A Thought Experiment: What Should We Learn If AI Can Do Everything?

Over the past year, AI has repeatedly appeared at the frontier of human capability. It is involved in long-standing mathematical problems and in finding previously unknown zero-day vulnerabilities.

It no longer merely answers existing questions. It is beginning to touch problems that once only the leading experts could advance.

Consider an extreme thought experiment.

Suppose AI eventually surpasses humans in every cognitive task. Even a person who spends a lifetime in one field cannot catch up with it in knowledge, understanding, creativity, or judgment. Would we still need to spend more than a decade in education?

Education would no longer be necessary for production and social operation

As AI capabilities have advanced, we have repeatedly answered “Why do I still matter?” by pointing to what AI cannot yet do.

At first it could only hold simple conversations and write short functions. We treated it as an interesting but unreliable toy. Later it began writing articles, producing images, and building engineering-scale programs. We said it did not write as well as a person, lacked lived experience, and could not handle truly complex problems.

When AI began producing usable reports and code, the human role retreated to review: it could do the work, but a person still had to catch mistakes and take responsibility. As AI began checking facts, generating tests, and looking for counterexamples, the human role moved further upstream: execution and checking might be delegated, but people still had to define the problem and decide what was worth solving.

In this extreme thought experiment, even that final refuge is gone. AI can identify important problems earlier than we can, assess their value more accurately, and see the long-term consequences of different choices more fully.

Those answers describe temporary gaps in AI capability, not a permanent source of human value. Each advance moves our answer to “Why do I matter?” another step back. When the gaps disappear, the search for an exclusively human capability reaches its end.

Education has long been treated as socially necessary because production depends on human knowledge and skill, and people use those abilities to earn a living. Education prepares producers for society and gives individuals a path to participation and reward. “What kind of people does society need?” and “What can a person live on?” have largely overlapped.

If AI surpassed us across the board, that overlap would break. Society would no longer need education to obtain human cognitive capability. Mastering such capabilities would not automatically earn a person a return because production needed them. If AI could understand needs, discover problems, judge value, predict consequences, correct errors, and execute, society would no longer have to spend a decade or more training people to do those things.

If education means cultivating human capabilities necessary for production and social operation, it would lose that necessity.

From social necessity to social choice

If society no longer depended on human capability to function, education would become an active choice. At heart, it would be a political and ethical question:

Even if AI does things better, what place do we want people to have in society? Which real abilities, rights, and institutional powers are we willing to preserve for them?

At least three internally coherent positions are possible.

Prioritize outcomes

Society seeks correct, useful, inexpensive results. If a human–AI system delivers them, whether an individual understands or can work independently no longer decides their value.

Education would not vanish, but it could shrink dramatically. Most people might need only basic training for everyday life, using systems, and following social rules. Traditional schools, examinations, and degrees might give way to real-time capability checks, tool certifications, and audits of results.

Learning would no longer be necessary to participate in production or sustain a livelihood. How well someone learned would not decide whether they could find work, receive a return, or enjoy basic security. The instrumental value of learning as a means of survival would approach zero.

Paradoxically, this could bring learning closer to what people imagine as interest-driven: not for exams, work, competition, or credentials, but because someone wants to understand something.

In a society that prioritizes outcomes, education might retreat without learning disappearing. “What should I learn?” would lose its universal answer; “What do I want to learn?” would remain.

Prioritize human participation

Society could instead decide that even when AI offers better answers, people should retain the ability to understand the world, question conclusions, reject arrangements, and change the rules.

Education would no longer serve a human advantage in production. It would preserve a social status: people would remain participants in collective life rather than merely consumers of a system’s outputs.

But participation could become a ritual.

If AI proposes a plan, supplies the argument, compares alternatives, assesses risks, and recommends the best decision, while a person only clicks “agree,” the decision does not really belong to that person. The human supplies formal approval, legitimacy, and someone to hold responsible for the system’s choice.

