The AI Compute Dilution Hypothesis, Part III
In March, I thought the binding constraint was computational. By August, I thought it was physical: infrastructure, energy, and the speed at which anything can actually be built. I now think both of those sit above something else. Every constraint in the first two parts gets resolved by someone deciding how to resolve it, and that is the layer about to change.
Part II ended on an assumption rather than a conclusion. Every analogy I used there, from textile overproduction to the diffusion of robotics to the water infrastructure of a Dalmatian island, works because the systems being described share one constant: the consequential decisions within them were made by human beings.
Humans decided where to invest. Boards decided strategy. Entrepreneurs decided what to develop. Consumers decided what to buy. Investors decided how much risk to accept. Governments decided what to regulate.
And humans do not decide through pure optimization. We use data, but we also use intuition, fear, ambition, ego, status, experience, loyalty, greed, hope, herd behavior, and emotion. Across centuries, the technologies changed enormously. The decision-making architecture did not.
This is the quiet reason historical analogy works at all. It is not that circumstances repeat. It is that the agent processing those circumstances has remained recognizable.
Artificial intelligence may break that continuity. I would state it as the Decision-Maker Discontinuity Hypothesis: historical patterns retain some predictive power because the behavioral architecture that produces consequential decisions has remained relatively stable. If AI becomes a dominant decision-making agent, that architecture changes, and historical patterns may no longer behave in the same way.
That is a larger claim than it first appears, and it deserves to be examined rather than asserted.
The Transfer Nobody Signs
The obvious objection is that none of this is happening. Boards still vote. Executives still sign. Regulators still hold named individuals accountable. No organization I know of has formally delegated strategic authority to a model, and few would survive the attempt.
I think that objection misses where the transfer actually occurs. Formal authority is the last thing to move, not the first. What moves earlier is the default position, and there are at least five mechanisms by which it moves without anyone deciding it should.
The first is default inversion. In an organization without analytical systems, a proposal requires effort to produce and effort to approve. Once a system generates the recommendation, the recommendation arrives already formed. Human action is then required to deviate rather than to comply. Nothing has been delegated. The direction of effort has simply reversed, and effort is one of the strongest predictors of what organizations actually do.
The second is the asymmetric burden of justification. Following a well-documented recommendation typically requires no written explanation. Overriding it usually does. Someone has to record why the analysis was rejected, and that record will be read later, in hindsight, by people who know the outcome. This asymmetry does not need to be designed. It emerges from ordinary governance practice, and it makes agreement structurally cheaper than disagreement.
The third is speed. Once systems produce analysis at machine cadence, human deliberation becomes the slowest step in the process. Organizations under competitive pressure do not usually respond by slowing the machine. They respond by compressing the review, and a compressed review converges toward approval, because approval is the only response that can be given quickly without appearing arbitrary.
The fourth is liability. In regulated and litigated environments, the defensible position tends to be the documented one. If a decision follows the output of a validated system and the outcome is poor, responsibility is distributed across a process. If a decision contradicts that output and the outcome is poor, responsibility concentrates on the individual who overrode it. Rational actors notice this quickly, and it does not require cynicism to act on it.
The fifth, and the one I consider most consequential, is control of the option set. Formal decision rights concern which option is selected. They say nothing about who determines which options appear. When a system produces the shortlist, the frame, the comparison criteria, and the base case, the human choice remains entirely real while operating inside boundaries the human did not set. This is not new; it is the oldest form of influence in any organization, exercised by anyone who has ever prepared a board pack. What is new is the scale and speed at which the framing can now be generated.
There is a sixth effect that follows from the first five rather than standing beside them. Judgment is a practiced capability. If the option to independently evaluate is available but rarely exercised, the capacity to exercise it erodes, and the erosion is invisible until the moment it is needed. Lisanne Bainbridge described this in 1983 as one of the ironies of automation: the more of a process you automate, the more critical the operator's residual judgment becomes, and the fewer opportunities that operator has to keep it sharp (Bainbridge, 1983). Her subject was process control and power generation. The pattern is not specific to AI. AI simply extends it from the control room into the boardroom.
None of these mechanisms involves a transfer of authority. Each is individually reasonable. Together they can move effective decision weight a long way while every formal structure remains exactly as it was.
Who Gets 51% of the Decision?
