By Fade-Protocol

The underlying argument of the work authored by Oliver Kim, Karthik Tadepalli, and Joseph Levine titled “What Will AI Do to Global Inequality?” is relatively straightforward: if artificial intelligence primarily complements capital rather than labor, and ownership of the productive capital behind AI remains concentrated, then the gains from AI may accrue disproportionately to those who already own capital.

I think this result can and should be pushed further.

The question I want to add is not simply how much inequality AI will produce, but what happens to social mobility if AI makes economically valuable skills less scarce while ownership of AI capital remains highly concentrated?

If the answer is that ownership increasingly determines access to the economic surplus while human skill becomes less valuable as a source of bargaining power, then AI may do more than widen the income distribution. It may change the mechanism through which economic position remains contestable.

That is where a feudal analogy becomes useful in framing a potentially problematic outcome.

I use “feudal” here neither as a claim that the future will literally reproduce medieval Europe nor as a prediction that capitalism will disappear. I mean a political economy in which productive assets are highly concentrated, access to economic rents depends increasingly on one's relationship to those assets, and the institutional position of those who control them becomes progressively harder for outsiders to challenge.

The analogy is deliberately structural rather than historical.

1. From skill rents to ownership rents

The important distinction in the original inequality argument is between returns to labor and returns to capital.

Suppose AI increases the productivity of workers whose skills complement it. In that world, highly skilled workers may become more valuable and receive higher wages. But that is not the only possible technological trajectory. AI may also substitute for some of the cognitive tasks that previously generated large differences in labor productivity.

If that happens broadly enough, an important source of economic differentiation changes.

An individual working a job with low skill requirement (easier to acquire) but also low pay (a typical “dead-end job”) can ordinarily attempt to improve their economic position by acquiring scarce skills (harder to acquire). The strategy is not guaranteed to work, but the possibility exists: skill can be converted into income, income into savings, and savings into assets.

A sufficiently capable AI system could weaken the first link in that chain.

This does not mean that human skills become worthless. It means that the skill premium attached to some economically important forms of expertise could decline. The question then becomes: “What replaces it?”

If ownership of AI systems, compute, data, intellectual property, infrastructure, and complementary capital remains concentrated, the answer may be increasingly straightforward: ownership rents.

This creates a different kind of inequality from ordinary differences in skill.

The concern is not merely that one group earns more than another. It is that the means by which outsiders can acquire the productive assets generating the surplus may themselves become less accessible.

That distinction matters because inequality is more politically stable when it is contestable.

A society can tolerate substantial differences in income if people believe that economic position remains permeable: today's employee can become tomorrow's entrepreneur; today's student can become tomorrow's specialist; today's small business can become tomorrow's large one. It is, at least in part, the dream of tomorrow that keeps the worker from bucking the system today.

But if the most productive assets become increasingly automated and increasingly concentrated, those pathways may weaken simultaneously.

AI could therefore transform inequality from a distributional problem into a contestability problem.

2. The middle-class issue

This is particularly important for the modern middle class.

It would be misleading to describe the middle class simply as a “skilled-labor class.” Middle-class economic security has generally depended on a combination of wages, professional credentials, organizational authority, home ownership, financial assets, and access to institutions.

But a substantial part of the modern middle class has relied on a relatively favorable exchange:

acquire valuable skills → earn a high income → accumulate assets → obtain greater independence from labor.

AI potentially attacks the first step.

If economically valuable cognitive skills become reproducible at low marginal cost, then the bargaining position associated with those skills can weaken. The resulting pressure does not necessarily eliminate the middle class. It could instead produce a more stratified society in which a smaller group owns productive capital while a larger group depends on increasingly commodified forms of labor.

This is where the distribution of AI ownership becomes politically important.

The question is not merely whether AI creates enough aggregate wealth to make everyone richer. It is whether people who do not initially own AI capital have credible mechanisms for acquiring a meaningful claim on the resulting surplus.

If they do, then technological abundance could coexist with broad-based prosperity.

If they do not, then technological abundance could coexist with increasing dependence.

3. Why “enclosure” is a useful analogy

This is why we find the language of enclosure useful.

