We’re Teaching Students to Use AI Before Teaching Them to Disagree With It

AI literacy is quickly becoming part of education.

Students are being encouraged to learn how to use AI tools, how to write better prompts, how to improve their productivity, and how to work alongside systems that can generate text, code, images, analysis, and ideas within seconds.

All of that is useful.

But I think we may be skipping an important step.

We are teaching students how to ask AI for answers before teaching them how to question those answers.

And the more I see AI being used in education, the more I feel that this distinction matters.

It is now remarkably easy to produce something that looks convincing.

The difficult part is no longer always producing the answer.

The difficult part is knowing whether the answer deserves to be trusted.

A correct-looking answer is not necessarily a correct answer

This problem did not begin with generative AI.

Students have always been able to use calculators without understanding mathematics, follow a statistical procedure without understanding why it was appropriate, or copy a method from another paper without really knowing whether it suited their own research question.

AI has simply made this easier, faster and much more convincing.

A student can now generate an analysis that appears technically sophisticated without understanding the assumptions behind it.

They can receive a recommendation for a machine-learning model without knowing why that model was selected.

They can ask AI to interpret precision, recall or a confusion matrix and receive an explanation that sounds perfectly reasonable.

But if we ask one more question — why? — the weakness sometimes becomes visible.

These are not questions about AI.

They are questions about judgement.

And judgement is becoming more important, not less.

We may be focusing too much on prompting

There has been a lot of discussion around “prompt engineering”.

Students are often told that better prompts produce better results.

That is true to some extent.

But I am not convinced that prompting should be treated as the central skill of AI literacy.

AI systems will change.

Interfaces will improve.

Models will become easier to use.

Many of the prompting techniques that feel important today may eventually become unnecessary.

The ability to evaluate an answer, however, will remain important.

A student who knows how to write an excellent prompt but cannot recognise a poor answer is still vulnerable.

A student who asks a mediocre question but has enough subject knowledge to challenge the response may actually be in a much stronger position.

The uncomfortable part is that AI often sounds right

One reason this is difficult is that AI does not always look uncertain when it is wrong.

It can give an incorrect answer in the same calm and confident language that it uses for a correct one.

For an experienced person, this can still be problematic.

For a learner, it is much harder.

Students are still developing the knowledge that allows them to recognise when something does not make sense.

If an AI system suggests an inappropriate statistical test, an experienced researcher may notice immediately.

A student may not.

If it generates a plausible but inaccurate interpretation of a clinical concept, an experienced clinician may feel that something is wrong.

A learner may simply see a fluent explanation.

This creates an interesting educational problem.

We often expect students to use AI critically.

But critical use depends partly on having enough knowledge to challenge the system in the first place.

So perhaps AI literacy cannot be separated from subject knowledge as easily as we sometimes assume.

You need to know something before you can confidently say:

“No, that does not look right.”

Disagreement may be an important learning skill

I increasingly think that one useful way to teach AI is to deliberately ask students to disagree with it.

Give them an AI-generated answer and ask them to find the weaknesses.

Ask them what information is missing.

Ask them which claims need verification.

Ask them whether the conclusion goes beyond the evidence.

Ask them what they would change.

Ask them to explain where they agree with the system and, more importantly, where they do not.

That exercise may teach more than simply asking students to generate another answer.

Because once students become comfortable disagreeing with AI, the relationship changes.

AI stops being an authority.

It becomes a tool.

And there is an important difference between the two.

Tools assist judgement.

Authorities replace it.

Education should probably be very careful about allowing the second to happen.

This also changes how we think about assessment

There is another reason this matters.

AI has made the final product less informative than it used to be.

A polished report does not necessarily tell us how much the student understands.

Excellent grammar does not necessarily indicate excellent reasoning.

Correct code does not necessarily mean that the student understands the method.

A strong-looking answer may now tell us surprisingly little about the thinking that produced it.

That means assessment may need to place greater value on explanation.

Not simply, “What did you produce?”

But questions such as:

These questions do something useful.

They make thinking visible.

And perhaps that is what assessment increasingly needs to do.

The goal should not be to make students afraid of AI

None of this means students should avoid AI.

That would probably be unrealistic, and in many cases unnecessary.

AI can be genuinely useful.

It can help students explore unfamiliar ideas, simplify difficult explanations, generate alternatives, debug problems, organise information and receive immediate feedback.

Used well, it can support learning.

The problem begins when convenience becomes dependence.

If a student reaches a point where every uncertainty is immediately handed to AI, something important may be lost.

Struggling with a problem is sometimes part of learning.

Being unsure is sometimes useful.

Making a mistake, recognising it and correcting it can teach more than receiving the correct answer immediately.

Education has never only been about arriving at the answer.

It is also about developing the person who can arrive there.

That distinction becomes easy to forget when answers are available instantly.

Maybe AI literacy needs a different definition

Perhaps AI literacy should not be measured by how comfortably someone can use an AI system.

Maybe it should be measured by what happens after the system responds.

Those questions feel more important to me than whether someone knows the perfect prompt.

Because the future will probably not suffer from a shortage of people who know how to ask AI for answers.

AI systems themselves will make that increasingly easy.

What we may need are people who can look at a convincing answer and still have the confidence to ask:

“Does this actually make sense?”

That is not resistance to AI.

It is probably one of the most responsible ways of using it.

And perhaps, before we teach students to become better users of AI, we should make sure they are also comfortable disagreeing with it.