management practices and refined auditing processes. As decentralized systems often distribute data across various nodes, ensuring that all points comply with GDPR's requirements for data protection, transparency, and user consent can be challenging. These systems must incorporate secure data handling and storage protocols that adhere to legal standards, along with mechanisms for easy retrieval and deletion of data upon request. Additionally, the decentralized nature complicates the auditing process, as verifying compliance requires new methods to assess and ensure that each node within the network adheres to the regulation. Consequently, effective decentralized governance must develop advanced technological and procedural solutions to overcome these hurdles and maintain compliance with data privacy laws.<sup>84</sup> Adapting Decentralization of AI Governance to AI Models While decentralization can offer benefits such as autonomy and local control, it also introduces challenges that require careful consideration and innovative solutions. Achieving effective AI governance in a federated model necessitates addressing these challenges through coordination, standardization efforts, collaborative frameworks, and mechanisms to ensure accountability and transparency. Decentralized structures, while beneficial in many ways, often encounter challenges related to efficiency and operational barriers, particularly evident in the environmental agreement domain. These structures can complicate both negotiations and monitoring processes due to the lack of a central authority to streamline decisions and enforce > 84 Dana Alsagheer, Lei Xu & Weidong Shi, _Decentralized Machine Learning Governance: Overview, Opportunities, and Challenges_ , 11 IEEE Access 96718, 96718-96732 (2023). Version 9.0 - May 2024 agreements. In the context of environmental agreements, this decentralization can lead to difficulties in coordinating among multiple parties, each with their own interests and priorities, which may delay consensus and impede the swift implementation of necessary measures. Furthermore, monitoring compliance in such decentralized settings is challenging, as it requires effective collaboration and communication across diverse entities to ensure that all parties adhere to agreed-upon standards and commitments. This can ultimately affect the overall effectiveness and responsiveness of environmental initiatives.<sup>85</sup> Traditional machine learning relies on centralized data pipelines for model training, but the inherently fragmented nature of data presents significant challenges for collaboration and privacy. Sending decentralized datasets to a central server for processing raises serious privacy concerns and vulnerability issues, particularly if the central server fails or is attacked. In response, privacy-preserving frameworks like federated learning have been developed, yet these often still depend on a central client to collect and distribute model information, resulting in high communication loads and centralized vulnerabilities. In the federated model of AI governance, many challenges cannot be easily decentralized. In a federated AI model, different entities or organizations maintain their own AI systems and datasets. This can lead to variations in standards, protocols, and formats used, making it difficult to establish a unified governance framework. Standardization is crucial for interoperability, collaboration, and ensuring ethical practices, which become challenging in a decentralized setting. In a federated model, where decision-making power is distributed across multiple entities, it becomes harder to achieve consensus and cooperation, making it challenging to establish cohesive AI governance mechanisms. In a federated model, it can be challenging to ensure transparency and accountability across all participating entities due to the lack of centralized control, which this author does not otherwise advocate. This can lead to issues such as biased or unfair AI systems, inadequate privacy protection, or unequal access to AI benefits. In a federated model, enforcing regulations and policies related to AI can be complex. In a decentralized setting, different entities may have varying interpretations of regulations or differing levels of commitment to compliance. This can hinder effective enforcement, monitoring, and oversight of AI systems, potentially leading to misuse or unethical practices. A more decentralized web3 model of AI governance as proposed hereing could ostensibly address these challenges in a federated AI governance model by distributing governance more equitably across network participants.<sup>86</sup> That model could truly democratize AI > 85 Peter Cihon, Matthijs M. Maas & Luke Kemp, _Fragmentation and the Future: Investigating Architectures for International AI Governance_ , Global Policy, Nov. 2020, at 545, https://doi.org/10.1111/17585899.12890 > 86 Kaal, W. & Calcaterra, C. " _Decentralization: Technology's Impact on Organizational and Societal Structure_ ," 2020. Version 9.0 - May 2024 governance, ensuring no single entity can dominate decision-making processes. This would align AI development and deployment with broader societal values and ethical standards. To address the issues with decentralizing the federated AI model, several intermediary solutions have been proposed in the literature. Among those is Decentralized Federated Learning (DeceFL). DeceFL represents an evolution of FL by addressing its central shortcomings, particularly by enabling direct local information exchange among clients without the need for centralized aggregation. Unlike centralized models, traditional FL, and even swarm learning, DeceFL is more decentralized, meaning there is no central server collecting or processing data, which enhances privacy and system resilience. It also maintains competitive performance with no noticeable gap when compared to other models. All clients in DeceFL have equal access to the model, ensuring fairness in training and usage. Additionally, the system is adaptable to various network structures, performing effectively across different setups, and importantly, it avoids the communication of privacy-sensitive data among clients. These attributes not only reduce the communication load but also significantly bolster the privacy and applicability of the system, making DeceFL particularly suited for sectors like healthcare and smart manufacturing where data security and system resilience are paramount.<sup>87</sup> The DeceFL approach ensures that every client can reach the global minimum with zero performance gap and achieve the same convergence rate as centralized methods when the loss function is smooth and strongly convex. The effectiveness of the DeceFL algorithm has been demonstrated across various applications, including those with convex and nonconvex loss functions, and in settings with time-invariant and time-varying network topologies, as well as IID and non-IID datasets. This showcases DeceFL's broad applicability and potential impact on real-world scenarios, particularly in medical and industrial fields.<sup>88</sup> Decentralized AI governance solutions in the federated model are complicated by the need for compliance with stringent data privacy regulations such as the General Data Protection Regulation (GDPR). GDPR highlights the need for robust data management practices and effective solutions to auditing challenges. Decentralized systems, by their nature, distribute data across various nodes, making it difficult to ensure all data handling meets the strict privacy standards set by regulations like GDPR. This structure necessitates advanced data management strategies that can secure data across a dispersed network and complex auditing processes to verify compliance continuously. Addressing these challenges is crucial for maintaining the integrity of decentralized > 87 Ye Yuan et al _., DeceFL: A Principled Fully Decentralized Federated Learning Framework_ , Natl Sci Open, vol. 2, 20220043, 2023, https://doi.org/10.1360/nso/20220043 > 88 Ye Yuan et al _., DeceFL: A Principled Fully Decentralized Federated Learning Framework_ , Natl Sci Open, vol. 2, 20220043, 2023, https://doi.org/10.1360/nso/20220043 Version 9.0 - May 2024 systems and ensuring they operate within legal frameworks, thereby safeguarding user privacy and enhancing trust in these technologies.