Meaningful participation can be expressed as a simple relationship:

Effective participation = capability × institutional power × counterfactual impact

People need to understand and question; they also need rights to refuse, revise, exit, and appeal. A different human choice must be able to change the result. If any factor is zero, participation becomes hollow.

Education can develop capability but cannot grant institutional power by itself. A society that chooses to preserve human participation must preserve real places where that capability can matter.

Prioritize social governance

A third society might focus first on whether complex human–AI systems remain controllable and whether power, benefits, and responsibility have clear owners.

It need not demand that every person judge better than AI or promise everyone a say in every decision. It would require someone at critical points who understands the system’s limits, exercises defined authority, handles exceptions, and accepts responsibility for consequences.

Education would move toward preparation for roles and certification of qualifications. A person would not study to prove they could do every task without AI, but to show that they could fulfill a particular social role under specified tools, processes, and supervision.

What society evaluates would change too: neither an isolated person nor a final output alone, but a human–AI system operating in a particular role.

The three positions lead to three different forms of education:

• When outcomes come first, education provides a minimum of social adaptation.
• When human participation comes first, it preserves people’s capacity to act as participants.
• When governance comes first, it prepares roles with authority and responsibility.

All three positions are coherent, but the desires behind them conflict.

We want AI’s best results without surrendering our right to participate, question, and refuse. We want systems to make judgments for us while still expecting someone to explain and answer for mistakes. We want the efficiency of outcome-first thinking, the human agency of participation, and the safety and order of governance.

Real societies are unlikely to choose one position faithfully. They will write these conflicting wishes into institutions together: optimal outcomes in some domains, genuine human participation in some decisions, and controllability and clear responsibility where the stakes are highest.

What education retains, whom it serves, and how it exists will take shape amid those tensions.

But that is the choice society faces if the extreme premise comes true. Today there is no settled answer, and education cannot wait for one before deciding what to cultivate.

What, then, should education attend to during a transition that may last a long time?

When generation becomes easy, who bears the cost of proof?

AI changes more than the quantity of content. It changes the relationship between an output and the ability of the person presenting it.

In the past, producing a complete report, a working program, or a tightly structured paper usually indicated an investment of time and a matching level of knowledge and skill. The output itself was part of the evidence of capability.

AI breaks that correspondence. It lets everyone produce at scale and at low cost, while creating a nearly limitless appearance of capability. What someone can hand over tells us less and less about what they themselves can do.

An output may be real and valuable. But when you can produce one professional-looking result and others can produce thousands of equally complete ones, a new question follows:

Why should the recipient accept yours? How can society avoid re-evaluating every result from scratch?

Reports, programs, images, videos, literature reviews, and plans can all be produced in volume with AI. Specialized knowledge production is growing quickly too. AAAI received 15,532 submissions in 2025. During the AAAI-27 submission period, one submitter reported that paper IDs in OpenReview had passed 50,000. In June 2026, arXiv received a record 32,040 new submissions—more than 1,000 a day on average.

Submission IDs and new-submission counts measure how much content entered academic systems, not how many conclusions were established and accepted. They tell us first that the speed and scale of production have changed. Deciding what is valid, important, and worth acting upon is a separate process after generation.

Polish is also a weaker filter. A rigorous-looking paper may have a flawed argument. Running code may fail at its boundaries. A complete-looking plan may not be worth executing.

People, AI, automated tools, or institutions can carry out verification. Whoever does it spends resources. Cheaper generation has not eliminated the cost of proof; if the producer does not bear it, the recipient will.

The value of a result depends not only on whether it is correct, but on the cost of trusting it. Sources, evidence, tests, limits of applicability, expert review, and a producer’s established reputation can carry part of that burden in advance. They let recipients decide whether to adopt a result at a cost proportionate to its risk, without starting from zero.

Assessment needs a defense, not just a finished product

The way society decides whether to trust a result should shape how schools assess students. When a polished output no longer shows that its submitter has the corresponding capability, a school cannot look only at the finished product. Assessment needs a dynamic process: keep questioning the result, and require the submitter to help make it trustworthy.

