Exam prep · Updated 20 September 2026

Test Anxiety and AI: What Actually Works

Midterms are two weeks out, and the advice on offer is a breathing exercise or a chatbot telling you it will be fine. The thing with the most evidence behind it is neither.

There is no AI trick that settles exam nerves, and the research is fairly clear about it: the one meta-analysis that has pooled AI interventions on academic anxiety found no significant effect for chatbots, and none for university students either. What does have a measured effect is older and duller — taking practice tests. Across 24 studies, practice testing reduced test anxiety by a medium amount, and the three things researchers say make quizzing feel less threatening (ungraded, retakable, nobody watching) happen to describe a chatbot-generated practice quiz exactly. Use the tool as the quiz, not as the counsellor.

The chatbot pep talk has not been shown to work

In June 2026, Weihao Wang, Long Diao and Jingting Yu published a meta-analysis in BMC Psychology pooling 22 studies and 30 effect sizes on whether AI interventions reduce academic anxiety. The headline is encouraging: "AI interventions showed a significant moderate overall effect on reducing academic anxiety (g = − 0.65)." Test anxiety specifically came out at g = −0.98.

Then you look at which studies produced that number. Intelligent tutoring systems worked (g = −1.08, nine effects) and so did robotics. Chatbots did not reach significance (g = −0.33, p = .33, seven effects). The significant age groups were primary and middle school students; the university subgroup was the largest in the whole analysis at eighteen effects, and it was not significant (g = −0.35, p = .15). Eastern contexts were significant and Western ones were not (g = −0.27, p = .35).

Every cell that describes a university student in the US or UK typing at a chatbot came back non-significant. Be careful how you read that, though. The between-group tests were non-significant too, and the authors say plainly that "observed differences are descriptive rather than statistically robust." This is an absence of evidence, not evidence of absence — nobody has shown a chatbot calms exam nerves, and nobody has shown it makes them worse.

Neither does the famous writing trick

If you have looked this up before, you have met the ten-minute writing exercise: write out your worries just before the exam, which supposedly frees up working memory. It comes from a 2011 Science paper by Ramirez and Beilock, and it is repeated in nearly every study-skills article, including ones published this year.

It did not survive testing. In 2026, O'Meara and Lovett published a meta-analysis in Anxiety, Stress & Coping covering 21 studies and 1,457 participants, and found "only a negligible, nonsignificant effect of expressive writing on anxiety (r = −.05, p = .57)", with "very little evidence of publication bias" to explain it away. Their conclusion runs to one sentence: "The available research literature fails to support single sessions of expressive writing as a treatment for test anxiety."

Two earlier results point the same way. Myers, Davis and Chan tested it across four real exams in a cognitive psychology course at Iowa State and reported in Cognitive Research: Principles and Implications that "neither intervention was effective at reducing test anxiety or improving exam performance." They also note that the Social Sciences Replication Project "attempted twice to replicate Ramirez and Beilock's study", and "both of their amply-powered experiments failed to find a benefit of expressive writing on test performance."

This matters beyond one exercise. Ask a chatbot for evidence-based ways to handle test anxiety and it will hand you this technique, confidently and with a citation, because it is one of the most-repeated findings in the study-skills corpus. Repeated is not the same as true — the problem we cover in studying social psychology after the replication crisis.

The thing that does have a number attached

In 2023, Chunliang Yang, Jiaojiao Li, Wenbo Zhao, Liang Luo and David Shanks asked a question nobody had pooled before: do practice quizzes reduce test anxiety, or cause it? Their meta-analysis in Educational Psychology Review covers 24 studies, 25 effects and 3,374 participants, and the answer is not close. Practice testing reduced test anxiety at "Hedges' g = −0.52, 95% CI = [−0.69, −0.36], p < .001", with 22 of the 25 effects (88%) pointing that way. Correcting for publication bias barely moved it, to g = −0.50. In classroom settings specifically — which is your situation, not a lab — it was g = −0.55.

That is larger than anything the chatbot subgroup managed, and it is the same activity that, in a separate review of 222 studies and 48,478 students by mostly the same group, raised academic achievement by g = 0.499. One activity, both effects. Nothing else on this page does as much work.