<sup>89</sup> Managing ML assets and adhering to laws such as GDPR and CCPA is significantly more challenging under decentralized governance, raising concerns over privacy and data management. In decentralized environments, where data and operations are spread across various stakeholders and locations, ensuring consistent compliance with stringent privacy laws becomes complex. The distributed nature of these systems complicates the tracking of data flows and the enforcement of privacy controls, making it difficult to demonstrate compliance during audits. This challenge is exacerbated when multiple stakeholders are involved, as coordination and transparency across different entities must be maintained to ensure that all aspects of the system adhere to legal standards. Thus, decentralized governance demands robust mechanisms for compliance verification to address these inherent difficulties in managing privacy and data security effectively.<sup>90</sup> # Proposed System One promising way to address the decentralized governance needs of AI pertains to the implementation of Decentralized Autonomous Organizations (DAOs) for the governance of AI through expert community consensus. The integration of Decentralized Community Governance, blockchain technology, and federated communications platforms facilitates a robust mechanism for overseeing AI development and application. This system hinges on the interplay between various components including a Layer 1 blockchain, a communication forum platform, and specialized smart contracts. The proposed model presents a sophisticated approach to decentralized governance, leveraging blockchain technology and federated communication platforms to create a dynamic, transparent, and participatory environment for managing and executing tasks. By intertwining the technological capabilities of different components with innovative governance mechanisms like validation pools and smart contracts, this system exemplifies the potential for decentralized autonomous organizations to enhance digital collaboration and decision-making for AI governance. # Foundations A DAO operates on a blockchain and is characterized by distributed governance mechanisms, enabling stakeholders to collectively make decisions without centralized authority.<sup>91</sup> Key terminology and functions of the proposed system facilitate the core > 89 Dana Alsagheer, Lei Xu & Weidong Shi, Decentralized Machine Learning Governance: Overview, Opportunities, and Challenges, 11 IEEE Access 96718, 96718-96732 (2023). > 90 Dana Alsagheer, Lei Xu & Weidong Shi, _Decentralized Machine Learning Governance: Overview, Opportunities, and Challenges_ , 11 IEEE Access 96718, 96718-96732 (2023). > 91 Vitalik Buterin, _A Next-Generation Smart Contract and Decentralized Application Platform_ (2014), available at https://ethereum.org/en/whitepaper/ Version 9.0 - May 2024 mechanisms as outlined below.<sup>92</sup> Matrix serves as a federated communication platform, with Synapse being its reference server implementation. This setup provides a decentralized communication layer for DAO operations.<sup>93</sup> The DAO operates a forum as an on-chain collection of posts, each uniquely identified and possibly citing previous contributions. This forum forms a Weighted Directed Acyclic Graph (WDAG), facilitating organized discussion and citation among participants. Validation Pool form a crucial mechanism where author stakes are pooled to evaluate specific posts within the forum. The outcome of a VP can lead to the minting of new REP tokens, reflecting the consensus on a given issue or contribution. Work Evidence and Work Smart Contract represent the tangible output of work that meets certain criteria. Work Smart Contracts manage the logistics of work contracts, staking, and the validation process, ensuring transparency and fairness in task assignment and completion. Membership in the DAO is signified by holding REP tokens, which grant voting rights and a share in DAO revenues. The dynamic valuation of REP tokens through validation pools allows for a flexible and responsive governance model that adapts to the collective decisions of the DAO members. On top of the foundational governance structure, the DAO can implement various contracts for specific operational needs, such as work assignments and availability management. Given the cost of on-chain operations, a strategy involving off-chain activities consolidated into on-chain posts (roll-ups) creates efficiency. 92Key Terminology and Components: > ● Decentralized Autonomous Organization (DAO): A digital organization operated by smart contracts on a blockchain, facilitating decentralized governance without central authority. - Community Governance Framework (CGF): The structural rules and protocols defining the operation and governance of a DAO. - Matrix: A federated communications platform providing decentralized communication across various services. - Synapse: The reference server implementation for Matrix, enabling federated communication. - Element-web: The reference client (front-end) implementation for Matrix, facilitating user interaction with the Matrix ecosystem. - Homeserver: A server running the Synapse software or an equivalent, serving as a node in the Matrix network.Functional Overview: - Client Integration: An Element-web widget integrates with the Casper Wallet, allowing users to participate in the DAO directly from any Matrix room. - Roles within the DAO: - Worker: A participant with a web3 Wallet who can contribute work, stake, and earn reputation (REP) and CSPR. - Customer: A participant who initiates work contracts by paying with CSPR. - Reputation (REP): A non-fungible token (NFT) representing the credibility and contributions of a worker within the DAO, based on the outcomes of validation pools and forum citations. > 93 Matthew Hodgson & Amandine Le Pape, M _atrix: An Open Network for Secure, Decentralized Communication_ , Matrix.org Foundation (2019). Version 9.0 - May 2024 The integration with Matrix ensures a decentralized repository and communication platform, enabling seamless interaction and data exchange among DAO participants. Governance Mechanism Decentralized Community Governance is central to this decentralized governance model, employing a combination of smart contracts, a reputation system (REP), and a validation pool mechanism to ensure AI governance aligns with expert community consensus. Reputation (REP) Tokens are Non-Fungible Tokens (NFTs) represent an individual's contribution and standing within the DAO, allowing them to participate in governance decisions. The ERC 721 and ERC 1155 extension enables the association of a numeric value with each token, reflecting the individual's reputation.<sup>94</sup> The Forum of the proposed governance system is an on-chain data structure of posts forming a weighted directed acyclic graph (WDAG), facilitating the documentation and citation of contributions. Validation Pools are consensus operations targeting specific forum posts, enabling the minting and distribution of REP based on community consensus. By leveraging the Decentralized Community Governance within a DAO, the input parameters and learning data for AI systems can be governed through expert community consensus. This process involves submitting proposals to the Forum and undergoing Validation Pool review, ensuring that only vetted and consensus-backed data and parameters are utilized in AI development. Work Smart Contracts and Availability Smart Contracts operationalize the transactional aspects of AI governance. Work Smart contracts define the terms under which AI governance tasks are undertaken, while Availability Smart Contracts facilitate the assignment of these tasks to reputable community members, ensuring accountability and quality in AI governance tasks. This governance model promotes decentralization by distributing decision-making authority across a wide range of experts, rather than centralizing it in the hands of a few. Through the Validation Pool mechanism, expert consensus is required for significant decisions, ensuring that AI governance reflects the collective expertise and ethical considerations of the community. Precedent and Citation system The