Imagine a corporate board evaluating a major investment ten years from now. Its systems have analyzed every relevant market, competitors, patents, technological trajectories, consumer behavior, regulations, energy, supply chains, macroeconomics, geopolitics, and the outcomes of comparable historical investments. The analysis recommends one strategy. Several directors disagree, and their disagreement is largely intuitive. Who wins?
Perhaps humans formally retain every vote. But suppose the board follows the recommendation in ninety-five percent of major decisions. Humans still legally control the company. Do they still control the decision?
The relevant variable may not be formal voting rights at all, but effective decision weight. Does the relationship remain 51% human and 49% machine, or does it eventually become 49% human and 51% machine? And if it does, at what point does anyone notice? The transition has no ceremony attached. AI simply becomes correct often enough that overriding it becomes progressively harder to justify to a board, a regulator, an insurer, or a shareholder.
The uncomfortable part is that this can happen even as every individual decision improves. A system can raise the average quality of outcomes and simultaneously relocate the point at which the outcome is actually determined.
The Robotaxi in Zagreb
I noticed a small illustration of the problem while watching robotaxi testing in Zagreb. When the autonomous vehicle needed to change lanes, it appeared to struggle to find an opportunity that satisfied its safety requirements. The human drivers around it behaved differently. They accepted smaller gaps, accelerated, closed spaces, signaled intention through vehicle movement, and sometimes simply accepted more risk. The vehicle was rationally searching for a sufficiently safe opening in a system populated by agents operating under an entirely different risk function, resulting in hesitation and delay.
The interesting question is not whether the robotaxi drives better or worse than the humans around it. On any reasonable safety metric, it may well drive better. The point is that, according to one decision architecture, rational behavior can become inefficient in a system dominated by another.
Now scale that observation. AI allocates capital while humans speculate. AI manages procurement while humans negotiate emotionally. AI recommends terminating an R&D program, while its founder refuses, believing the breakthrough is one experiment away. AI purchasing agents optimize price and utility, while humans continue to buy products that communicate identity rather than deliver function.
For some period, and it may be a long one, we will inhabit neither a human-governed economy nor a machine-governed one. We will inhabit a mixed-agency complex adaptive system in which two decision architectures with different risk functions, time horizons, and tolerance for ambiguity operate in the same market simultaneously. The friction in such a system is not a transitional inconvenience. It is a structural property of the mixture.
The Question a Machine Would Answer Differently
There is one decision where the difference between the two architectures becomes concrete rather than theoretical, and it is the decision most product and R&D leaders will face this decade: when to enter a disruptive technology.
Two patterns dominate the history of that decision, and both are familiar. The first is waiting. Let the technology move through adoption, let the diffusion curve resolve, enter once the uncertainty is priced. Wait too long, and there is nothing left to take. Kodak is the case everyone reaches for, though what Kodak lacked was never the technology. It held digital patents for years and could not activate them without destroying the margin that paid for everything else.
The second is jumping. Commit early, publicly, and at scale, before the infrastructure, the market, or the adoption rate justifies it. Several established carmakers did exactly this in the shift to electric vehicles, and did it fast, and are living with the consequences now.
What is worth noticing is that both patterns run on the same fuel. Waiting is driven by the fear of stranding an investment. Jumping is driven by the fear of becoming the next cautionary tale. Neither is analysis. Both are fear, pointed in opposite directions, and between them they account for a very large share of the strategic decisions recorded in business history.
Now remove the fuel. A system optimizing against stated objectives, constraints, probabilities, and available data does not experience either fear. It has no ego invested in the previous strategy, no status to defend in front of a board, no loss aversion, no instinct to move because a competitor moved. If that system decides when to enter, there is no particular reason for it to select either of the two patterns humans have been alternating between for a century. It may well select a third thing, and the reason history offers no guidance about what that third thing looks like is that history contains no agent that decided without fear.
This is the sharper form of the discontinuity, and it cuts against a common objection. One might argue that machines will simply herd faster, joining the same stampedes for the same reasons. That holds only if the machine is given human inputs to work with. Simulate ego, loss aversion, social pressure, and intuition, randomize them, and you can certainly reproduce human decision-making in silicon. But at that point, it is a computational model of a human decision, not a machine decision, and the interesting case has been ruled out.