Why you avoid the thing that works

Here is the detail that explains the whole problem. The same paper pooled two studies comparing how stressful practice testing feels against re-reading: "practice testing was more stressful than restudying, g = 0.60, 95% CI = [0.15, 1.05], p = .009." So quizzing yourself lowers your anxiety about the exam while being more unpleasant in the moment than re-reading your notes. That is backwards from how anyone picks what to do at 9pm, and it is why the highlighter wins. Re-reading feels calm and competent, which is the feeling, not the preparation.

The authors, writing for instructors, say the fix is to make quizzes less threatening rather than rarer, and the examples they cite are specific: students rated ungraded quizzes as less stressful than graded ones; 89% reported positive experiences with quizzes allowing multiple attempts; and over 90% agreed that "the option to retake a quiz makes me feel secure and reduces anxiety."

Start easier than feels serious

One more finding, and it is the counterintuitive one: within that meta-analysis, easier practice tests reduced anxiety more than hard ones, with every 10% increase in practice-test score corresponding to a further g = −0.17. Only five of the 25 effects reported practice-test performance, so the authors "warn readers to interpret these results cautiously". Treat it as a sensible default rather than a law.

It cuts against instinct, because a practice test you pass feels like a wasted evening, and the paper names the trade-off: "easy practice tests tend to be more effective in reducing TA but less effective in enhancing learning." You do not have to pick. Their resolution is successive relearning — "repeatedly taking practice tests on study materials with corrective feedback until successful retrieval is achieved" — which leaves students with "robust learning but also end up with very little anxiety." In practice: an easy first pass, then repeating only what you missed, with the answer explained, across several days. Prompts 1 and 3 do that.

Split the job

The workflow, step by step

Four prompts, in order, starting about a week out. None offers reassurance or predicts your grade, which is deliberate: a chatbot will agree that you are going to do fine whatever the truth, so that output tells you nothing.

1. The deliberately easy first pass. The one to run tonight. Resist the urge to make it hard.

Here are my notes for [course / topic]:

[paste them]

Write 15 EASY recall questions from this material only. Easy
means: one fact or one step per question, answerable in a
sentence, no multi-part questions, no trick wording, nothing
that needs material I did not paste.

Ask me ONE at a time and wait for my answer. Do not show me
the answers first. After each answer say only "correct" or
"not quite" plus the correct answer in one line, then move on.

No encouragement, no score commentary, no "you've got this".
At the end, just list which question numbers I missed.

Keep that list of missed numbers. It is the only output of this step that matters.

2. The worry triage. The honest replacement for writing out your feelings: sort them into the two kinds that have an action.

Below is everything I am worried about for this exam, in no
particular order:

[dump it - messy is fine]

Sort each item into exactly one of three buckets:
A) A knowledge gap - something I could turn into a practice
   question and drill.
B) A fact about the exam I do not know (format, length,
   materials allowed, weighting, what happens if I fail).
C) Neither.

For A, rewrite it as a specific practice question. For B,
write the one-line question I should email my instructor.
For C, write "no action" and say nothing else about it.

Do not reassure me, do not rank my worries, do not tell me
which ones are reasonable, and do not add worries I did not
write.

Bucket B is the one that surprises people. A good share of what feels like anxiety turns out to be not knowing how long the exam is or whether a formula sheet is provided — a two-line email, not a mood.

3. The relearning loop. Run this on days two, four and six — not all in one sitting.

These are the questions I missed last time:

[paste the missed items]

Re-ask me only these, one at a time. When I get one wrong,
give me the correct answer AND one sentence on why, then put
that question back in the pile and ask it again later in the
same session.

A question is retired only when I have answered it correctly
on two different days. Track that for me and tell me at the
end which questions are still not retired.

Do not add new questions. Do not rephrase a question to make
it easier once I have missed it.

4. The dress rehearsal. Last, once the easy material is solid. This is the only one that should feel like the real thing.

My exam is [length], [format: MCQ / short answer / essay /
problems], [open or closed book], covering [topics].

Build one practice paper matching those conditions exactly.
Do NOT invent details I did not give you - if something is
missing, tell me to go and ask my instructor rather than
guessing.

Give me the whole paper at once with no answers. I will time
myself and send my answers back in one message.

Then mark it against the material I gave you, show which
topic each mistake came from, and tell me the single topic to
spend tomorrow on. Do not predict my grade.

