Why Students Are Getting Better at Finding Answers But Not Always at Asking Better Questions

---

A student can now turn a vague assignment into a polished explanation in minutes. A chatbot can suggest topics, summarize papers, explain difficult concepts, generate examples and even help refine a prompt. The friction that once stood between a question and an answer has fallen dramatically.

That convenience is useful. But it creates a less obvious educational problem: when answers become easier to obtain, the quality of the question becomes more important—not less.

There is growing evidence that generative AI can improve how students perform on individual tasks without necessarily producing equivalent gains in learning. The OECD’s Digital Education Outlook 2026 notes that general-purpose AI can improve task performance while cognitive offloading may reduce the learning that happens underneath that performance.

This does not establish that students generally are becoming worse at asking questions. The evidence is more nuanced. What is changing is the environment in which students learn: AI increasingly rewards questions that produce clear, immediate and usable outputs. Education, however, also depends on questions that are difficult, ambiguous, incomplete or initially poorly framed.

The distinction matters because finding an answer is a task; knowing what is worth asking is a form of judgment.

Key Takeaways

• Generative AI can improve students’ immediate task performance without guaranteeing deeper learning.
• Good questions often emerge through uncertainty, revision and investigation rather than instant clarity.
• AI can encourage students to optimize prompts instead of examining the problem behind the prompt.
• Question generation is closely connected to metacognition, including planning, monitoring and evaluating one’s thinking.
• Schools may need to assess how students reason and revise questions, not only the quality of their final answers.
• AI can strengthen inquiry when it becomes a tool for challenging assumptions rather than simply producing responses.

The New Advantage: Answers Are Cheap

For generations, students faced a relatively predictable sequence.

They encountered a problem, searched for information, read several sources, compared explanations and eventually constructed an answer.

Generative AI compresses that process.

A student researching renewable energy, for example, can ask an AI system for an explanation of grid storage, request simpler language, ask for competing viewpoints and then obtain a list of possible research questions all within minutes.

That is a genuine productivity gain.

But productivity and learning are not identical.

The OECD’s 2026 review of research on generative AI in education makes this distinction particularly important. It reports emerging evidence that students can produce better work with general-purpose GenAI while not necessarily achieving corresponding learning gains. In some studies, performance advantages seen while AI was available disappeared when students later had to work without it.

One example cited by the OECD involves a randomized controlled trial in which access to GPT-4 improved students’ performance while practising mathematics. But in closed-book examinations, students who had used the general-purpose version performed worse than students who had not had access.

The lesson is not that AI makes students incapable of learning.

It is that successfully completing an AI-assisted task can be a poor proxy for how much the student has actually learned.

That distinction changes the role of the question.

The Question Used to Be Part of the Learning

Before a student could easily obtain an answer, the difficulty of research often forced the student to define the problem.

A broad topic such as How does social media affect teenagers? was not yet a good research question.

The student had to narrow it.

Which teenagers? Which platforms? What type of effect? Over what period? Is the question about mental health, communication, education or social relationships? What evidence would establish an effect?

Those decisions were themselves educational work.

Research and inquiry frequently involve failed questions, incomplete searches and contradictory evidence. A student may begin with an assumption, discover that it is unsupported and then reformulate the question.

UNESCO authors Victoria Livingstone and Jeppe Klitgaard Stricker recently described this process as a form of “cognitive friction.” Their argument is that AI can remove some of the productive difficulty through which students learn to refine questions, evaluate sources and tolerate ambiguity.

This is an important distinction.

A bad question is not necessarily wasted effort.

Sometimes discovering why a question is bad is part of learning how to ask a better one.

Better Prompts Are Not Always Better Questions

There is another subtle shift taking place.

AI systems encourage users to become better at specifying what they want from the machine.

That skill is useful.

Students need to learn how to give context, define constraints, provide relevant information and ask for an appropriate format. UNESCO’s AI competency framework explicitly includes the ability to understand, apply and critically engage with AI, rather than treating AI literacy as simple tool usage.

But prompt quality and question quality are not the same thing.

Consider two questions:

“Give me five causes of declining biodiversity.”

and:

“If biodiversity loss is driven by several interacting factors, which of those factors can be distinguished reliably, and what evidence would allow us to separate correlation from causation?”

The second question is harder for a reason.

It does not merely request information. It exposes assumptions, identifies uncertainty and asks what evidence would be necessary to reach a conclusion.

An AI system can help formulate such a question. But the student must still decide what uncertainty matters.

That is the deeper skill.

The Missing Skill May Be Metacognition

Question asking is closely related to metacognition the ability to think about and regulate one’s own thinking.

The Education Endowment Foundation describes metacognition in terms that include planning a task, monitoring learning and evaluating whether an approach worked. It also highlights research suggesting that having students generate their own questions can support reading comprehension.