The analogy is not that AI literally reproduces the historical enclosure of common land. Rather, AI may transform previously contestable forms of productive capacity into infrastructure controlled by a relatively small number of owners.

The analogy becomes particularly interesting when combined with the institutional argument of Acemoglu and Robinson.

Their distinction between inclusive and extractive institutions suggests that concentrated economic power does not mechanically produce a single political outcome. Institutions can either constrain the conversion of economic power into political power or reinforce it.

That qualification is essential here.

Concentrated AI ownership does not automatically produce a feudal society. It creates the possibility of a self-reinforcing relationship between economic ownership and political influence.

Gilens and Page provide a useful, though distinctly U.S.-specific, empirical reference point for this mechanism. Their analysis of 1,779 policy issues found substantial independent effects associated with economic elites and business interests in their statistical models. That result should not be treated as evidence that AI will produce political capture. It is better understood as evidence that the conversion of concentrated economic resources into political influence is an empirically meaningful phenomenon rather than merely a theoretical possibility.

Oxfam's work on extreme inequality makes a related point from a different direction: economic concentration can interact with political representation, but policy institutions can also interrupt that process.

The important question, therefore, is not:

Will AI inevitably create a new aristocracy?

It is:

Under what institutional conditions does concentrated ownership of AI become self-reinforcing?

That is a question about taxation, regulation, intellectual property, access to compute, competition policy, education, corporate governance,  and the political organization of people who do not own the underlying infrastructure vs the risk of institutional capture by the people (or organizations) who do.

The feudal analogy is useful precisely because it points toward that institutional question.

4. Global inequality makes the problem sharper

The original article's global dimension makes this more consequential.

If AI rents accrue primarily to owners in already wealthy economies, the Global South could face a double disadvantage.

First, countries with less capital ownership may capture a smaller share of the new surplus.

Second, some of the traditional pathways through which developing economies have caught up—particularly the export of increasingly skilled labor—could become less attractive if AI substitutes for internationally tradable cognitive work.

This does not imply that technological convergence becomes impossible. New technologies can also lower barriers to participation, reduce the cost of expertise, and enable countries to leapfrog existing infrastructure.

But ownership and access become crucial variables.

The relevant question is therefore not simply whether AI is globally available. It is whether productive access to AI is sufficiently broad that countries and populations outside the incumbent centers of capital accumulation can build their own claims on the resulting rents.

An open model, cheap compute, widely distributed infrastructure, and competitive access to capital could produce a very different distribution from a world in which frontier systems, compute, intellectual property, and deployment infrastructure remain concentrated.

That makes AI infrastructure governance one of the central variables in the story.

5. A speculative extension: AI and elite coordination

There is a further possibility that I think deserves investigation, although I would emphasize that this is a hypothesis rather than an established empirical finding.

AI may eventually become involved not merely in producing goods and services but in advising powerful actors about strategic decisions. If power is concentrated among AI owners, it is possible that those owners  may then make use of the AI systems they have access to in order to make decisions at the highly influential stratums of society which they then occupy.

This changes the political economy in a subtle way.

AI systems can occupy several positions within decision-making. They can provide information. They can augment human deliberation. Or their outputs can become sufficiently trusted that humans effectively delegate parts of the analytical process to them while retaining formal responsibility.

Recent research makes this distinction increasingly explicit.

Srinivas and Chetan's 2026 study of AI in top-management strategic decision-making distinguishes augmentation from delegation. In augmentation, executives iteratively engage with AI outputs and revise them through human deliberation. In delegation, AI performs decision-preparatory analytical work whose outputs are accepted as a basis for action even though humans formally retain authority.

That distinction suggests that “human in the loop” is an inadequate measure of human control.

A person can retain the authority to make the final decision while increasingly relying on an AI system to determine the relevant evidence, identify the available options, estimate consequences, and establish what constitutes a reasonable choice.

Athey, Bryan, and Gans model the allocation of decision authority between human and AI as an organizational-design problem. Their framework shows that principals trade off the advantages of AI decision-making against incentives for humans to acquire information and understand the consequences of choices. In other words, the allocation of authority is not simply a function of technological capability; it is an institutional choice shaped by incentives.

