When Answers Become Cheap, What Is Education For?

A student submits a clear explanation of a difficult concept. The argument is coherent, the vocabulary appropriate, and the conclusion correct.

What can a teacher infer?

Before generative AI, the answer was never entirely straightforward. A student could receive help, imitate a source, or memorise an explanation without understanding it. But producing a competent piece of work generally required some engagement with its contents.

AI makes that relationship less dependable. A finished assignment can now conceal a much wider range of underlying abilities.

This is often discussed as a cheating problem. It is also a measurement problem, a teaching problem, and eventually an economic one. Education is expected both to develop competence and to provide evidence that competence exists. When polished output becomes easier to obtain, those functions become harder to align.

Performance and learning are different outcomes

Consider two students using an AI tool during practice.

One asks for an explanation, attempts the next step, and checks where their reasoning failed. The other requests a completed solution and reproduces it.

Both may submit correct work. Their immediate performance can look similar even if the practice leaves them with different abilities.

A field experiment reported in PNAS in 2025 examined this distinction among nearly a thousand high-school mathematics students in Turkey. Researchers compared a standard GPT-based interface, a tutor designed with learning safeguards, and a control condition.

AI access improved performance during supported practice. Once the tools were removed, students with the standard interface performed worse than the control group. Safeguards designed to preserve learning largely mitigated that negative effect.

The result does not establish that AI generally harms education. It shows why successful task completion cannot automatically be treated as evidence of learning.

It also points to a design conflict. A general assistant is often expected to reduce the effort needed to obtain an answer. A tutor sometimes needs to preserve effort long enough to discover what the learner does not understand.

The evidence also contains genuine reasons for optimism

Other research has found benefits from AI-supported instruction.

A randomised study in a Harvard undergraduate physics course compared a purpose-built AI tutor with classroom active learning for specified lessons. The tutor group achieved stronger immediate learning outcomes while spending less time on the material. The system incorporated instructional design and structured guidance; the intervention was more than unrestricted chatbot access.

In Nigeria, a World Bank working paper evaluated a six-week, teacher-supported after-school programme using generative AI. It reported improvements across assessed areas including English and digital skills. Those findings concern a particular programme with organised participation and support. They do not isolate the effect of simply making a chatbot available.

These results are not necessarily contradictory.

A tool that supplies answers, a tutor built around instructional principles, and a teacher-supported programme are different educational interventions. The model is one component of each.

The relevant research questions concern what students do, what support surrounds them, and whether the acquired ability lasts beyond the lesson.

Cheaper explanations do not guarantee equal opportunities

The prospect of widely available tutoring is important. A learner who lacks access to individual help may benefit substantially from an explanation they can request repeatedly and without embarrassment.

But availability and effective use are separate conditions.

Imagine two hypothetical students with access to the same model. One has a quiet place to study, a clear curriculum, and someone who notices persistent confusion. The other has limited time, unreliable access, and little guidance about what to ask.

The software is identical. The learning environments are not.

This suggests a risk that deserves investigation: as explanations become cheaper, other resources may become more decisive. Attention, study time, feedback, and the ability to judge an answer could account for a greater share of the difference between learners.

That is an inference about how inequality could develop, not a finding established by the studies above. It matters because it changes the policy question.

Providing a tool may be worthwhile. Evaluating access alone would still leave unanswered whether students can turn that access into understanding.

The assessment problem will not disappear with better detection

If a written assignment can be generated, institutions need to reconsider what it demonstrates.

Attempts to identify AI-written work address authorship. Even perfect authorship detection would not fully resolve whether a student understands the subject.

An educational system also needs ways to observe reasoning: asking learners to explain a choice, respond to a changed condition, diagnose an error, or apply a concept in an unfamiliar case.

These approaches have costs. Oral assessment takes time. Closely observed practical work requires staffing and organisation. Large courses cannot simply replace every written assignment with an individual conversation.

An economic tension follows. AI can lower the cost of producing an impressive answer while institutions may need to spend more effort establishing the competence behind it.

That could shift the value of assessment towards activities that are harder to scale. Whether institutions can provide those activities fairly is a question of resources as much as technology.

The same problem continues after graduation

Workplaces also teach through tasks.

A junior employee develops knowledge while drafting, checking, debugging, and being corrected. If AI completes these activities, the organisation may gain output while the employee gains less experience.

A 2026 preprint by Judy Hanwen Shen and Alex Tamkin tested developers learning an unfamiliar programming library. AI assistance impaired several measured aspects of mastery without delivering a significant average efficiency gain in the experiment. Interaction patterns involving active cognitive engagement were associated with better-preserved learning. The study concerns a bounded task, rather than the long-term careers of developers.

Its implications are nevertheless relevant to education.

Students are often told that the important skill will be supervising AI. Effective supervision requires enough knowledge to recognise when a plausible result is wrong. A curriculum that removes too much foundational practice may weaken the very ability it expects graduates to exercise.

The appropriate amount of independent practice will vary by subject. There is no reason to preserve every old task simply because it predates AI. But replacing a task requires understanding what it taught.

What should remain difficult?

Education has always used tools to remove some kinds of effort. Access to books, calculators, search engines, and software changes which activities deserve time.

Generative AI makes that decision newly urgent because it can perform tasks that previously served as visible evidence of thought.

Some difficulty is unnecessary: unclear instructions, inaccessible materials, long waits for feedback. Removing it can help learning.

Other difficulty comes from attempting an explanation, discovering a contradiction, or revising a mistaken understanding. A well-designed educational tool needs to distinguish these situations.

The unresolved question is how to make that distinction at scale, across subjects and learners, without reducing education to either constant surveillance or effortless completion.

A student leaving an AI-supported lesson should have more than a correct answer in the conversation history. There should be something they can now notice, explain, or do that they could not do before. Establishing that difference is becoming one of education’s most important jobs.