The limit of this, and it is a real one, is that removing fear from the reasoning does not remove it from the objective. A system optimizes toward a goal, and that goal is written by people with careers, boards, and reputations. Ego is not eliminated. It relocates one step upstream, into the definition of what counts as a good outcome. That is the quiet reason the selection of the objective, discussed further below, matters more than any question about capability.
More Rational Agents Do Not Produce a More Rational System
There is a further paradox, and it is the one I find hardest to argue away.
Suppose machine decision-making becomes extraordinarily rational. Thousands of corporate systems analyze the same technological opportunity. They draw on overlapping data, share much of their underlying architecture, and reason toward similar conclusions. They identify the same superior investment. They all recommend entering.
Capital floods the sector. Competition explodes. Supply exceeds demand. Margins collapse. And the investment that was rational for every individual agent becomes unattractive precisely because every other rational agent reached the same conclusion.
We are back to the robotics example from Part II, with one difference: convergence happens much faster, and the diversity that previously slowed it is gone.
I should be careful here, because the contrast is easy to overstate. Human decision-makers converge too, and often violently. When a competitor moves first and appears to be creating value, most organizations follow within a year, driven less by analysis than by the fear of becoming the next cautionary tale. The result is already visible in most industries: a market filling with near-identical products that nobody quite chose. Herding is not a machine invention.
The difference is mechanism, speed, and floor. Human herding runs on status and fear, propagates through people who talk to each other, and stays noisy because the fear is unevenly distributed. Some are too proud to follow, some are too slow, some are protecting a business the trend would cannibalize. That noise had a systemic function nobody designed, and nobody valued: it worked as crude diversification across an entire economy. Machine convergence has no such floor. It arises from agreement rather than fear, propagates at the speed of an API call, and whatever dissent remains will reflect differences in data and configuration rather than differences in temperament.
This sits alongside the previous section rather than contradicting it, and the pairing is the uncomfortable part. Machines may well diverge from humans, reaching entry decisions no human pattern would have produced. There is no reason to expect them to diverge from each other. A population of agents that all decline the stampede and all identify the same better moment has still converged. It has simply converged somewhere new. Removing fear removes the noise it generates, and that noise was doing structural work nobody had accounted for.
Complex adaptive systems do not require irrational actors to produce unexpected outcomes. Individually rational behavior is entirely sufficient. AI may therefore remove some human biases while dramatically accelerating the feedback loops that produced the instability in the first place.
Where the Human Decision Should Remain
The argument so far runs in one direction, so it is worth stating the counterweight properly rather than as a concession.
There are categories of decision where human authority should remain decisive, and not because humans decide them better. In several of these cases humans clearly decide them worse. They should remain human because the decision has properties that make optimization the wrong frame.
The first category is irreversibility. Optimization operates on expected value, which is the correct instrument when outcomes are recoverable and the process repeats. It is a poor instrument when a single outcome ends the process. Closing a facility, exiting a market permanently, releasing something that cannot be recalled, and any decision whose downside is unrecoverable belong to a different class. The relevant question is not which option has the best expected outcome, but which failures the organization can survive.
The second is the selection of the objective, and it carries more weight than the others because everything above depends on it. Any system optimizes against an objective it is given. Choosing that objective, deciding what the organization is actually for, which tradeoffs are acceptable, and whose interests are counted, is a value judgment and not a computation. This is also where the fear I claimed had been removed comes back. The reasoning may be free of ego and loss aversion; the goal is still written by people who have careers, boards, and reputations to protect. Ego is not eliminated by better analysis. It relocates one step upstream, into the definition of what counts as a good outcome, where it is much harder to see. Delegating the optimization while believing you have delegated the strategy is the most likely governance failure of the next decade, and it will not look like a failure while it is happening.
The third is non-transferable accountability. Some obligations attach to persons rather than processes. A regulator sanctions a named individual. A court holds an officer responsible. A community holds a company answerable. Where accountability cannot be distributed, decision authority should not be distributed either, because a system that cannot be held responsible cannot be trusted with a decision for which someone must be.