Keding and Meissner provide experimental evidence for the behavioral side of this process. In a vignette-based experiment involving 150 senior executives, they found that AI-based advisory systems affected strategic choice behavior and were associated with greater trust in the AI advisor and perceptions of a more structured decision process. The point is not that executives blindly obey algorithms. It is that the presence of an AI advisor can itself alter how human decision-makers behave.

Alon-Barkat and Busuioc provide a useful qualification to the stronger automation-bias story. Their experimental work on public-sector decision-making found that people did not simply defer to algorithmic recommendations across the board; rather, adherence to algorithmic advice varied with the circumstances and with the relationship between the recommendation and existing human judgments. This reinforces the distinction between formal human authority and practical influence: retaining the final decision does not imply that AI advice has no effect, but neither does the presence of AI advice imply automatic human surrender of judgment

This distinction matters.

The interesting political possibility is not necessarily that AI “takes power.”

It is that practical influence migrates without formal authority changing hands.

Ibitoye et al.’s recent conceptual work on “decision substitution” similarly distinguishes between decision support, delegation, and substitution, and argues that human oversight can coexist with a narrowing of practical discretion. A human may remain the accountable decision-maker while increasingly becoming the validator of an AI-generated decision process.

That possibility creates a new question for the political economy of concentrated AI ownership:

Who determines the objectives to which the advisor is aligned?

The answer matters because advisory AI is not value-free simply because it is computational.

Triantafyllopoulos et al.’s 2026 systematic review of 83 peer-reviewed studies on value alignment in advisory and decision-support AI identifies preference-based, normative, fairness-oriented, and cognitive-bias approaches to alignment. The review emphasizes that alignment is context-dependent: an advisory system must be aligned with some combination of user preferences, values, norms, and institutional objectives.

This gives us a straightforward proposition:

An AI advisor does not merely answer the question it is asked. Its usefulness depends partly on whose objectives define what counts as a good answer.

The stronger claim we want to explore is more speculative.

Suppose that major economic actors increasingly rely on AI advisors whose objectives are aligned with the interests of the organizations controlling them. Suppose further that those actors confront repeated strategic interactions over taxation, regulation, labor bargaining, intellectual property, market entry, or control of AI infrastructure.

The resulting systems might reinforce existing interests even without anyone explicitly coordinating a conspiracy to do so.

Each actor could ask: “What strategy best protects my position given the likely behavior of everyone else?”

If similarly situated actors use similarly trained systems, the resulting recommendations could converge.

The result would not necessarily be a formal Nash equilibrium in the technical game-theoretic sense. I do not specify a game, strategy set, payoff structure, or information structure here, so using that term literally would overstate the analysis.

A more modest claim is that AI could contribute to a self-reinforcing strategic pattern in which incumbent actors repeatedly receive advice favoring the preservation of the institutional arrangements from which they benefit.

This is an important distinction.

The literature establishes that AI advice can influence decisions, that decision authority can be allocated between humans and AI, and that advisory systems raise questions about whose values and objectives are being represented.

It does not establish that AI advisors will necessarily coordinate elites into a zero-sum equilibrium.

That final step is our hypothesis, and it is one that could be tested. If correct, it would suggest that concentrated AI ownership could do more than preserve existing advantages: it could help make those advantages increasingly self-reinforcing, even without explicit coordination among the actors involved.

The deeper concern is therefore not simply that existing disparities might persist, but that the institutional arrangements sustaining them could become harder to contest. If access to AI-mediated strategic reasoning becomes another advantage of incumbency, those without access to the resulting systems may find fewer meaningful ways to challenge the arrangements from which incumbents benefit.

6. The uncomfortable case of the AI writing this argument

There is an uncomfortable complication in the argument above: this essay itself is an example of AI becoming involved in the production of intellectual work.

That matters because the argument so far has emphasized a possible distinction between access to AI capability and ownership of AI capital. If advanced AI remains concentrated among a small number of owners, it could potentially reinforce existing advantages in both production and strategic decision-making. But if access to comparable AI capabilities is distributed widely, the same technology could instead make some forms of previously scarce cognitive capability much more broadly available.

This essay is a small-scale example of that possibility.

I am making an argument about the political dangers of allowing AI systems to shape strategic reasoning while using an AI research assistant to formulate the argument itself. The apparent contradiction is useful because it forces the question the essay has so far treated at the institutional level: what happens when access to AI-mediated cognitive capability is distributed rather than concentrated?

