The Machine Will Not Make Us Equal

Artificial general intelligence is being sold as a universal ladder: give everyone access to the same system and the distance between the gifted and the ordinary will collapse. The evidence available so far points to a harder truth. AI can lift people at the bottom on defined tasks, but it can also enlarge the rewards for workers who already know how to frame problems, detect errors and turn answers into decisions.

That distinction is moving from seminar rooms into public policy. Politicians are debating public ownership of AI companies, dividends from machine-generated wealth, free access through libraries and rules modelled on net neutrality. Yet most of these proposals address the price of the tool or the ownership of the profits. They do not answer the central question: what happens when the same machine becomes more valuable in the hands of a person who has better judgement?

The promise of equality begins with a misleading picture

The equalising story is attractive because it sounds both modern and familiar. A worker who cannot write a polished report can ask a language model to produce one. A junior programmer can request code that would once have required years of experience. A small business owner can obtain market research, legal drafting and a marketing plan without hiring a team of specialists. If the machine supplies the missing expertise, the argument goes, the expertise itself loses its scarcity.

That argument confuses access to an answer with possession of the ability to use one. A model can produce a plausible report, but it cannot decide whether the central premise is false. It can generate code, but it cannot guarantee that the code solves the right problem. It can propose a strategy, but it cannot bear the cost when the strategy fails. The human being remains responsible for selecting the question, defining the objective, judging the evidence and deciding when to stop.

Those responsibilities are not minor details left over after the machine has done the real work. They are the work. In routine settings, an AI assistant may turn a weak performer into a competent one by spreading the habits of more experienced colleagues. In open-ended settings, the value of the tool depends on whether the user can distinguish a useful possibility from a confident piece of nonsense. The person who lacks that distinction may receive more words, more options and more apparent certainty without receiving better judgement.

The word “equality” hides this problem. It can mean equal access, equal treatment, equal opportunity, equal capability or equal results. These are not the same objective. A public library that offers the same model to every visitor may achieve equality of access. It cannot guarantee equality of output. Nor should it pretend that equality of output is possible without restricting the people who can make especially productive use of the system.

The political temptation is to treat the machine as a neutral quantity that can be distributed like electricity or broadband. But AI is not merely a pipe carrying an identical commodity. It is a responsive instrument whose return changes with the quality of the person directing it. The more capable the user, the more ambitious the task that can be delegated, checked and recombined. That is why universal access may be necessary for fairness while remaining insufficient for equal outcomes.

What the early evidence actually shows

The first serious workplace studies do not support a simple story of either universal liberation or universal displacement. They show variation. In a study of 5,179 customer-support agents, researchers found that access to a generative AI assistant increased issues resolved per hour by 14 per cent on average. The improvement was much larger among novice and lower-skilled workers, while experienced and highly skilled workers saw little effect in that particular setting.¹

That result matters because it explains why the equaliser claim has survived contact with evidence. A tool that gives inexperienced workers access to the patterns used by their stronger colleagues can narrow a gap. The novice can be shown how to phrase a response, identify a common problem and follow a reliable sequence. The technology does not need to possess a complete theory of customer service; it needs to make useful practice easier to imitate.

But the same study does not establish that AI will equalise all work. Customer support is structured, repetitive and measurable. The objective is often clear: resolve the customer’s issue accurately and quickly. Many occupations contain tasks of that kind. Many others do not. A business leader deciding whether to enter a market, a scientist choosing which hypothesis to test, a lawyer judging the strength of an untested argument or a journalist assessing a source faces a problem in which the correct question is part of the challenge.

Other evidence points in the opposite direction. A study of taxi drivers using an AI system that predicted where demand would be strongest found that productivity gains accrued to lower-skilled drivers and narrowed the productivity gap by 14 per cent.² Here again, the system supplied a form of local knowledge that inexperienced workers could use. It reduced wasted cruising time without requiring the driver to invent a new business model.

These findings are not contradictory. They show that the distributional effect of AI depends on the task, the technology and the relationship between the two. A system that supplies missing information may help the weaker performer catch up. A system that multiplies the user’s ability to set goals, combine evidence and make decisions may reward the stronger performer. It is reckless to speak of “the effect of AI” as though there were one effect waiting to be discovered.