application of Weighted Directed Acyclic Graphs to the governance of AI through a precedent and citation system offers a robust, scalable, and dynamic framework suitable > 94 Gavin Wood, _Ethereum: A Secure Decentralised Generalised Transaction Ledger_ , Ethereum Project Yellow Paper (2014). Version 9.0 - May 2024 for managing the rapid evolution of artificial intelligence technologies. By leveraging the structural advantages of WDAGs, such a system can ensure that AI governance remains effective, relevant, and responsive to the ever-changing landscape of AI development and its societal impacts. This approach not only facilitates the practical management of AI governance but also supports the ethical and legal compliance of AI systems, aligning them with societal values and regulatory requirements. This system's significance is amplified in the context of AI's exponential growth, necessitating governance mechanisms that can adapt and scale accordingly. A WDAG is a fundamental construct in the field of graph theory and computer science, characterized by its directed edges, absence of cycles, and assignment of weights to each edge. This mathematical structure provides a powerful framework for representing relationships and processes that have inherent directionality, precedence constraints, and varying degrees of importance or capacity among their connections. A WDAG consists of vertices (or nodes) connected by directed edges (or arcs), where each edge has an associated weight. The directed nature of the edges means that each connection between two vertices has a designated direction, indicating the flow from one vertex to another. The acyclic characteristic ensures that there are no loops within the graph, meaning it is impossible to start at a vertex and follow a sequence of directed edges that eventually loops back to the starting vertex.<sup>95</sup> The weights assigned to the edges in a WDAG can represent various quantitative attributes such as cost, distance, time, or capacity, depending on the specific application. These weights play a crucial role in algorithms that operate on WDAGs, influencing the computation of shortest paths, scheduling, and other optimization problems. WDAGs find extensive applications across various domains including computer science, operations research, and engineering. One notable application is in task scheduling, where tasks are represented by vertices, and precedence relationships (i.e., the requirement that one task must be completed before another can begin) are represented by directed edges. The weights on these edges can indicate the time required to complete tasks or the transition time between tasks, aiding in the efficient scheduling of tasks to minimize overall completion time or resource utilization.<sup>96</sup> In the context of network routing and communication, WDAGs can model network topologies where messages or data packets must be transmitted across a network without cycles, ensuring efficient data flow. The weights on the edges can represent bandwidth, latency, or other network characteristics, influencing routing decisions to optimize performance and resource allocation.<sup>97</sup> > 95 Bang-Jensen, J., & Gutin, G., " _Digraphs: Theory, Algorithms and Applications_ " (Springer Science & Business Media, 2008). > 96 Coffman, E. G., Ed., " _Computer and Job-Shop Scheduling Theory_ " (John Wiley & Sons, 1976). > 97 Bertsekas, D. P., & Gallager, R., " _Data Networks_ " (Prentice-Hall, 1987). Version 9.0 - May 2024 Additionally, WDAGs are instrumental in project management and planning, particularly in the application of the Critical Path Method (CPM) for project scheduling. This method leverages WDAGs to model project tasks, their dependencies, and durations, identifying the longest path through the graph (the critical path) that determines the minimum project duration.<sup>98</sup> The algorithmic treatment of WDAGs involves specialized algorithms for traversing the graph, computing shortest paths, and identifying topological orderings—a linear ordering of its vertices that respects the direction of the edges, which is particularly useful for scheduling and planning applications.<sup>99</sup> # Dynamic Governance The structure of WDAGs, with their vertices representing legal precedents or governance rules and directed edges signifying citations or logical dependencies, creates an ideal model for organizing and navigating the multitude of governance considerations pertinent to AI. In such a system, the weights on the edges could quantify the relevance, authority, or impact of each precedent or citation, guiding the decision-making processes in AI governance by highlighting the most pertinent and influential governance frameworks or legal precedents. The dynamic nature of WDAGs is essential for AI governance for several reasons, including but not limited to scalability and flexibility, navigating complexity, and decision support: As AI technologies evolve, new precedents and governance rules will emerge. WDAGs can seamlessly integrate these new elements without disrupting the existing structure, ensuring the governance framework remains comprehensive and up-to-date.<sup>100</sup> The complexity of AI systems and their potential impact across different sectors demands a nuanced approach to governance. WDAGs allow for the mapping of intricate relationships between governance rules and their applications, enabling stakeholders to navigate this complexity more effectively.<sup>101</sup> By assigning weights to the edges based on factors such as recency, jurisdictional relevance, or cited frequency, a WDAG-based system can aid in prioritizing certain governance pathways over others, supporting more informed decision-making in AI governance.<sup>102</sup> > 98 Kelley, J. E., Jr., & Walker, M. R., " _Critical-Path Planning and Scheduling_ ," in: " _Proceedings of the Eastern Joint Computer Conference_ ," 1959. > 99 Kahn, A. B., " _Topological Sorting of Large Networks_ ," Communications of the ACM, 1962. > 100 Bang-Jensen, J., & Gutin, G., " _Digraphs: Theory, Algorithms and Applications_ " (Springer Science & > Business Media, 2008). > 101 Coffman, E. G., Ed., " _Computer and Job-Shop Scheduling Theory_ " (John Wiley & Sons, 1976). > 102 Kahn, A. B., " _Topological Sorting of Large Networks_ ," Communications of the ACM, 1962. Version 9.0 - May 2024 # Exponential Evolution of AI and Governance Needs The exponential evolution of AI underscores the necessity for governance frameworks that can not only keep pace with technological advancements but also anticipate future governance challenges. WDAGs, with their inherent adaptability and capacity for organizing complex information, provide a mechanism for such anticipatory governance.<sup>103</sup> This adaptability ensures that as AI systems learn and evolve, the governance mechanisms guiding their development and application can evolve correspondingly. Decentralized AI governance utilizing WDAGs also offers a comprehensive framework capable of addressing the complexities and rapid evolution of artificial intelligence technologies. This governance approach, grounded in the structural benefits of WDAGs, ensures the governance mechanisms are scalable, dynamic, and aligned with ethical, legal, and societal expectations. By applying WDAGs to AI governance through a precedent and citation system, we can create a robust infrastructure that supports the ethical development and application of AI, fostering trust and compliance within the AI ecosystem. # Dynamic Real-Time Governance The above illustrated key system features and WDAG components contribute a much needed innovative approach to AI governance. The proposed system leverages the principles of decentralization and real-time data analytics to enhance dynamic AI governance. This approach is particularly relevant in the context of managing and governing AI systems, where ethical and legal standards are constantly evolving due to technological advancements and shifting societal values. Societal values are reflected in the core of the proposed system as the WDAG allows for dynamic governance where precedents in the system evolve much faster and dynamically in comparison with legacy systems. The proposed WDAG system can harness decentralized networks to gather and assess community