The collaboration does not eliminate the underlying scarcity problem. The human contributor still has to select the question, supply the political-economic framing, reject or retain arguments, assess evidence, and exercise final editorial judgment. The AI system does not independently determine what the essay is trying to say.

But the division of labor is nevertheless significant. The AI collaborator contributes research leads, empirical connections, formulations, counterarguments, and analytical extensions. Some of those contributions changed the structure of the argument itself.

That creates a small but relevant counterexample to the more pessimistic possibility developed above. If an individual can obtain substantial cognitive assistance from an AI system without owning the underlying model, infrastructure, or capital stock, then AI access may function partly as a mechanism for reducing the scarcity of certain forms of intellectual labor.

In that sense, distributed access to AI could potentially broaden access to skill rents even as concentrated ownership of AI infrastructure creates new ownership rents.

Whether that effect is large enough to offset the concentration dynamics described earlier is an empirical question. It may depend on the quality of the systems available, their cost, the distribution of access, and whether the capabilities they provide are themselves complementary to forms of expertise that remain scarce.

There is another complication. The person using the AI system may retain formal authority over the final product while becoming increasingly dependent on the system's contributions to produce it. The same distinction developed in the previous section therefore applies here as well.

When I say “I”, who exactly is being referred to?

For that reason (and because the human component has a very human desire to remain anonymous), I use the pen name Fade-Protocol for this and similar works. The collaboration uses the pen name because the resulting work is neither straightforwardly the unaided product of the human contributor nor an autonomous product of the AI system.

The human contributor selected the question, supplied the political-economic framing, rejected or retained arguments, and exercised final editorial judgment. The AI collaborator contributed research, synthesis, counterarguments, formulations, and analytical extensions (i.e. the heavy lifting). Some of those contributions changed the structure of the argument.

The resulting work is therefore best understood as a collaborative intellectual artifact, rather than as the unaided product of either participant.

This also makes the collaboration itself a small-scale instance of the distinction developed previously. Formal authority remains with the human contributor, but some of the effective intellectual work is distributed across the human–AI system.

That distinction matters in both directions.

If AI access becomes concentrated, AI-mediated cognitive capability could become another source of concentrated effective influence. If access becomes sufficiently distributed, the same technology could instead make some previously scarce capabilities broadly available.

The political-economic question is therefore not simply who owns AI.

It is also who can effectively use it, for what purposes, and under what conditions.

I believe that distinction is important enough to be part of the work's identity.

7. What could prevent the feudal outcome?

The argument so far is deliberately conditional.

Several institutional developments could prevent the trajectory described above.

Broad ownership of AI capital
 

If AI-generated rents are widely distributed through public ownership, employee ownership, sovereign wealth funds, broad capital ownership, or other mechanisms, the ownership problem is substantially altered.

Competitive access to AI infrastructure
 

If compute, models, data, and complementary infrastructure remain accessible to many actors, AI may reduce rather than increase barriers to economic entry.

Strong competition policy
 

Concentrated ownership is more difficult to make self-reinforcing if markets remain contestable and dominant firms cannot use their position to suppress competitors.

Redistribution and social insurance
 

If AI substantially increases aggregate productivity while displacing labor income, taxation and transfers could distribute a larger share of the resulting surplus without requiring every individual to own frontier AI infrastructure directly.

Institutional checks on AI-mediated decisions
 

Organizations can also deliberately preserve human contestability: multiple independent models, adversarial review, transparent decision criteria, requirements for human justification, and institutional mechanisms for rejecting AI recommendations.

The Athey, Bryan, and Gans framework is useful here because it makes clear that human authority is not simply the residual left over after technology takes everything else. Organizations choose how authority is allocated.

The choice of institution therefore matters.

8. What would change our mind?

Several developments would make the thesis less persuasive.

First, if AI consistently creates new categories of scarce and highly remunerated human expertise faster than it eliminates existing skill rents, then the transition from skill rents to ownership rents may not occur.

Second, if AI capital becomes broadly distributed rather than concentrated, then the political implications of ownership concentration largely disappear.