That provides a useful way to understand the AI challenge.

Imagine two students receive the same difficult assignment.

Student A asks AI:

“Explain this topic and give me the answer.”

Student B asks:

“What am I misunderstanding about this problem? What assumptions am I making? Which parts should I verify independently? What alternative explanations should I investigate?”

Both students are using AI.

But they are using it differently.

The first primarily delegates the task.

The second uses AI as an instrument for examining the task.

That distinction may become increasingly important as AI becomes embedded in ordinary study routines.

The OECD similarly argues that educational use of GenAI works better when it has a clear pedagogical purpose and does not simply replace cognitive effort.

What Happens When Students Stop Sitting With Uncertainty?

One of the less visible consequences of instant answers is the declining value of waiting.

A student who encounters an unfamiliar concept can immediately request an explanation.

That is convenient, but uncertainty itself has educational value.

Suppose a student encounters two contradictory explanations in academic sources. Without AI, resolving the contradiction may require additional reading. With AI, the student can ask a chatbot to explain which source is correct.

That may solve the immediate problem.

But another question follows:

Did the student learn how to evaluate the contradiction, or simply learn how to ask a machine to resolve it?

This is precisely why AI literacy cannot be reduced to knowing how to operate a chatbot.

UNESCO’s student framework emphasizes critical judgment, ethics, human-centred thinking and the ability to understand, apply and create with AI.

The goal is therefore not simply to produce students who can get better answers from AI.

It is to produce students who can question the answer they receive.

The Classroom May Need to Reward the Questioning Process

If AI makes final answers easier to produce, schools and universities face a practical assessment problem.

A polished essay may reveal less about a student’s thinking than it once did.

UNESCO’s discussion of assessment in the AI era argues for greater attention to higher-order thinking, creativity, ethical reasoning and the learning process rather than relying exclusively on final outputs.

That could change what an assignment looks like.

Instead of asking only:

“Write an essay about climate change.”

A teacher might ask students to submit:

• their initial question;
• why they chose it;
• the assumptions behind it;
• how the question changed during research;
• which sources challenged their initial view;
• what AI contributed;
• which AI-generated claims they rejected;
• and what remains uncertain.

The final essay would still matter.

But the intellectual journey would become visible.

That approach also reflects the reality of professional work. Researchers, engineers, analysts, journalists and business leaders rarely solve meaningful problems by receiving one perfect answer. They define problems, test assumptions, compare evidence, revise their questions and make decisions under uncertainty.

AI Can Also Teach Better Questioning

The problem is therefore not AI itself.

An AI system can be used to produce shallow answers, but it can also become a questioning partner.

A teacher could ask students to begin with their own question and then instruct AI to challenge it:

“What assumptions does this question contain?”

“What would make this question too broad?”

“What evidence would change the answer?”

“What important perspective is missing?”

“Give me three competing explanations, but do not tell me which one is correct.”

That changes the role of AI.

Instead of being an answer machine, it becomes a tool for intellectual resistance.

This direction is consistent with the OECD’s finding that GenAI can support learning when deliberately integrated into teaching, including uses that encourage argumentation, collaboration and active learning.

It also aligns with UNESCO’s broader emphasis on developing students who can critically and responsibly interact with AI rather than simply consume its outputs.

The Important Educational Shift

The educational challenge created by AI may therefore be less about whether students can find information.

They can.

The harder question is whether they can decide which information they need, why they need it, what assumptions surround it and what evidence would change their minds.

That is a different form of competence.

A student who can ask AI for 20 answers may appear highly capable.

A student who knows that none of those answers addresses the real problem may be demonstrating the more valuable skill.

This does not mean returning to a pre-AI classroom. Nor does it mean forcing students to struggle with tasks that technology can legitimately simplify.

The more useful approach is selective friction.

Let AI handle tasks where automation improves learning. But preserve opportunities where students must formulate problems, investigate uncertainty, defend reasoning, question evidence and revise their own thinking.

The emerging educational skill may therefore be neither “knowing the answer” nor “knowing how to prompt.”

It may be knowing which question deserves an answer in the first place.

Conclusion

Generative AI has changed the economics of information. Answers that once required substantial searching can now appear almost instantly.

Education must respond by placing greater value on something AI cannot simply substitute for: human judgment about what is worth asking.

The evidence does not justify saying that students as a whole are becoming worse questioners. But it does show why educators should be cautious about equating faster, better-looking outputs with deeper learning. Research reviewed by the OECD indicates that general-purpose GenAI can improve task performance without necessarily producing equivalent learning gains, while UNESCO’s educational frameworks emphasize critical judgment and human agency.

The answer may not be to keep AI away from students.

It may be to teach students to use AI after they have thought about the question and sometimes to use it specifically to make that question harder, deeper and more revealing.