That warning becomes more important as systems move beyond assistance with bounded tasks. The strongest economic gains are unlikely to come from producing one more email or summarising one more meeting. They will come from coordinating many steps: finding a market, designing a product, testing it, revising it, selling it and responding to competitors. The user who can see the whole chain will gain more from an agent that can operate across it.

The multiplier favours the person who can steer it

There is a crucial difference between adding capacity and multiplying capacity. If every worker receives the same fixed improvement, the absolute gap between them may remain roughly stable. If the improvement depends on what the worker already knows, the gap can widen even while everyone becomes more productive.

Consider two employees given the same model. The first asks for a generic answer and accepts the result when it sounds fluent. The second supplies a precise objective, gives the system relevant constraints, tests its assumptions, asks for competing explanations and checks the result against primary evidence. The second employee has not merely used better wording. That employee has brought a better model of the work to the interaction.

Skill in this setting is broader than intelligence measured by a test. It includes domain knowledge, memory, judgement, patience and the ability to notice when a problem has been badly framed. It includes what psychologists call metacognition: the capacity to monitor one’s own thinking and correct it. It includes emotional discipline. A manager who cannot tolerate uncertainty may force an AI system to produce a premature answer. A researcher who wants a particular conclusion may use the model as a machine for confirming a preference. The system will often provide enough fluent material to conceal the mistake.

Experienced workers also know what a good result looks like before they receive it. That gives them a filter. They can reject a weak draft, identify a missing variable and ask for a revision that moves the work forward. A novice may not know that the answer is weak. The machine can therefore reduce the cost of ignorance in one moment while allowing ignorance to persist across the project.

This is the mechanism behind a possible cognitive ratchet. A capable user obtains a useful result, incorporates it into a larger plan, learns from the iteration and returns with a sharper request. Each cycle improves both the output and the user’s ability to extract output. The less capable user may also improve, but if the stronger worker is learning faster, the absolute distance grows.

One should not treat that outcome as a law of nature. Better interfaces, carefully designed training and systems that force users to show their assumptions could spread more of the advantage. Organisations can require source checks, stage decisions and provide feedback rather than simply handing workers a chatbot. But such measures would have to teach the human side of the partnership. Giving everyone the same login is the easy part. Giving everyone the same judgement is impossible.

The fashionable language of “democratising intelligence” is therefore too loose to be useful. AI may democratise access to assistance. It may not democratise the ability to decide what assistance is worth having.

Machines have always rewarded human differences

The historical comparison is not that every new machine increases inequality. History is more complicated than that. The stronger lesson is that machines change the value of the abilities people already possess. A power loom did not add an identical amount of productive force to every person in a factory. It rewarded workers who could coordinate movement, maintain equipment and organise production. The automobile extended the reach of people who could navigate, plan and maintain a vehicle. The spreadsheet rewarded those who knew which assumptions mattered.

Writing, mathematics and computing followed the same pattern. A person who can organise an argument uses a word processor more effectively than someone who cannot distinguish a central point from a collection of sentences. A person who understands a mathematical model gains more from a calculator than someone who merely receives a number. A search engine is more valuable to a researcher who can identify a credible source and recognise a misleading one.

None of this means that tools are irrelevant to the less skilled. The opposite is often true. Technologies can make a previously difficult task accessible and allow people to learn by doing. They can remove barriers that have little to do with the underlying purpose of a job. A navigation system can help an inexperienced driver find a destination. A spellchecker can prevent a mistake that says nothing about a person’s ability to reason. AI may perform the same service on a much larger scale.

The mistake is to assume that removing one barrier removes all barriers. If a machine makes drafting cheap, the scarce resource may become editorial judgement. If it makes coding cheap, product definition and system design may matter more. If it makes information abundant, attention and verification become more valuable. Scarcity moves; it does not disappear.

The arrival of a powerful general-purpose tool can also change the meaning of experience. In the past, a junior employee learned by performing routine tasks that built a foundation for harder work. If AI absorbs those tasks, the organisation may save money while depriving newcomers of the practice through which expertise was acquired. The result could be a thinner pipeline into senior roles. Firms may have more output in the short term and fewer people who understand the work without the machine in the long term.

This is where the workplace debate should be more precise. The question is not simply whether AI helps or harms workers. It is whether firms use it to transfer knowledge, to intensify supervision, to remove entry-level jobs, to reduce hours, to increase output or to capture the gains for shareholders. The same technical system can support a novice, deskill a job or make a highly skilled employee extraordinarily productive.