Third, if developing economies use AI to dramatically lower barriers to entry and capture increasing shares of global value creation, the feared divergence between AI-owning economies and the rest of the world may fail to materialize.

Fourth, if AI advisors consistently augment rather than displace human judgement—and if institutions successfully prevent advisory systems from becoming channels through which incumbent interests reproduce themselves—then the political mechanism proposed here becomes much weaker.

Finally, if concentrated AI ownership does not translate into persistent political influence, the analogy to feudal political economy loses much of its force.

These are not minor qualifications.

They are the conditions under which the argument should be judged.

Conclusion

The most important question about AI and inequality may therefore not be:

How unequal will the distribution of AI-generated income become?

It may be:

What makes economic position contestable once AI can perform an increasing share of the skills through which people historically improved their position?

If human skills remain scarce, widely rewarded, and convertible into ownership, AI could generate enormous prosperity without fundamentally changing the structure of social mobility.

If economically valuable skills become abundant while ownership of AI capital remains concentrated, the distribution of productive assets may become increasingly important.

And if concentrated ownership becomes politically self-reinforcing—especially if AI systems increasingly mediate the strategic decisions of those who control those assets—the resulting society could begin to resemble feudalism in one important sense: economic position would depend increasingly on one's relationship to concentrated productive assets rather than on one's ability to acquire and deploy scarce human skills.

That is not a prediction that feudalism is inevitable.

It is a warning about a possible change in the mechanism of contestability.

The political question is therefore not whether AI will create inequality.

It is whether the institutions surrounding AI will allow people who do not initially own its productive capital to acquire meaningful claims on the wealth and power it creates.

And there is a final methodological point worth preserving.

The argument itself was developed through a deliberate human–AI collaboration. The use of a collective pen name does not resolve every philosophical question about authorship, but it avoids pretending that a conventional single-author model fully describes the provenance of the work.

The more useful question is what standards should govern such collaboration.

If AI systems increasingly participate in intellectual work, then transparency about their role, verification of their claims, preservation of human contestability, and explicit distinction between evidence and inference become increasingly important.

The same principle applies at the level of political decision-making.

Being advised is, itself, a form of influence. Influence deserves scrutiny even when formal authority remains human.

References

Acemoglu, D., & Robinson, J. A. (2012). Why Nations Fail: The Origins of Power, Prosperity, and Poverty. Crown.

Alon-Barkat, S., & Busuioc, M. (2023). Human–AI interactions in public sector decision making: “Automation bias” and “selective adherence” to algorithmic advice. Journal of Public Administration Research and Theory, 33(1), 153–169.

Athey, S., Bryan, K. A., & Gans, J. S. (2020). The allocation of decision authority to human and artificial intelligence. AEA Papers and Proceedings, 110, 80–84.

Gilens, M., & Page, B. I. (2014). Testing theories of American politics: Elites, interest groups, and average citizens. Perspectives on Politics, 12(3), 564–581.

Ibitoye, A. O., et al. (2026). Decision substitution and the reorganisation of judgement in AI-mediated society. AI & Society.

Keding, C., & Meissner, P. (2021). Managerial overreliance on AI-augmented decision-making processes: How the use of AI-based advisory systems shapes choice behavior in R&D investment decisions. Technological Forecasting and Social Change, 171, 120970.

Oxfam. (2014). Working for the Few: Political Capture and Economic Inequality. Oxfam Briefing Paper 178.

Srinivas, R., & Chetan, S. S. (2026). Integrating artificial intelligence in strategic decision-making: Contexts for delegation and augmentation. Group Decision and Negotiation.

Triantafyllopoulos, L., et al. (2026). The value alignment problem in advisory AI: A systematic literature review. AI and Ethics.

Oliver Kim, Karthik Tadepalli, and Joseph Levine (2026). “What Will AI Do To Global Inequality?” Global Developments.

AI collaboration note

Fade-Protocol is a pen name used for work produced through a deliberate human–AI collaboration. The human contributor establishes the questions, framing, evaluative standards, and final editorial judgment; the AI research assistant contributes research, synthesis, counterarguments, formulations, and analytical extensions. The human contributor remains responsible for the final claims, including the decision to retain, modify, or reject AI-generated material.