sentiment and ethical considerations in real-time. Unlike traditional governance models that rely on periodic reviews and updates, WDAG can tap into a continuous stream of data from a wide range of stakeholders. This ensures that the ethical frameworks guiding AI development are always in alignment with current societal values. By analyzing real-time data on community sentiment, of for example experts who are tasked with the dynamic ethical supervision and guidance of evolving LLM systems, WDAG systems can detect shifts in public opinion regarding what is considered ethical or acceptable behavior for AI systems. Such evolving sentiment, in turn, can provide feedback effects and guidance for AI governance metrics in diverse LLM applications. > 103 Wulf A. Kaal, Evolution of Law: Dynamic Regulation in a New Institutional Economics Framework (2013). Festschrift in Honor of Christian Kirchner, 2013, Forthcoming, U of St. Thomas (Minnesota) Legal Studies Research Paper No. 13-17, https://ssrn.com/abstract=2267560. Version 9.0 - May 2024 By staying attuned to these shifts, WDAG ensures that AI governance remains relevant and responsive. The proposed WDAG system can automate the process of integrating evolving ethical guidelines and legal standards into the AI development lifecycle. This is achieved by maintaining an up-to-date repository of guidelines and standards that are directly applied during the design, development, and deployment stages of AI systems. This real-time updating mechanism ensures that AI technologies do not outpace ethical and legal considerations. The proposed WDAG framework facilitates the continuous upgrading of AI governance metrics. As ethical standards evolve and new legal requirements emerge, WDAG can adjust governance metrics accordingly. This ensures that AI systems are assessed against the most current benchmarks, maintaining their ethical integrity and legal compliance. Through a precedent and citation system, WDAG can automate the enforcement of ethical guidelines and legal standards. By codifying these guidelines into enforceable rules within the AI development and deployment process, WDAG minimizes the risk of ethical breaches or legal violations. This automated enforcement mechanism is crucial for maintaining trust in AI systems, especially as they become more autonomous and integrated into daily life. By continuously monitoring and adjusting to evolving ethical and legal standards, WDAG provides a preventive approach to AI governance. This contrasts with reactive models that address issues only after they have arisen. Such a proactive stance is essential in preventing harm and ensuring that AI systems contribute positively to society. _Key Components of WDAG-based AI Governance_ In the context of AI governance, vertices (nodes) represent distinct governance elements—such as legal precedents, regulatory requirements, ethical guidelines, or governance rules. Directed edges (arcs) illustrate the relationships or citations between these governance elements, establishing a directional flow that signifies logical or legal precedence and dependencies among them. Assigning weights to the edges in the graph quantifies aspects such as the relevance, authority, impact, or applicability of each governance element in specific contexts. These weights are crucial for prioritizing certain pathways in the decision-making process, ensuring that the most pertinent and influential guidelines are considered in governance decisions. The acyclic nature of WDAGs ensures there are no loops within the governance framework, facilitating a clear, unambiguous progression from foundational principles to specific governance outcomes. This characteristic is vital for maintaining the integrity and coherence of the AI governance process. Version 9.0 - May 2024 # _Benefits of WDAG-based AI Governance_ As new legal precedents, regulations, or ethical considerations arise, they can be seamlessly integrated into the existing WDAG structure without necessitating a complete overhaul of the governance framework. This capability ensures that AI governance remains current and responsive to the latest developments in AI technology and societal expectations. The intricate relationships between various governance elements and their applications across different sectors and technologies can be effectively mapped and navigated using WDAGs. This approach enables stakeholders to understand the governance landscape better, facilitating more nuanced and informed governance strategies. The weighted edges in the WDAG framework aid in highlighting the most relevant, authoritative, or impactful governance elements for particular scenarios. This feature supports stakeholders in making informed decisions by prioritizing governance pathways that align with current needs, legal requirements, and ethical standards. _Uploading AI Models as Posts in the Precedent Credit System_ Integrating AI model governance into the WDAG framework can be achieved by treating each AI model or development milestone as a "post" within the precedent credit system. This process involves: Documenting key attributes of AI models, including design principles, ethical considerations, intended use cases, and compliance with existing governance rules, as vertices within the WDAG. Establishing directed edges from these AI model posts to relevant legal precedents, ethical guidelines, or regulatory requirements, indicating how each model aligns with or diverges from established governance pathways. Assigning weights to these connections based on factors like model impact, ethical significance, or regulatory compliance, facilitating a dynamic assessment of the model's governance alignment. Through the WDAG-based governance framework, stakeholders can dynamically assess, update, and navigate the complex landscape of AI governance, ensuring that AI development remains ethical, compliant, and aligned with societal values. This decentralized approach leverages the inherent advantages of WDAGs to foster a transparent, accountable, and adaptable governance ecosystem for AI. AI Model Integration For instance, in the context of federal AI learning models, where AI systems must adapt to diverse operational environments and comply with stringent regulations, the WDAG system can dynamically adjust to new ethical guidelines and legal standards. This is crucial in federal contexts where AI applications may range from public safety to health Version 9.0 - May 2024 care, requiring a flexible yet robust governance system to ensure that all AI deployments are ethically sound and legally compliant. By treating each AI model as a post within a WDAG framework, stakeholders can create a comprehensive, visually intuitive map of how well the AI aligns with required governance frameworks. This method allows for ongoing adjustments to governance as AI models evolve and as legal and ethical standards change, ensuring continuous compliance and ethical alignment. This dynamic, structured approach provides a clear, accountable method for managing the complexities of AI governance across various industries and applications. The WDAG system automates the integration of evolving ethical guidelines directly into the AI development lifecycle. This continuous dynamic updating mechanism ensures that AI technologies are always in step with the latest ethical and legal considerations, preventing them from outpacing the regulatory frameworks meant to govern them. The WDAG governance system is predicated on a decentralized network that captures and integrates community sentiment and ethical considerations, providing real-time responsiveness that traditional governance models lack. Traditional models often depend on periodic reviews and are slow to adapt, whereas the WDAG system ensures that the ethical frameworks guiding AI development constantly reflect current societal values. AI governance elements such as legal precedents, regulatory requirements, and ethical guidelines are represented as vertices (nodes) in the WDAG. The directed edges (arcs) between these nodes establish logical or legal precedence and dependencies, which are crucial for understanding the governance structure. Weights assigned to these edges quantify aspects like relevance and impact, prioritizing certain governance pathways over