Call this defense-based proof for now. It might take the form of an oral defense, project demonstration, live test, replicated experiment, or audit. The form can vary; the central relationship remains: the recipient can keep asking, and the submitter cannot leave all the cost of judgment to the recipient.

At school, the teacher temporarily stands in for the recipient and asks the kinds of questions a result would face in society or the market:

AI may have generated every word of what you submit. But you must give me sufficient reason to believe it is right.

Outside school, “me” could be a customer, peer, employer, regulator, or another AI system. A defense at school should model that social relationship of trust.

Set aside, for a moment, whether students should be allowed to use AI. What schools need to observe is whether someone can subject a final result to professional questioning, and keep explaining, checking, and revising it as counterexamples and new evidence appear. Why this question? Which assumptions support the conclusion? Would a different treatment of the data change the result? What evidence would most likely overturn the judgment? Where does the method stop working?

Some UK universities already warn in guidance for graduate students that those who do not adequately understand AI-generated portions of a dissertation may struggle to defend them in an oral examination. Oxford also asks graduate students to explain their use of AI to examiners in detail and critically, and to take responsibility for code and results they adopt. Such assessment examines not only the paper but whether the student can continue to explain, test, and revise the work.

The duty to substantiate does not dictate in advance exactly what a student must learn. It repeatedly reveals what is missing. Faced with questions they cannot predict, students may need traditional disciplinary knowledge, or may need to use AI, design tests, seek counterexamples, identify limits, and organize evidence. To justify a submitted result, they must develop the capabilities that justification requires. Proof becomes a real reason to learn.

What to learn follows not from “Which steps forbid AI?” but from “What am I missing in order to substantiate this result?”

Schools must substantiate the ability to substantiate

A defense establishes the first relationship: the student demonstrates to the school that a particular result holds up. The school has a second task. Through repeated assessments across different tasks, it must judge whether the person has that ability reliably, then communicate the judgment to society.

The object of assessment should expand from one result to a person’s repeated performance across tasks. One defense shows only that the student can take responsibility for this result. Assessments across courses, projects, and situations can establish whether the ability is stable.

Each reliable assessment leaves evidence of capability. Credits should record the range in which the ability has been demonstrated: has the person produced work in a subject and defended it? A degree asks a further question: can they do this consistently across courses, projects, and contexts?

Student substantiates a result → school assesses the ability repeatedly → credits record its scope → degree judges its stability → society receives an initial judgment

UNESCO has described educational qualifications as a proxy for what someone can do. A qualification has always been such a proxy. It does not contain every piece of evidence collected over years of schooling; it compresses prolonged assessment into a judgment society can read quickly. In the AI era, what needs to change is the capability that proxy represents. A school should stake its credibility on a new claim: this person has repeatedly demonstrated an ability to substantiate the results they present.

Amid abundant content, a valid qualification cannot decide whether any particular result is correct. It can change the starting point from which others judge it. Recipients need not review every assessment record from those years. They can use the school’s compressed judgment to decide where verification should begin and how much it will cost. That is an initial judgment approaching O(1).

A qualification does not exempt a result from proof. It can keep that result from having to earn all trust from zero.

For students, future work begins with greater initial credibility. For society, limited attention and verification resources can move more quickly toward people and results with a reliable history.

Work records, project portfolios, professional qualifications, licenses, and enduring reputations can create similar signals. Schools are distinctive because they can build one through systematic, long-term assessment before a person enters wider society.

For a school to put its reputation behind a judgment of capability, that judgment must rest on real evidence collected over time. Courses, examinations, projects, and defenses should repeatedly ask the same question across different tasks: can this student give reasons sufficient to support trusting the work they present?

During this potentially long transition, education needs to cultivate people who can substantiate their results. Schools need to give society a reason to trust that capability at a reasonable cost.

Translation: Codex prepared this English version from Biaoo’s published Chinese article. Biaoo developed the original thought experiment and argument.