Skip to content
NewsletterAI Learning, Keeping It HumanCall647-641-0643
Christopher Ross

Guides

Ask the AI last

By 8 min read

A new Finnish study watched 891 students take a take-home exam with AI permitted. The students who did their own thinking first and asked the model for feedback at the end outscored every other group, including the ones who never touched it. The order matters more than the prompt.

A paper landed on 6 October that I keep thinking about, because it answers a question I had not seen anyone test cleanly: among students who use AI on an assignment, does when they use it predict how well they do?

Outi Hakola, Jari Kukkonen and Teemu Valtonen at the University of Eastern Finland looked at 891 take-home exams written across 2023 and 2024, in a system where students were allowed to use AI as long as they disclosed it. 159 students disclosed using it. The authors sorted those disclosures into three families: instrumental (hand the AI the question, use what comes back), ideational (ask the AI to help think), and reflective (write your own answer first, then ask the AI for feedback). The reflective group scored higher than every other group in the study, including the students who did not use AI at all. The instrumental group scored lower than the non-users (University of Eastern Finland press release, 6 October 2026; reported by phys.org on 7 October; paper in Computers in Human Behavior Reports, DOI 10.1016/j.chbr.2026.101340).

Associate Professor Hakola, who led the study, put the finding plainly in the press release: "The use of AI is neither inherently good nor bad. What matters is whether it is used in a meaningful way." I would say it more specifically. For a student on an assignment, the order they bring the model in is doing more work than the prompt they feed it.

The order matters more than the prompt.

Keep the caveats in view before the argument goes anywhere. 159 students is a small cell, split three ways is smaller, and the use-type is self-reported from a disclosure field written after the fact. The students chose their own order, which means this is correlational and the people who ask AI last may be stronger students already. The authors say as much. What they have measured is a pattern that holds up within one cohort at one institution; what they have not measured is causation, and I will not pretend otherwise.

With that said, the pattern is consistent with what a different group found this year on the quality-versus-uptake question. Farrokhnia, Latifi, Papadopoulos and colleagues ran a randomized three-group experiment with 70 graduate students writing argumentative essays, and compared teacher feedback against AI feedback generated with two prompting approaches. The AI feedback produced with chain-of-thought prompting was rated higher in quality than the teacher's. The students who received it did not revise their essays any better than the students who got the lower-rated teacher feedback (International Journal of Educational Technology in Higher Education, Vol. 23, Issue 1, December 2026; working-paper record at Wageningen University). The authors read that honestly. What determines whether the writing improves is the student's engagement with the feedback, and the quality of the feedback has almost no bearing on that.

Two studies looking at different ends of the same question are arriving at the same answer. If the student has already done the cognitive work, feedback lands. If they have not, better feedback goes to waste.

Why the order does the work

A real draft, written by one person over an hour, carries that person's gaps. The paragraph that went soft and the claim they could not actually defend. Feedback aimed at those specific gaps teaches, because the student knows exactly where it hits. Feedback on a draft the model wrote lands on a surface the student did not shape, and the student has nothing in them to compare it against. One of those gets filed as a grade. Only one gets remembered.

The instrumental finding is the mirror image of this. Hand the question to the AI and submit what comes back, and the only thing you have practised is pasting. The paper's correlational, so the drop below non-users is not proof that AI made those students worse; it is at least compatible with the simpler reading that they skipped the exercise and the grade reflects it.

The copyable version, for one person on one assignment

Three ordering rules a student or anyone else learning something can test this week. My own distillation, drawn from the research and untested in it.

  • Write first, then ask. The first draft carries your own gaps, which is what feedback has to find. Fifteen minutes of writing before any prompt opens is the whole intervention.
  • Ask for the gap rather than the fix. "What is the weakest claim in this and why" produces a different lesson than "rewrite this." The second gets you a sentence; the first gets you something you can use next week.
  • Revise once before accepting anything. If the AI's suggestion is adopted verbatim, the version in the head has not changed. Edit the suggestion into your own words, or reject it.

None of this needs a tool budget or an institutional policy. It needs the student, or the professional learning anything, to do the first half themselves. Outside the exam room, where no one is forcing the cold start, that costs something the ordering rule does not mention: the willingness to begin badly. A student with three assignments due, a job and a child has that willingness on strict ration, and reaching for the model at the blank page is often the only way to begin at all. The ordering rule is still right. It is also harder than it looks the first time it goes on the page.

Who this leaves exposed

There is a separate finding worth sitting next to this one. Stanford researchers Mei Tan and Lena Phalen tested four large models on 600 eighth-grade persuasive essays this spring, giving each essay to each model under three conditions: no demographics, assigned demographics, and names coded for race or gender. The feedback changed. Female-coded writers got warmer, lighter critique. Students coded as people of colour got feedback focused on "polishing." Lower-expectation framings produced shallower comments (reported in Education Week, 18 June 2026). The study is from June, so treat it as backdrop rather than this week's news.

Put the two together, and the ordering rule changes shape. When the feedback is biased, "ask the AI last" does more than teach: it protects. A student who has already written the paragraph has somewhere to stand when the model nudges them toward a thinner version of it, and a draft of their own in front of them to argue the model back from.

How this was made. Full disclosure, and I mean it as a courtesy, not a confession. AI tools help me research, check my claims, tidy my grammar, and stay on top of far more reading than one person can. The thinking and the judgment are mine, I edit every word into my own voice by hand, and then I put my name on it. Telling you that is the whole point of keeping it human.

One question I would like answered, and I do not have one myself. If you teach, or if you learn alongside a model, where in a working session do you reach for it first? At the blank page, or after you have made a mess of something and want a second opinion? I suspect most honest answers sit closer to the blank page than anyone admits, and I would like to know whether I am right.

Next edition: what AI tutors are actually measuring when they claim to measure learning. The gap between the dashboard and the student.

These editions run alongside Learning to Learn With AI, a longer series working the same ground in order.

Key takeaways

  • A University of Eastern Finland study of 891 take-home exams, where AI was allowed if students disclosed it, sorted the 159 students who disclosed use into three groups: instrumental, ideational and reflective.
  • Students who wrote their own answer first and asked the AI for feedback afterwards scored higher than every other group, including students who didn't use AI. The instrumental group scored lower than the non-users.
  • The result is correlational and self-reported, from one cohort, so stronger students may simply be the ones who ask last. A randomized trial of 70 graduate students found AI feedback rated better than the teacher's still didn't improve revisions.

Common questions

Does asking the AI last improve grades? In the Finnish study, students who wrote their answer first and used AI for feedback afterwards scored higher than every other group. The study is correlational and self-reported, so it can't show the order caused the higher scores. The authors say stronger students may already be the ones asking last.

What were the three kinds of AI use in the study? Instrumental, where the student hands the AI the question and uses what comes back. Ideational, where the student asks the AI to help think. Reflective, where the student writes their own answer first and then asks the AI for feedback.

What can I try this week? Three ordering rules: write first and then ask, ask for the gap rather than the fix, and revise once before accepting anything. They're the author's own distillation of the research and haven't been tested in it.

References