When Answers Become Cheap, What Should Education Teach?

AI exposes a problem that existed long before ChatGPT: completing an assignment and understanding something are different achievements.

A student can now produce an impressive essay, solve a problem or build a presentation without developing the capabilities those assignments were meant to teach. At the same time, AI can provide explanations, practice and feedback that many students previously couldn’t access.

Both are possible. The difference is how we design education.

Here are eight problems. These are an editorial synthesis of the research, rather than a claim that every education system has the same weaknesses.

1. We still confuse access to information with learning.

Putting lectures online increases access. Giving students a chatbot increases access again. Neither automatically creates understanding.

Learning requires people to connect ideas, attempt problems, receive feedback and use what they know in unfamiliar situations. The OECD’s 2026 review makes this distinction explicit: better performance with AI does not necessarily translate into learning. More content is valuable only if students can turn it into capability. ^[1]

2. We reward the finished answer, even when the thinking has been outsourced.

In a study involving nearly 1,000 high-school mathematics students in Turkey, Hamsa Bastani and colleagues found that a general-purpose GPT interface improved practice performance. But once AI was removed, those students performed worse on a subsequent unassisted test than students who had never received it. A tutor designed to give hints largely avoided that harm.

This was one setting, not a verdict on all AI learning. But it demonstrates the problem clearly: a tool can improve the work while weakening the learning behind it. Assessment needs to reveal reasoning, revision and independent understanding alongside the final result. ^[2]

3. The economic value of knowledge is changing, but knowledge is still necessary.

The ability to retrieve a fact or produce a conventional first draft becomes less distinguishing when machines can do it cheaply. Choosing the right problem, recognizing a bad answer and making a defensible decision become more consequential.

Economist David Autor argues that AI could extend the reach of human expertise by helping more people perform demanding decision-making tasks. His argument depends on people having complementary knowledge. Education therefore faces a difficult double requirement: build strong foundations while teaching students how to work beyond them with AI. ^[3]

4. Attention is under pressure, and education cannot simply assume it is available.

Gloria Mark’s research documents increasingly frequent switching between screens. The widely repeated “47 seconds” figure concerns observed screen attention; it does not establish that every person’s underlying capacity to concentrate has collapsed.

The useful connection is the attention economy: learning takes place amid products competing to interrupt it. Making every lesson shorter may accommodate fragmentation without helping students develop sustained focus. Education needs opportunities to practice reading deeply, tolerating uncertainty and staying with a difficult problem. ^[4]

5. We risk removing the effort through which expertise develops.

An experienced person can use AI to accelerate work because they already recognize what good work looks like. A beginner may skip the very experiences that develop that judgment.

This creates a potential apprenticeship problem: if machines handle the early tasks, how do people become capable of supervising the later ones? The implication is to distinguish unnecessary frustration from productive effort, and design AI assistance that gradually gives responsibility back to the learner. Bastani’s comparison between answer-giving and hint-giving shows why that design choice matters. ^[2]

6. We teach tool use more readily than intellectual independence.

Prompting is only one part of AI literacy. Students also need to know when to distrust an answer, what evidence would change their minds, whose interests a system serves and when to refuse its use.

Education researchers Stephanie Smith Budhai and Marie Heath argue for this broader, critical approach. It connects AI education to citizenship and agency: people should be able to question the systems shaping their lives. ^[5]

7. We organize education into stages, while adaptation is becoming a continuing responsibility.

School, qualification, employment: this sequence provides structure, but it is insufficient when people must repeatedly learn unfamiliar tools and reconsider established practices.

A useful direction comes from Karen Brennan’s work on creative, self-directed projects: students practice setting goals, experimenting and evaluating their own progress. Those habits matter beyond any particular application. The educational aim is someone who can keep learning when the syllabus runs out. ^[6]

8. We risk treating technology access as educational equality.

Two students can receive the same chatbot and have very different experiences. One gets an answer to copy; the other gets carefully structured practice, a teacher’s guidance and help identifying misconceptions.

The concern is that unequal educational support could persist beneath apparently equal access. The OECD emphasizes teaching principles and implementation, which makes teacher preparation and institutional capacity central to the AI question. ^[1]

Where edtech and digital education differ

The terms overlap, so we should avoid presenting them as strict opposites. For this piece, a useful distinction is:

	What it concerns	Example
Edtech	The technologies used to support education	An AI tutor, learning platform or assessment tool
Digital education	How teaching and learning are organized using digital technologies	How feedback, practice, participation and assessment work together
Education for an AI world	What people need to learn to exercise judgment and agency in that world	Evaluating evidence, developing expertise, collaborating and understanding AI’s consequences

An institution can buy excellent edtech while leaving its educational model largely unchanged. And some of the most valuable responses to AI (discussion, practical work, mentorship and protected concentration) may involve less screen time.

The question running through the piece is: When technology can produce the answer, what must the learner still become capable of doing?

Sources & notes

Dates span 2023–2026. The eight problems and three-way taxonomy are our editorial synthesis; education systems differ and the categories overlap. Diagrams are illustrative, not measured results.

1. OECD. How to effectively use generative AI in education. January 2026. Institutional synthesis; supported performance is not necessarily learning. ↩︎ ↩︎

2. Bastani et al. Generative AI without guardrails can harm learning. PNAS 122(26), e2422633122. June 25, 2025. One high-school mathematics setting in Turkey. The hint tutor largely mitigated harm, not demonstrated superior independent learning. An August 2025 correction concerned author affiliation. ↩︎ ↩︎

3. David Autor. Applying AI to Rebuild Middle Class Jobs. NBER Working Paper 32140. February 2024. Economic argument, not a peer-reviewed causal forecast. ↩︎

4. UC Irvine. Regaining Focus in a World of Digital Distractions. January 26, 2023. Interview with Gloria Mark; 47 seconds describes observed screen switching, not universal concentration capacity or a 2026 population measure. ↩︎

5. Jill Anderson. Teaching Students to Think Critically About AI. Harvard Graduate School of Education. October 8, 2025. Expert interview with Stephanie Smith Budhai and Marie Heath, not experimental causal evidence. ↩︎

6. Elizabeth M. Ross. Tips for Using AI, From Grad Students and Professors. Harvard Graduate School of Education. January 17, 2025. Interview-informed guidance including Karen Brennan; not proof of lifetime adaptability. ↩︎

All articles ↗