The machine does not decide the distribution on its own. Management does. Markets do. Ownership does. But the human differences that the machine acts upon remain part of the equation, and public policy cannot erase them by changing the label on the technology.

The new argument is about ownership, not cognitive rationing

Political proposals in the United States have begun to acknowledge that access to a powerful technology may not be enough. The most striking proposal came from Senator Bernie Sanders, who announced the American AI Sovereign Wealth Fund Act in June 2026. His plan would impose a one-time 50 per cent tax, paid in stock, on large AI companies and place the resulting ownership in a public fund. Sanders argues that the public should share in the wealth created from the accumulated knowledge and creative work of society.³

That is a proposal about ownership and proceeds. It says that if AI produces enormous returns, citizens should receive a financial stake rather than watch the gains accrue to a narrow group of founders and investors. Other versions of the idea would provide dividends, public investment or a national fund that holds equity in technology companies. The argument has also attracted interest from parts of the technology industry, which see public participation as a way to manage the political backlash created by rapid automation.

There is a sound reason for concentrating the debate there. A government can tax income, hold shares, distribute dividends or subsidise access without pretending that it can measure the exact amount of intelligence amplification each citizen deserves. Those instruments are imperfect, but they are familiar. Their rules can be debated in public. They do not require an official to determine whether one person’s questions are sufficiently sophisticated to justify a stronger model.

The alternative would be a licensing regime for cognitive power. A state might attempt to allocate more capable models according to income, occupation, educational achievement, identity or a supposed social need. It would need to authenticate users, monitor their use, prevent sharing and decide what counts as an unfair advantage. Open-weight models, local hardware and private networks would make enforcement difficult. The people who controlled the allocation system would acquire a power more intimate than the power to distribute money: the power to decide who may think with which machine.

That prospect should trouble anyone who takes equality seriously. A programme designed to prevent a new hierarchy could create a bureaucratic hierarchy of its own. It could also freeze people into categories. The student judged unlikely to make productive use of a frontier model might be precisely the person who would make an unexpected discovery. The small company denied advanced access might be the one that develops a valuable product. A system that claims to correct unequal outcomes could end by rationing opportunity.

Universal access is a better starting point. So are public libraries, schools, apprenticeships and training programmes that teach people to check and direct AI rather than merely consume its output. But even those measures should be described honestly. They can raise the floor. They cannot guarantee that the floor will meet the ceiling.

Redistribution can moderate the gap without abolishing it

The debate over AI wealth is likely to resemble the older debate over progressive taxation. Tax systems can reduce the distance between incomes after tax. They can fund education, health care and security against shocks. They can limit the speed at which wealth passes from one generation to the next. They cannot make the underlying production of value equal.

That distinction is often treated as a moral failure of taxation, when it is in part a mathematical fact. If one person uses capital, judgement and technology to create a much larger surplus, a tax can claim a portion of the surplus. It cannot make the original output identical without taking the entire surplus or restricting the activity that produced it. The more powerful the productive tool, the more pressure there will be to decide how much of the gain should remain with the creator and how much should be shared.

AI may intensify that pressure because its marginal cost can fall as its reach expands. A strong model can help one employee complete a task, but an agentic system could help a small team coordinate thousands of tasks. The owner of the system, the data, the computing infrastructure and the distribution channel may capture returns that are out of proportion to the number of people employed. The resulting wealth concentration would not be proof that every wealthy person made an illegitimate contribution. It would be proof that ownership matters when a technology scales.

There are also limits to the claim that taxes can halt inequality across generations. Wealth can be transferred through companies, trusts, gifts and assets whose value is difficult to measure. Skilled people and capital can move. Rules designed to capture future AI gains will produce avoidance strategies of their own. This does not make redistribution pointless. It makes it a continuing contest rather than a final solution.

The more defensible aim is to prevent economic security from depending on whether a person owns a slice of the machine economy. Public dividends, wage insurance, portable benefits, strong education and worker ownership could help. So could shorter working hours if productivity gains are shared as time rather than only as cash. Employee representation could influence whether AI is used to remove jobs, improve jobs or simply increase the pace of work.

