Ask for a hint, not the answer: taking the help without losing the skill

Overview

Why now. In July 2026 researchers built a benchmark that watches an AI assistant help a simulated student solve a problem, and compared its choices against a small sample of human helpers. The models intervened more often, and earlier — and where a person tends to offer a nudge, the assistants “provide complete solutions rather than targeted hints”. That is not a bug; it is the default behaviour of a helpful tool. On any task you need to stay good at, ask for a hint or a critique of your own attempt — not the finished answer.

The content

The usual version of this worry is about you: don’t get lazy, keep your judgment sharp. That puts the whole burden on willpower, at nine o’clock on a Thursday, against a tool built to be maximally helpful. It loses.

The more useful read is that this is a setting, not a character flaw — what changes the outcome is the role you give the assistant before you type. The cleanest evidence is a randomised trial of nearly a thousand high-school students in PNAS. Three groups practised maths: one with no AI, one with an unrestricted assistant, and one with a teacher-configured assistant that gave hints instead of answers, primed with the correct solutions and the common wrong turns. Both AI groups did far better while the tool was in front of them.

Then the researchers took it away and ran an unassisted exam. The unrestricted group scored below the students who had never had AI at all — a 17% reduction in grades, or as the paper puts it, “students actually perform worse than those who never had access”. The hint-constrained group came out statistically indistinguishable from control. Same tool, same students, different role.

Read that precisely, because it is easy to oversell. The hint version did not make anyone better than the control group. It stopped them getting worse. Preserved, not improved. And it was more than a hint instruction: teachers built it, with the answers and common mistakes already loaded in, and the interface enforced the rule. The trial cannot tell you how much of the protection came from the hints alone.

There is a second finding in the same study, quieter and more uncomfortable. The hint-arm students did not outperform the control group — but they believed they had. The feeling of having learned survives the help intact, whether or not the learning does.

The same gap shows up in aggregate. A 2026 meta-analysis of programming studies — 23 in all, 10 measuring learning — found gains depend on whether the assistant is there at the test: g = 0.76 where AI was available during assessment, g = −0.06 where it was not. Pooled across everything the learning effect was not significant at all. What you can do with the tool tells you little about what you can do without it.

So the practice is not “use AI less”. Decide in advance which competencies you intend to keep, and bound the assistant’s role on those only. Everywhere else, let it do the work.

Two honest limits. This discipline is not popular: in a study of interventions that force you to engage before accepting an AI’s output, the designs that most reduced over-reliance were the ones participants rated least favourably. And if you are a genuine beginner, a worked answer is probably the better teacher — a hint about a subject you have no footing in just leaves you stuck.

Try it

Pick one real task this week you would be embarrassed to be bad at in two years — an analysis, a kind of drafting, a judgment call you are known for. Then:

1. Write your own attempt first, however rough — three bullet points is enough. Skipping this makes the rest cosmetic.
2. Give the assistant one job, and say what it is not. For example: “Here’s my draft analysis. Don’t rewrite it and don’t give me your version. Name the single weakest assumption in it, and one thing I’ve left out.”
3. Judge its answer against your own reasoning rather than replacing yours. Where you disagree, work out which of you is right before moving on.

The deadline version. Sometimes you will just take the answer. When you do, write one line predicting what it will say before you read it, then check where you were wrong. Ten seconds, and it turns reading into a test.

Where it breaks. On throughput work — formatting, first-pass summaries, tidying — this is pure friction, and you should skip it. And note what the trial tested: a tool built to withhold answers, not an adult choosing to ask for less and free to change their mind mid-prompt. Nobody has measured the self-imposed version — treat it as a sensible handle, not a proven mechanism.

Additional reading

• Generative AI without guardrails can harm learning — Bastani et al., PNAS (June 2025) — the randomised trial behind this module.
• AI Assistants Overassist — arXiv:2607.21306 — when AI helpers step in, against when people do. Preprint.
• GenAI, productivity and learning in programming — arXiv:2605.04779 — where the gains go once the tool is taken away. Preprint; small subgroups.
• To Trust or to Think — Buçinca, Malaya & Gajos — peer-reviewed (CSCW 2021): the designs that best reduce over-reliance are the ones people like least.

Editor’s note

I’m currently working through a very big project, and I have adopted this method of work to try and get it done. Like the students who were given hints instead of answers, I feel like I’m getting better work out of myself because of the process, though whether that is true on an individual basis is unknowable. It is also frustrating. Knowing the tool could simply do the work for me, and choosing to do it myself anyway, is an exercise in restraint. If you want to keep your edge, I believe it’s worth it.