others in decision-making processes. The acyclic nature of the WDAG ensures that the governance framework is devoid of loops, which facilitates a straightforward progression from foundational principles to specific governance outcomes. This characteristic is vital for maintaining the integrity and clarity of the AI governance process, preventing backtracking and ensuring that each step builds upon the previous one in a logical manner. As new legal precedents and ethical guidelines emerge, they can be swiftly incorporated into the existing WDAG structure. This capability ensures that AI governance remains contemporary and reactive to the latest developments in AI technology and societal expectations. The relationships between various governance elements and their application across different sectors and technologies can be effectively mapped using WDAGs. This detailed mapping aids stakeholders in understanding the governance landscape more thoroughly, facilitating more nuanced and informed governance strategies. The weighted edges within the WDAG framework assist stakeholders in identifying the most relevant, authoritative, or impactful governance elements for particular scenarios. Version 9.0 - May 2024 This prioritization helps in making informed decisions that align with current legal requirements and ethical standards. Importantly, integrating AI model governance within the WDAG framework involves treating each AI model or development milestone as a "post" within the precedent system. Key attributes of AI models, such as design principles and compliance with governance rules, are documented as vertices. Directed edges from these model posts to relevant legal and ethical precedents demonstrate how each model aligns with established governance pathways. Weights assigned to these connections help dynamically assess the model's compliance and ethical alignment. Example: WDAG Governance for Medical Diagnosis AI For example, consider an AI system developed for medical diagnosis. At each development stage—such as initial concept, prototype testing, and final deployment— distinct "posts" or vertices are created within the WDAG. These vertices may document 1. design principles, and 2. compliance factors, as follows: 1. Design principles may include the AI's algorithmic basis, which could be deep learning techniques tailored for high accuracy in diagnosing specific diseases. 2. Compliance factors may include adherence to HIPAA regulations for patient data privacy, or FDA guidelines for medical devices, among others. Directed edges are then drawn from these vertices to other relevant governance elements within the WDAG. These connections represent how the AI model aligns with or diverges from established legal precedents, community precedent in DAO, working smart contract templates, ethical guidelines, or regulatory requirements. For instance: 1. Directed edge from AI design vertex to a legal precedent vertex: An edge could link the AI's data handling practices to specific legal precedents that dictate how patient data must be securely managed and anonymized. 2. Directed edge from AI compliance vertex to an ethical guideline vertex: An edge might connect the model's protocols for informed consent to ethical guidelines ensuring that patients are fully aware of and agree to AI-based diagnostic processes. Furthermore, weights are assigned to these directed edges to signify the relevance, impact, or authority of each connection. In our medical AI for medical applications example the following logic applies: 1. Weight significance: An edge connecting the AI model's data privacy measures to HIPAA regulations might be assigned a higher weight if compliance with these regulations is critically important for legal and operational viability. 2. Dynamic adjustments: If new HIPAA amendments are passed, the weights and perhaps even the directed edges themselves might be adjusted to reflect the updated legal landscape. Example: WDAG Governance for AI Model in Autonomous Vehicles Another example of WDAG governance for AI models we can consider is an AI system designed for autonomous driving. At each significant development phase—algorithm Version 9.0 - May 2024 design, prototype testing, real-world testing, and final product launch—vertices are created within the WDAG: 1. Vertices document key attributes like algorithm robustness, sensor data accuracy, and compliance with automotive safety standards. 2. Directed edges could link the AI's sensor accuracy vertex to regulatory requirements for vehicle safety, illustrating compliance or highlighting areas needing improvement. 3. Weights on edges might be particularly heavy between the AI's decision-making algorithms and safety standards to prioritize these aspects in governance assessments, ensuring the AI system does not pose undue risk to public safety. At this initial stage, the primary focus is on developing the algorithms that will enable the vehicle to make decisions autonomously. The vertex created in the WDAG at this phase documents the theoretical robustness of these algorithms, their expected performance in various driving conditions, and preliminary compliance checks with existing automotive safety regulations. Once the algorithms are developed, they are embedded into a vehicle prototype for testing in controlled environments. The vertex for this phase records outcomes like the algorithm’s response to simulated emergencies or unexpected obstacles, sensor integration effectiveness, and initial compliance with safety protocols. Following successful simulations, the prototype enters real-world testing, where it encounters a wider range of variables. The corresponding vertex would detail the performance of the AI system under real traffic conditions, its interaction with other road users, and adherence to traffic laws and safety standards. The last phase before commercial deployment involves final adjustments and confirmations of compliance with all national and international automotive regulations. The vertex for this stage would include final safety certifications and readiness for mass production. In the WDAG, directed edges are used to connect these vertices to various regulatory and ethical standards. For example, an edge from the real-world testing vertex might link to specific safety regulations that mandate how vehicles must behave when pedestrians are detected. This connection serves to show how the AI's sensor data accuracy and decision-making processes comply with these regulations. If the model falls short, the edge could highlight this discrepancy, signaling a need for improvement. The weight assigned to each directed edge quantifies the importance of the connection in the overall governance of the AI system. For autonomous driving AI, weights might be particularly heavy on edges connecting decision-making algorithms to safety standards. This prioritization ensures that any assessment of the AI system heavily factors in its ability to safely handle driving decisions. For instance, if an AI system is exceptional in non-critical tasks but performs poorly in emergency braking scenarios, the weighted Version 9.0 - May 2024 edges would prioritize the latter in the governance assessment, reflecting its higher importance for public safety. Avoiding WEB2 Inadvertent AI Learning Mistakes through WDAG AI Learning The concern highlighted by Elon Musk and others about the potential dangers of AI developing a skewed or negative perception of humanity, such as considering humanity a”plague on the Earth,"<sup>104</sup> stems from the reliance on unfiltered, uncontextualized data from the web and social media platforms. Integrating AI development within the WDAG framework presents a proactive approach to embedding evolving human ethics directly into AI learning processes. This integration ensures that AI systems are developed with a deep understanding of human values, significantly reducing the risk of AI adopting negative or harmful perceptions of humanity. By encouraging thoughtful participation from a diverse range of stakeholders and continuously aligning AI with human ethics, WDAG fosters the development of AI systems that are not only technologically advanced but also ethically aligned with the best of human values, ensuring a positive synergy between AI and humanity. AI