None of these policies would eliminate differences in talent, ambition or judgement. That is not a reason to reject them. It is a reason to state their purpose accurately. Social policy can protect dignity and broaden the ownership of gains. It cannot promise that every worker will receive the same return from a tool whose value depends on how it is used.

The public argument becomes dishonest when it promises both extraordinary productivity and equal outcomes without explaining the trade-off. A society can choose to share more of the surplus. It cannot share what it has first prevented from being created.

Net neutrality offers a warning about grand promises

The recent history of net neutrality illustrates how a narrow technical rule can be asked to solve a much larger political problem. In 2015, the Federal Communications Commission reclassified broadband providers under Title II and adopted rules against blocking, throttling and paid prioritisation of lawful content. The order concerned the companies that carried data across the physical network. It did not control the platforms that hosted, ranked or removed speech.

The distinction was decisive. Internet service providers operated the pipes. Search engines, social networks and content platforms operated the layer above them. A rule preventing an ISP from slowing a lawful service did not require a social-media company to give every viewpoint equal prominence. It did not prevent a platform from suspending an account or removing a post under its own terms. The policy’s public language was broader than its legal mechanism.

The FCC’s own 2024 announcement, when it restored a national open-internet standard, described the rules in concrete terms: providers would again be prohibited from blocking, throttling or engaging in paid prioritisation of lawful content.⁴ That is a meaningful objective. It is not the same as guaranteeing viewpoint neutrality across the digital public square.

Yet many supporters and critics treated net neutrality as though it would settle the broader question of who controls speech online. When platforms later made decisions that inflamed political controversy, disappointment was directed at the neutrality regime even though the regime had never governed the relevant actors. A pipe could be neutral while the service delivered through it remained selective.

The lesson for AI is not that regulation is futile. It is that labels do not do the work of definitions. “Fair AI”, “equal AI” and “neutral AI” can refer to access, price, refusal rates, training data, political bias, safety rules or ownership. Each problem requires a different instrument. A law aimed at one cannot be judged by whether it solves all the others.

There is a second lesson. Public emotion tends to expand the mission of a policy. People who fear exclusion hear “equal access” and imagine equal influence. People who fear censorship hear “neutrality” and imagine a government compelled to provide every answer. The resulting expectations are almost impossible to satisfy because they were never tied to a precise mechanism.

AI policy will face the same danger. A public fund may redistribute financial returns while leaving unequal human-AI productivity untouched. A library may offer a premium model while leaving questions of accuracy and institutional power unresolved. A rule about the model provider may do nothing about the employers who decide how the tool changes work. The gap between the promise and the instrument will become tomorrow’s political grievance.

“AI neutrality” contains several different demands

It is possible to separate the idea of AI neutrality into at least four demands. Access neutrality would mean that users receive the same basic capability, rate limits and service quality regardless of wealth or political affiliation. Query neutrality would mean that a provider does not degrade or refuse lawful requests because of who is asking. Output neutrality would require the model to avoid systematic preference for one political or cultural view. Provider neutrality would limit the ability of a company to use a frontier system as an instrument of institutional power.

Only the first demand resembles the classic net-neutrality rule. It can be approached through public licences, school and library access, open models and competition among providers. Even here, “the same model” does not mean the same capability. One user may have better data, more computing time, a private toolchain and a team of specialists. Formal equal access can coexist with large practical differences.

Query neutrality is more difficult because legitimate safeguards often require distinctions. A system should not assist with certain forms of violence, fraud or privacy invasion. If a government prohibits a provider from refusing any lawful query, it must decide whether the legality of the query is the only relevant standard. If a provider may refuse on safety grounds, someone must define the boundary. The bright line disappears.

Output neutrality is harder still. A model does not simply transmit a packet whose contents were fixed elsewhere. It generates an answer from training data, system instructions, safety rules, commercial incentives and statistical patterns. Two answers can both be factually defensible while giving different emphasis to history, risk or responsibility. A regulator asked to decide whether the difference is bias would be entering the argument rather than standing outside it.

The European Union’s AI Act demonstrates the direction taken by one major regulatory system. Its framework is built around risk, with obligations attached to systems according to their potential effects on health, safety and fundamental rights rather than a general promise that every output will be neutral.⁵ The approach has its own difficulties, but it recognises that not every use of AI presents the same risk and that regulation must attach to a use case.

