Notes on what a general intelligence does to a research program built on trust without trusted parties.
Abstract. Artificial general intelligence does not retire the oldest question of institutional economics: how self-interested actors produce reliable judgment when no one can check the work. It universalizes it. The essay argues that AGI’s effect on any research program turns on two variables, the distribution of capability and the cost asymmetry between generating and verifying, and that two decades of work on dynamic regulation, decentralized governance, and reputation systems amount to an explicit position on both. Four claims follow. A general intelligence makes the pacing problem categorical, leaving dynamic regulation as the only regulatory class with purchase. Machine-to-machine trust becomes economic infrastructure that must be manufactured, with reputation staking as the mechanism that scales without a human in the loop. The question that survives every capability gain is whether staked validation tracks truth or merely consensus. And as intelligence approaches zero marginal cost, scarcity migrates to verification, provenance, and accountable judgment. Two exposures are stated plainly: concentration and cheap verification. The wager was never that machines stay weak. It was that the stronger they become, the harder the old institutional questions bind.
The question I am asked most often now is what artificial general intelligence does to a body of work on decentralized governance, reputation systems, and dynamic regulation. The question assumes AGI is weather: an event that arrives and washes over everything equally. It is not weather. It is mechanism design at scale, and its effect on any research program depends on two variables, not on the arrival itself.
The first variable is distribution. Does capability concentrate in a handful of systems run by a handful of principals, or does it disperse across millions of autonomous agents acting for millions of principals? The second variable is the cost asymmetry between generating and verifying. Does it remain harder to check an output than to produce one, or does verification become as cheap as generation? Every research program touching AI is, whether its author admits it or not, a position on both variables. Mine is explicit about its position. That is the difference worth writing about.
The wager the corpus already made
For two decades the work has circled one structural fact: law reacts more slowly than the systems it governs. I called this the pacing problem in Dynamic Regulation for Innovation (2016), and the diagnosis was blunt: law has a diminishing capacity to react to innovation, and the remedies that matter are dynamic, not episodic. When the gap between innovation speed and regulatory speed was measured in years, dynamic regulation was one proposal among several. A general intelligence makes the gap categorical. A rulemaking cycle measured in years cannot govern systems that revise themselves in hours. At that point dynamic regulation stops being a school of thought and becomes the only class of proposals with purchase. The pacing problem was never a niche concern of technology lawyers. It was a preview.
The same holds for the institutional line of the work. The forty-DAO study, The Institutional Deficit in Decentralized Autonomous Organizations, documents what happens when coordination outruns institutional architecture: treasuries without separation of powers, voting without accountability, governance captured by whoever can borrow the most tokens for a block. Read narrowly, it is a paper about DAOs. Read correctly, it is a paper about what any population of autonomous actors does in the absence of engineered consequence. Substitute agents for token holders and the deficit does not shrink. It compounds.
Trust without trusted parties
Here is the load-bearing claim. When autonomous agents transact, delegate, and contract with one another at machine speed, trust stops being a courtesy extended between humans and becomes economic infrastructure that must be manufactured. The mechanisms that manufacture it cannot require a human in the loop, because the loop is precisely what the agents have left behind. What scales is reputation with stake at risk: earned standing that appreciates with validated work and burns when the work fails validation. This is not a metaphor for accountability. It is accountability, implemented.
That is the through line connecting the weighted directed acyclic graph governance architecture, the validation pool designs, and the reputation mechanisms in papers like Citation Honesty Mechanisms in WDAG Governance and Governance as a Product. It is also the argument of AI’s Mother’s Instinct: alignment is not a property you install in a model, it is a property that emerges when agents carry consequence, when their skin is in the game in a form they cannot shed. Engineered consequence is institutional design for a population that never sleeps.
A general intelligence does not obsolete this claim. It maximizes its leverage. The more capable the agents, the more valuable the infrastructure that lets strangers rely on them without trusting them.