Do it once, early enough that a bad result is useful rather than frightening. A rehearsal the night before is not practice — it is just an extra exam.

When this is bigger than study skills

Everything above is a study-skills answer to a study-skills-sized problem, and test anxiety is not always that. If it is costing you sleep, making you avoid classes or exams, or producing physical symptoms, the thing worth knowing is that it responds well to actual treatment. Huntley and colleagues pooled 44 randomised controlled trials of 2,209 university students in the Journal of Anxiety Disorders and found interventions reduced test anxiety at g = −0.76, with "most support for behaviour therapy". They flag publication bias and uneven reporting, so hold the exact number loosely — but the effect is larger than anything a prompt here will get you.

That is your campus counselling or student wellbeing service, and the wait is usually shorter in September than in December. If the anxiety is tied to a diagnosed condition, our guide to AI use and disability accommodations covers how to get it documented rather than argued about.

Where the line is

Nothing here goes near an integrity rule: you produce the answers, and nothing you make gets submitted. Check only whether your course restricts uploading its materials, which our privacy checklist and class AI policy checklist cover. One real risk remains and it is not a rules risk: a quiz built from AI-generated content rather than your notes will teach you wrong facts with total confidence, so keep the source material yours.

Related reading

• Build the quiz properly. The full exam prep system, practice exams from class notes, and quizzing yourself from notes.
• Spacing and recall. Anki and active recall and flashcards from notes, which are the same mechanism on a longer timetable.
• If your exam is a conversation. Preparing for oral exams and vivas, where unfamiliarity rather than knowledge is the documented risk.
• The week itself. A seven-day exam plan and the over-reliance checklist.

FAQ

Does asking ChatGPT for reassurance before an exam help?

There is no evidence that it does. The 2026 meta-analysis of AI interventions on academic anxiety found the chatbot subgroup non-significant (g = −0.33, p = .33) and the university-student subgroup non-significant as well, despite that being the largest subgroup in the analysis. The authors caution that their subgroup comparisons are descriptive rather than statistically robust, so this is an absence of evidence rather than proof it fails. It is also worth remembering that a chatbot will tell you that you are well prepared whether or not you are, so the answer carries no information about you.

Should I write out my worries before the exam?

Probably not, if you are doing it because you read that it works. A 2026 meta-analysis of 21 studies and 1,457 participants found "only a negligible, nonsignificant effect of expressive writing on anxiety", the original 2011 Science result failed two well-powered replication attempts, and a classroom study across four real exams found no benefit to anxiety or scores. If writing settles you personally, nothing is lost by doing it — just do not use it instead of practice questions.

Will making practice tests just make me more anxious?

Less anxious about the exam, more uncomfortable during the quiz. Across 24 studies, practice testing reduced test anxiety at g = −0.52, with 88% of effects pointing that way. But the same paper found practice testing felt more stressful in the moment than re-reading (g = 0.60), which is why students avoid it. Making the first quiz deliberately easy, ungraded and private is the documented way to keep it bearable.

How hard should my practice questions be?

Easier than you think at first. Within the same meta-analysis, every 10% increase in practice-test score corresponded to a further g = −0.17 reduction in test anxiety, though that estimate rests on only five effects and the authors ask readers to treat it cautiously. Harder questions are better for learning and easier ones are better for anxiety, so run an easy pass first, then repeat only what you missed with the answers explained, over several days.

When should I stop using study tips and talk to someone?

When the anxiety is not tracking your preparation — if you are well prepared and still unable to sleep, avoiding exams, or getting physical symptoms. A meta-analysis of 44 randomised trials with 2,209 university students found interventions reduced test anxiety at g = −0.76, with most support for behaviour therapy. That is a bigger effect than any prompt on this page, and it is what your campus counselling service does.

Bottom line

Nobody has shown that talking to a chatbot calms exam nerves, and the writing exercise everyone recommends does not hold up. What holds up is quizzing yourself — a medium-sized reduction in test anxiety that is also the best-evidenced way to learn the material. It feels worse than re-reading while you do it, which is exactly why it gets skipped. So make the first quiz easy, keep it ungraded and private, repeat only your misses over a few days, and rehearse under real conditions once. The chatbot's job in all of that is writing questions, not offering comfort.