A broad neutrality mandate could also strengthen the gatekeepers it claims to restrain. Large companies would be able to afford compliance teams, lawyers and testing laboratories. Smaller developers might be pushed out by the cost of proving that their systems had not displayed an impermissible pattern. The government would acquire a continuing supervisory role over the production of language and knowledge. A rule intended to reduce private discretion might transfer discretion to public officials.

The more practical demand is transparency. Providers can disclose model limitations, explain the categories of safety intervention, publish evaluation methods and allow independent testing. Employers can tell workers when AI is being used to assess or direct them. Users can be given a meaningful route to challenge an automated decision. These measures do not produce a neutral machine. They make the machine’s choices less mysterious.

The workplace will be shaped by decisions made before the model arrives

The central conflict will not be settled by model architecture. It will be settled in offices, factories, hospitals and public agencies where someone decides what the technology is for. A company can use AI to give an inexperienced employee a reliable starting point, then invest in training and preserve time for judgement. It can also use the same system to increase targets, reduce headcount and turn every worker into a supervisor of an unreliable automated process.

Ownership determines who has the first claim on the gains. Governance determines whether workers have a voice in the use. Training determines whether AI spreads expertise or merely hides its absence. Competition determines whether customers receive lower prices or whether a small number of firms retain the surplus. These are old economic questions made sharper by a technology that can operate across more kinds of work.

The public debate should therefore stop asking whether AI will make people equal. It should ask which inequalities are tolerable, which are dangerous and which can be reduced without suppressing useful activity. It should distinguish a worker who earns more because a tool allows better service from a platform that earns more because it controls the only route to customers. It should distinguish an employer who removes pointless paperwork from one who removes the training ground on which a profession depends.

There is a case for making advanced tools broadly available. There is a case for public participation in the wealth they generate. There is a case for rules that prevent arbitrary discrimination, undisclosed surveillance and automated decisions that cannot be challenged. But these arguments grow weaker when they are wrapped in a fantasy of equal results.

A society that gives everyone a calculator does not make everyone a mathematician. A society that gives everyone a search engine does not make everyone a historian. A society that gives everyone an AI assistant will not make everyone equally able to frame a problem, judge an answer or build a company around an insight. That is not an indictment of access. It is the condition that access must confront.

The realistic ambition is not to abolish the human differences that advanced tools magnify. It is to ensure that those differences do not determine whether people can live secure and dignified lives. That requires sharing more of the surplus, protecting the ability to learn, keeping markets open and refusing to let either corporations or regulators turn “fairness” into a licence for unaccountable control.

The question regulators cannot avoid

Every proposal for AI equality eventually reaches the same point. If the goal is equal access, the policy can be designed and tested. If the goal is a share of the wealth, ownership, taxation and dividends provide instruments, however imperfect. If the goal is protection from discrimination, specific rules can identify prohibited conduct and create a process for appeal.

If the goal is equal productive power, however, the policy must confront the human partner. It must decide whether to limit the most capable user, subsidise the least capable, or accept that equal tools will produce unequal results. The first option invites rationing. The second demands serious education and institutional reform. The third requires political honesty.

The evidence does not justify declaring that AI will inevitably create a permanent cognitive aristocracy. Some systems are already narrowing gaps on structured tasks. Nor does it justify the opposite promise that general access will dissolve hierarchy. The effects are conditional, and the conditions are becoming more important as AI moves from drafting assistance to broad coordination.

What can be said with confidence is narrower and more useful. AI is a multiplier, not a magic equaliser. It can transfer good practice to people who lack experience, but it can also give exceptional users a larger field on which to exercise exceptional judgement. Public policy can spread access and distribute wealth. It cannot distribute judgement by decree.

References

1. Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, “Generative AI at Work,” National Bureau of Economic Research Working Paper 31161. ↩
2. Kyogo Kanazawa, Daiji Kawaguchi, Hitoshi Shigeoka and Yasutora Watanabe, “AI, Skill, and Productivity: The Case of Taxi Drivers,” National Bureau of Economic Research Working Paper 30612. ↩
3. Bernie Sanders, “The Public Should Own Half of the Big A.I. Companies,” U.S. Senate, June 1, 2026. ↩
4. Federal Communications Commission, “FCC Restores Net Neutrality,” April 25, 2024. ↩
5. European Union, Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence. ↩