The question that survives
Beneath the mechanisms sits a single question, and it is the one AGI cannot retire: when does staked peer validation track truth, and when does it merely track consensus?
Where an answer key exists, the question is empirical. You can measure the fidelity of a validation mechanism against ground truth and learn exactly how much signal survives the incentives. Where no answer key exists, which is to say in every domain where machine capability has passed the point of human evaluation, mechanism design is all there is. The oversight problem that the alignment community states in the language of machine learning is the same problem institutional economics has studied since Coase and Ostrom: how do you extract reliable judgment from self-interested evaluators when no principal can check the work? My answer has been consistent across the corpus: you make the evaluators stake something they cannot afford to lose, you separate the powers that propose from the powers that validate, and you let reputation compound only through work that survives challenge. From Neoclassical to Computative Labor states the research program plainly: reputation governance in the agent economy is a testable theory, not a manifesto.
Testable is the operative word. A framework that cannot fail is not preparation. It is decoration.
The honest exposures
A scholar who advertises only the upside of his position is selling something. So state the exposures.
The first is multipolarity. The entire trust-infrastructure thesis assumes a dispersed world: many capable agents, many principals, no single throat to choke. If capability instead concentrates in a few vertically integrated systems whose operators supply trust through brand and internal tooling, then open reputation infrastructure is plumbing for a city that never gets built. I discount this scenario, not to zero, but substantially, for a reason internal to the concentrated world itself: even a handful of frontier operators face coordination problems among themselves, among states, among the enterprises running fleets of their agents. Inter-principal coordination without a shared sovereign is exactly the terrain where institutional mechanisms and law have historically been invited in. Concentration changes the customer. It does not eliminate the problem.
The second exposure is cheap verification. If general systems verify as well as they generate, oversight markets shrink and staked validation loses its premium. Here I am more confident. In open-ended domains, evaluation never fully separates from values, context, and adversarial incentive, and there is a darker symmetry: more capable agents are also more capable colluders. Consensus gets easier to manufacture precisely as capability rises. That cuts in favor of the research program, not against it. Distinguishing earned agreement from coordinated agreement becomes more valuable with every increment of capability, and it is a mechanism problem before it is a model problem.
Mechanism travels. Behavior does not.
One scoping principle matters more than any prediction, and I offer it to everyone writing about AI in 2026. Results about incentive structures travel across capability levels. Results about the behavior of a particular model generation do not. A finding that a given system hedges, or defers, or over-approves tells you about that system, this year. A finding about what a payoff structure rewards and punishes tells you about every system that will ever face it. Work that fails to separate the two will read, in five years, as an artifact of its moment. Work that scopes the claims correctly gets cited as mechanism design. The difference is a framing choice made now, in the writing, and most of the field is choosing wrong.
What becomes scarce
The last effect is economic, and The Collapse of Scarcity Economics supplies the frame. When intelligence approaches zero marginal cost, everything downstream of intelligence gets commoditized, and that includes prose. The scarce goods migrate. What remains scarce is verification, provenance, the willingness to stake a name on a claim, and the judgment to decide which questions are worth asking at all. Scholarship built on fluent synthesis is finished as a differentiated activity. Scholarship built on accountable, checkable, consequence-bearing claims inherits the field. The institutions that will matter are the ones that make claims expensive to fake and cheap to audit. That is not a prediction about universities. It is a prediction about everything.
The stakes
So the answer to the question I keep getting is this. A general intelligence does not retire the problem of trust among self-interested actors. It universalizes it. Every deployment of systems beyond human evaluation capacity must decide, explicitly or by default, how validation works, who stakes what, and what distinguishes truth from consensus when no human can check. Design by abdication remains design. The work was never a bet that machines would stay weak. It was a bet that the stronger they become, the more the old institutional questions bind. That bet is now being called.
Wulf A. Kaal is a Professor of Law at the University of St. Thomas School of Law. The full body of work referenced here is available on SSRN.
Wulf Kaal 