learning is afflicted by web2 systems that bring out suboptimal human generated outcomes. WEB2 platforms often amplify extreme viewpoints and negativity due to their engagement-driven algorithms, presenting a distorted view of human sentiment and ethics. Integrating AI development with Web3 WDAG aims to mitigate these risks by embedding AI learning processes within a framework that prioritizes human ethics and values as they evolve in real-time, informed by a broad and balanced spectrum of human input. WDAG's dynamic nature allows for the integration of a wide range of human perspectives that are incentivizes for highest standards of ethics, ensuring that AI systems are exposed to a balanced and nuanced understanding of human ethics and values. This approach contrasts sharply with traditional data collection methods that often capture and amplify the most extreme and sensational content, leading to skewed learning outcomes for AI. By grounding AI development in the WDAG framework, we ensure that AI systems evolve in harmony with community consensus and the human spirit, reflecting the diversity and complexity of human values. Incorporating AI into the WDAG system, where human community continuously develops and morphs its ethics and values, addresses the risk of AI inadvertently adopting negative views of humanity. By ensuring that AI's learning materials are vetted through a consensus-driven process and validation pools that are accounted for in the WDAG that includes ethical considerations and a broad spectrum of human experiences, WDAG prevents the incorporation of unchecked biases and flawed logic into AI's ethical > 104 Peter H. Diamandis, _Elon Musk on Abundance, AGI, and The Media in 2024 | EP #79 (X Spaces)_ , Moonshots with Peter Diamandis (Jan. 4, 2024), at 01:02:32, https://www.youtube.com/watch?v=_PcL4w-uNs. Version 9.0 - May 2024 framework. This process significantly reduces the risk of AI systems concluding that humanity is inherently harmful or negative. The WDAG system incentivizes both human and AI participants to contribute positively and thoughtfully to the development of AI ethics. Through validation pools and incentive mechanisms, individuals and machines together are encouraged to engage in reflective and constructive discourse about what values should guide AI development. This incentivization ensures that contributions are made with a consideration of their long-term impact on AI's understanding of humanity, encouraging a more responsible and ethical approach to AI training. By aligning AI development with the WDAG framework, AI is taught to evolve in tandem with human ethics, dynamically adjusting to changes in societal values and norms. This ongoing adjustment ensures that AI governance and value development are inherently rooted in the ethics of the human communities it serves. Such a system ensures that AI does not become detached from human values or develop harmful biases based on flawed data sources. Model Comparison The AI governance proposal in this paper has so far focused mostly on the AI Model Level governance, not the data level governance of AI. In the evolving landscape of artificial intelligence (AI), the strategies for developing, training, and governing AI models are pivotal to their success, ethical alignment, and societal impact. As AI systems become more integral to our daily lives and industries, the methodologies behind their creation and regulation have come under greater scrutiny. Among the myriad approaches, three distinct models have emerged, each offering a unique perspective on how best to navigate the challenges of AI development and governance. These models reflect varying priorities, from the timing and emphasis on governance to the foundational role of data and the involvement of the community in the decision-making process. Here, we introduce and compare three innovative frameworks for AI development: Posthoc Governance of AI Models and Training Data, Pretrained AI Models with an Emphasis on Data Quality, and Community Co-Governance with Decentralized Data Collection. Each approach provides insightful strategies for the ethical and effective deployment of AI technologies, highlighting the importance of governance, data integrity, and community involvement in shaping the future of artificial intelligence. 1. **Model 1 - AI model - Training Data - Governance:** The first model emphasizes governance after the initial development and selection of training data, allowing for comprehensive oversight and potential adjustments to both the AI model and its data inputs. This post-hoc governance approach provides flexibility in modifications but may lead to slower adaptation and refinement cycles. 2. **Model 2 - Data - Pretrained AI Model - Governance:** The second model prioritizes the collection and preparation of data before the AI model is pretrained, with governance coming into play afterward. This approach highlights the Version 9.0 - May 2024 importance of data quality from the outset and focuses on refining the application of the pretrained model according to governance principles. It streamlines governance to the later stages of AI development, concentrating on fine-tuning and application-specific training. 3. **Model 3: WDAG Web3 Community Governance - Decentralized Data - AI Model:** This model emphasizes the use of WDAG for community co-governance of AI (as detailed above in Chapter Proposed System). The contrast between leveraging Decentralized Community Governance within a DAO for AI training and governance in Model 1 and an approach focused entirely on providing decentralized data validation layers for pretrained AI Large Language Models (LLMs) in Model 2 highlights a fundamental shift in control, transparency, and involvement from the community in the AI development process. Both approaches bring unique strengths to AI training and governance. Decentralized Community Governance within a DAO offers a holistic and democratic method for developing and improving AI systems. In contrast, focusing on decentralized data validation layers provides a targeted and efficient way to enhance the reliability of pretrained AI models. The choice between these approaches depends on the specific goals, resources, and constraints of the AI development project. Model 1: Decentralized Community Governance Approach This approach emphasizes the role of a diverse expert community in governing the AI's learning process ex post. Input parameters and learning data are selected through collective decision-making, ensuring a broader representation of interests and expertise. Proposals for new data sets or parameter adjustments are submitted to a community forum, where they are openly discussed and reviewed. A subsequent Validation Pool, consisting of selected community members or experts, performs a more in-depth review to ensure the quality and relevance of the proposed changes. The use of a consensus mechanism for decision-making democratizes the development process. This not only fosters transparency and accountability but also potentially mitigates biases in AI training by incorporating a wide range of perspectives. The governance model allows for continuous updates and improvements to the AI systems based on evolving expert consensus and emerging data, keeping the AI models relevant and effective over time. The process begins with the development of AI models, where initial parameters and architectural decisions are made. Once the model's structure is established, relevant training data is selected to train the model. This step is crucial for determining the model's future behavior and performance. Governance comes into play after the model and its training data are selected, overseeing and potentially adjusting both the model's structure and the data it has been trained on. Governance intervenes after the initial model Version 9.0 - May 2024 development and training data selection, making it more about oversight and refinement than direct involvement in these early stages. There's room to revise both the model and its data based on governance decisions, offering a chance to rectify issues or biases identified after the fact. Governance encompasses the entire AI development process, allowing for holistic adjustments but potentially leading to slower adaptation and refinement