The fourth is legitimacy, and the objection to including it is that acceptance is a communication problem rather than a question of who decided. I do not think that holds. Consent is usually withheld for a specific reason: the people asked to bear a cost do not believe the decider will bear any part of it. That is a claim about exposure, not about explanation, and it cannot be answered with better messaging. A system has nothing to lose, which is precisely why a decision attributed to one is harder to accept. And a decision nobody accepts is not a superior decision in any operational sense, whatever its quality on paper.
The fifth is genuine novelty. Systems trained on the past perform well when tomorrow resembles yesterday, which it usually does. The decisions that determine the fate of an organization are disproportionately drawn from the cases where it does not. This is the smallest category by volume and the largest by consequence.
What these five have in common is that none of them is a claim about capability. They are claims about the structure of the decision. That distinction matters, because capability arguments expire as systems improve. Structural arguments do not.
What Would Show This To Be Wrong
A hypothesis that cannot fail is not a hypothesis, so it is worth saying what would count against this one.
If effective decision weight is genuinely shifting, the mechanisms described above should leave traces. Override rates should decline over time in organizations that adopt these systems, and the reasons recorded for overriding should get shorter and more procedural. Review cycles should compress. Dissent inside firms running similar systems should start to look similar across firms. And when things go wrong, the post-mortems should show that the decision was effectively settled before it reached the people formally responsible for making it.
If instead override rates hold steady, if disagreement stays as varied between organizations as it is today, and if boards routinely reject well-supported recommendations without that being treated as a governance failure, then the mechanisms I have described are weaker than I think, and human judgment is more robust to procedural pressure than this argument assumes.
I would rather be wrong in that direction.
Conclusion: The Question That Actually Matters
Artificial intelligence may be technologically unprecedented while still operating inside economic and behavioral patterns humanity has lived through many times before. That was the argument of Part II, and I still hold it.
But it carries an expiry condition. The patterns held because the agent remained constant. Change the agent and the historical analogy weakens, not because history stops being instructive, but because the mechanism that made it instructive has been altered.
Perhaps the greatest disruption in the history of artificial intelligence will not arrive when a machine becomes capable of producing something better than we can. It will arrive when humans begin trusting machines to decide what should be produced, what should be developed, where resources should be invested, and which risks should be taken. At that point we are no longer using AI to optimize a complex adaptive system. AI has become one of the agents determining how the system evolves.
The useful response is not to resist the shift, which is neither realistic nor desirable given how many decisions it will improve. It is to be explicit about which decisions are being moved, and to notice that the mechanisms doing the moving are procedural rather than deliberate. Defaults, justification burdens, review speed, liability exposure, and control of the option set are all governable. They are governable, however, only by an organization that has looked at them, and most organizations are currently looking at capability instead.
Read together, the three parts look like three subjects and are really one question: who gets the next unit of a scarce resource. In March that resource looked computational, and the argument was a supposition. Five months later it was a description of what the industry was saying out loud, and the scarce thing had become physical: chips, power, plant, and the years it takes to build any of them. None of that is going to resolve quickly, which means the coming decade is largely a decade of allocation decisions. Who gets capacity, what it gets used for, how much, and when. Those are the same four questions whether the resource is GPU hours or the water supply of an island.
What changes is not the list of questions. It is who answers them, and this part has argued that the answer is already moving without anyone voting on it.
Compute, energy, infrastructure, attention, agency. The chain does not stop, and there is no reason to assume agency is the last station on it. What sits behind it is less comfortable. If every agent decides optimally on overlapping data, it becomes an open question what happens to the mechanism that has been coordinating those decisions all along, because exchange presupposes disagreement about value. For now the objectives are still written by people, and while they are, the disagreement does not disappear. It moves from assessment into preference. That is a subject for another article.
Humans may still hold 51% of the votes. The question that matters is whether we will still hold 51% of the decision, and whether we will be able to tell the difference in time to choose.
This article represents the author's personal observations and conceptual hypotheses. It is a conceptual argument rather than an empirical study, and the examples are used to explore recurring patterns in decision architecture and complex adaptive systems, not to establish causal relationships.
Concepts: Decision-Maker Discontinuity Hypothesis; Effective Decision Weight; Default Inversion; Asymmetric Burden of Justification; Option-Set Control; Mixed-Agency Systems; Complex Adaptive Systems.
References
Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775-779. https://doi.org/10.1016/0005-1098(83)90046-8
Other personal observations and conceptual models: Author.
Comments