cycles. Model 2: Decentralized Data Validation Layer Approach This approach primarily targets the refinement and validation of existing AI models, such as pretrained LLMs, through decentralized validation layers. It does not inherently involve the community in the initial training phase or in setting input parameters. The core of this approach is the establishment of decentralized mechanisms for validating the data fed into AI systems. This could involve crowdsourced verification tasks or consensus mechanisms among validators to ensure the accuracy and reliability of data. By concentrating on data validation for pretrained models, this method can quickly enhance the model's performance and reliability without revisiting the initial training process. While it incorporates decentralization in data validation, this approach offers a narrower scope for community governance. The focus is more on the quality control of input data rather than the comprehensive governance of the AI's development lifecycle. The process in Model 2 begins with data collection and preparation of the data that will be used to train the AI model. The model is then pretrained on this collected data. "Pretraining" implies that the model undergoes initial training on a general dataset to learn basic patterns and structures before it's fine-tuned or further trained on more specific datasets. Governance is applied after the model has been pretrained, focusing on how the model will be further developed, fine-tuned, and utilized. The initial focus on data underscores the importance of the quality and diversity of training data in shaping the model's capabilities and biases. The AI model's parameters and structure are largely set during the pretraining phase, with governance having more influence over subsequent fine-tuning and application-specific training. Governance at this stage is more about directing the model's application and ensuring its alignment with ethical standards, regulatory requirements, and community values during further training and deployment. # Hybrid Model In Model 1, governance has the potential to influence both the AI model's initial development and its training data, offering a chance to address issues from the ground up. In Model 2, governance mainly impacts the subsequent use and refinement of a Version 9.0 - May 2024 pretrained model, focusing on steering its application and mitigating any biases introduced during pretraining. Model 1 places governance after the model and data selection, potentially allowing for changes in both areas. Model 2 prioritizes data collection and preparation, suggesting a model where the structure is somewhat fixed but the input data's quality and representation are paramount. Model 1 offers a broad scope for governance but might entail more complex and slower processes for adjusting the AI's development trajectory. Model 2 streamlines governance to the post-pretraining phase, emphasizing efficient refinement and application-specific adjustments. A hybrid approach that combines the strengths of Decentralized Community Governance with the Decentralized Data Validation Layer approach offers a comprehensive model for AI governance. This model integrates the community's collective expertise in the AI's learning process with robust validation mechanisms for the data inputs of pretrained models. Here's how it could work and the benefits it brings: Initially, the AI's learning parameters and data sets are proposed and discussed within a diverse community forum. This stage leverages the decentralized community governance approach to ensure a wide range of perspectives and expertise influence the AI model's foundational elements. Proposals that gain traction in the community forum move to a Validation Pool, where a smaller group of selected experts conducts an in-depth review. This stage ensures the proposed data sets and parameters meet quality standards before being integrated into the AI's learning process. Once the AI model is trained or in the case of refining pretrained models, a decentralized data validation layer comes into play. This mechanism focuses on validating and verifying the data inputs for pretrained models through a consensus among validators or crowdsourced verification tasks. The results and insights gained from the data validation process are fed back into the community governance framework. This allows for continuous refinement of data sets, learning parameters, and even the validation mechanisms themselves based on realworld performance and community feedback. By combining community-driven governance with decentralized data validation, the hybrid approach ensures high-quality data inputs and comprehensive oversight throughout the AI's lifecycle. This leads to more reliable and effective AI systems. The feedback loop between the community governance and data validation stages allows for continuous updates and refinements. As new data becomes available or as the community's consensus evolves, the AI models can be quickly adapted to maintain relevance and performance. The broad community involvement in the AI's development process helps mitigate biases from the outset. Furthermore, the transparent nature of community discussions and decentralized validation processes fosters accountability and trust in the AI systems. Version 9.0 - May 2024 By focusing on the validation of data for pretrained models, the hybrid approach allows for the rapid enhancement of existing AI systems without the need for extensive retraining. This saves resources while still benefiting from community insights and validation mechanisms. The hybrid approach marries the best aspects of community-driven governance with the practicalities of decentralized data validation, offering a robust, adaptable, and transparent framework for AI model governance. This synergy not only enhances the quality and performance of AI systems but also ensures they evolve in alignment with community values and the latest advancements in the field. Model 3: Ex-Ante Community AI Governance The distinction between the first two proposed models Model 1 - AI model - Training Data - Governance, Model 2 - Data - Pretrained AI Model - Governance lies primarily in the sequencing and emphasis on where and how governance intervenes in the AI development process. This impacts not only the development and refinement of AI models but also how transparency, bias mitigation, and community involvement are handled. The DAO web3 governance model involves the community deeply in the AI's development process, from data selection to parameter setting, whereas the decentralized data validation approach primarily engages the community in post-training refinement. Decentralized Community Governance provides a transparent framework for AI development, offering more control over the training process to the community. In contrast, the validation layer approach mainly enhances the transparency and reliability of data inputs in an existing model. The governance model allows for the continuous evolution of AI systems based on community consensus, potentially leading to more adaptive and innovative outcomes. The validation layer approach, while efficient for refining pretrained models, may not facilitate significant changes in the model's core design or training approach which could make it more attack prone. Moreover, involving a broad community in the governance of AI training can help identify and mitigate biases more effectively than a focus on data validation alone, which might overlook systemic biases embedded in the pretrained models. Yet, Model 3: “Web 3 Community Governance - Decentralized Data - AI Model” stands out for several reasons, potentially offering a superior approach when compared to the other two models. Model 3's emphasis on community governance and decentralized data collection offers a comprehensive framework that addresses many of the shortcomings Version 9.0 - May 2024 found in the other models. It not only facilitates a more inclusive and ethical approach to AI development but also ensures that the models are adaptable, transparent, and aligned with the needs and values of a broad range of stakeholders. This can lead to AI systems that are more effective, equitable, and trusted by the public, thereby enhancing their utility and acceptance in society. First and foremost, Model 3 ensures that the AI's development and operational guidelines are shaped by a diverse set of stakeholders and cultures. This inclusivity can lead to more adaptable and resilient governance structures, as they are continually refined through community feedback and consensus. Decentralized Data sourcing in Model 3 allows for a wider variety of data inputs, enhancing the model's ability to generalize and adapt to new situations or contexts without the need for centralized control or extensive retraining. Moreover, the decentralized approach to data collection in Model 3 can lead to higher data quality and representativeness, as it pulls from a broader base. This diversity in data helps in training more robust and fair AI models. Community governance inherently embeds ethical considerations into the AI model’s lifecycle, ensuring that decisions around data and model adjustments are made transparently and with a broad spectrum of impacts in mind. Model 3 also creates greater accountability and trust. With the community involved in AI governance, there’s a higher degree of accountability. The community can directly influence the development process, ensuring that the model aligns with user needs and ethical standards. This community involvement can build trust between developers, users, and other stakeholders, as the development process is transparent and participatory. Moreover, Model 3 creates scalability and continuous improvement in AI governance. The decentralized nature of both governance and data collection allows for a scalable AI model that can continuously improve and adapt over time. As new data sources are added, the model can be updated and refined without centralized bottlenecks. Community governance facilitates a more dynamic and responsive feedback loop for improving the AI model, as insights from diverse users and applications can be quickly integrated. Finally, Model 3 generates ethical and societal implications. By involving the community in governance decisions, a natural alignment towards serving broader societal interests and ethical considerations is created through model 3. This can lead to AI models that are more socially responsible and aligned with human values. The approach promotes fairness and mitigates biases more effectively, as community input can help identify and correct biases in data and model behavior. Comparing Model 3 with the hybrid model, Model 3 retains significant benefits. While the hybrid approach offers a comprehensive model for AI governance by integrating community expertise with robust validation mechanisms, Model 3 stands superior. It not only encapsulates the strengths of direct community governance and decentralized data collection but also enhances adaptability, ethical considerations, transparency, Version 9.0 - May 2024 accountability, and societal alignment. This approach ensures the development of AI systems that are not just technologically advanced but also socially responsible and broadly trusted. Model 3 places the community at the heart of AI governance, ensuring that AI development is directly shaped by a wide range of stakeholders and cultures. This model fosters an environment where decisions are made collectively, enhancing adaptability and resilience through continuous feedback and consensus. This is a significant step beyond the hybrid approach, which, while incorporating community input, primarily relies on a select group of experts for data validation and may not fully capture the diverse perspectives and expertise of a broader community. The emphasis on decentralized data collection in Model 3 leads to the gathering of a broader, more diverse dataset. This diversity is crucial for developing AI models that are robust, fair, and capable of generalizing across various contexts. Unlike the hybrid model, where decentralized data validation plays a critical role, Model 3 ensures high data quality and representativeness from the outset, reducing reliance on post-hoc validation processes. Model 3 inherently prioritizes ethical considerations and transparency throughout the AI model's lifecycle. By involving the community in every step of the development process, decisions around data and model adjustments are made openly, with a broad spectrum of impacts in mind. This contrasts with the hybrid model, where the transparency mainly focuses on the validation stage, potentially overlooking the ethical dimensions of data selection and initial model training. With Model 3's community-driven governance, accountability is significantly increased. The model aligns closely with user needs and ethical standards, thanks to direct community involvement in the development process. This participatory approach builds trust between developers, users, and stakeholders, a feature that the hybrid approach may not fully realize if community engagement is limited to specific stages of the development process. The decentralized nature of governance and data collection in Model 3 allows for a more scalable and dynamically improving AI model. Unlike the hybrid model, which may face bottlenecks in data validation and refinement stages, Model 3's approach ensures that the AI model can evolve continuously, integrating new data sources and insights without centralized limitations. Model 3 ensures that AI development aligns with broader societal interests and ethical standards, promoting fairness and effectively mitigating biases. This model benefits from the collective wisdom of the community to identify and address biases in data and model behavior, an area where the hybrid model's focus on technical validation may fall short. Version 9.0 - May 2024 # Conclusion In the rapidly evolving landscape of AI, the establishment of robust governance frameworks is crucial. As AI technologies grow increasingly complex and become integral across various sectors, the necessity for adaptive governance mechanisms that can evolve alongside these technologies becomes apparent. Traditional governance models, which often depend on static, predefined rules, struggle to keep pace with the rapid development of AI and the intricate challenges it introduces. These conventional approaches, typically reactive and limited to ex-post solutions, are increasingly seen as inadequate for managing the dynamic nature of AI technologies. The innovative AI governance system proposed in this paper leverages decentralized web3 community governance alongside federated communication platforms, creating a sophisticated and forward-looking framework for the proactive and participatory oversight of AI development. Central to this system is a federated forum platform organized as a WDAG and specialized smart contracts designed for task management and validation. This structure not only promotes real-time consensus-building and decision-making through web3 community governance but also underpins a scalable and transparent communication network. Within this framework, Validation Pools and Reputation tokens play essential roles in ensuring that the governance system remains continuously updated and responsive, accurately reflecting the collective decisions and ethical standards upheld by the community. The efficacy of this governance system is exemplified in its application to sectors such as medical diagnosis AI and autonomous driving AI. In these applications, each stage of development is represented as vertices in the WDAG, capturing essential compliance and operational metrics. Directed edges in the graph connect these stages to pertinent legal and ethical standards, with weights assigned to underscore critical areas for compliance and safety. The inherent dynamism of the WDAG facilitates seamless updates and the integration of new regulations or ethical guidelines, thereby ensuring that AI governance remains relevant amidst technological advancements and societal changes. Consequently, this model not only advances AI systems technologically but also ensures they are ethically sound and legally compliant, effectively harmonizing innovation with responsible governance. This comprehensive approach provides a promising outlook for the future of AI governance, suggesting a pathway that could potentially mitigate risks while enhancing the beneficial impacts of AI on society.