An AI study helper works best when it asks questions instead of supplying answers. Lab research since 2006 shows that quizzing on material improves delayed recall more than rereading it, and the few studies on chatbots point the same way: one cohort comparison linked chatbot-generated practice questions to higher exam scores, while a randomized trial found lower retention after students used ChatGPT to research a topic. The evidence is small and recent, so the findings are signals, not settled rules.
Key takeaways
- A 2006 experiment found that recalling material beat restudying it on tests given two days and one week later, though restudying won on a test five minutes later.
- A 2025 Malaysian study found higher final exam scores among students whose practice questions were generated by ChatGPT, but it compared different cohorts.
- A randomized trial of 120 Brazilian undergraduates reported lower 45-day retention for students who researched a topic with ChatGPT.
- A three-arm Virginia Tech trial of AI-supported retrieval practice was still unreported as of October 2026.
- Using AI to generate questions, then answering them from memory, matches the strongest learning research.
What the testing effect says about studying
In "Test-Enhanced Learning," published in Psychological Science in 2006, Henry Roediger and Jeffrey Karpicke had students read prose passages and then either restudy them or take free-recall tests without feedback. A final test came 5 minutes, 2 days, or 1 week later. At 5 minutes, repeated studying gave better recall. On the delayed tests, prior testing produced substantially greater retention, even though repeated studying made students more confident that they would remember.
That mismatch between feeling and result matters for AI tools. A chatbot that summarizes a chapter can produce the comfortable feeling of understanding, while one that asks questions forces the retrieval that the experiment tied to long-term memory.
What studies of chatbots in studying found
Ibnatul Jalilah Yusof compared final exam scores in a Measurement and Evaluation in Education course in Malaysia. Earlier cohorts studied without ChatGPT, and the current cohorts used it to generate objective practice questions and get immediate feedback. With 64 randomly selected students per group, the article in the Journal of Information Technology Education: Research (2025) reports significantly better exam results for the ChatGPT group. The published record gives no group averages that could be checked here, and different cohorts may differ in ways unrelated to the tool.
A second study points the other direction. André Barcaui randomly split 120 business undergraduates in Brazil into two groups of 60. Both researched basic AI concepts and gave a short presentation; one group used ChatGPT, the other used notes, academic databases, and search engines. In a surprise 20-question test 45 days later, 85 students returned. As reported in PsyPost's coverage of the paper in Social Sciences and Humanities Open, the search-and-notes group averaged 68.5 percent and the ChatGPT group 57.5 percent. The ChatGPT group also spent less time, 3.2 hours against 5.8 hours. About 30 percent of participants did not return, and all students came from one university.
Which way of using AI matches which evidence
The two studies used the chatbot differently, which is the most useful way to read them side by side. The table is a reading aid built from the sources, not a ranking of products.
| How the AI is used | What the learner does | Closest evidence |
|---|---|---|
| Generates practice questions with feedback | Answers from memory, then checks | Testing effect (2006); Malaysian cohort study (2025) |
| Researches and explains a topic | Reads the answer, prepares a presentation | Brazilian randomized trial: lower retention at 45 days |
| Summarizes a chapter | Rereads the summary | No direct trial found; rereading lost to testing in the 2006 study |
Ongoing work may sharpen this picture. A registry entry (AEARCTR-0018528) describes a Virginia Tech trial in a macroeconomics course with 141 consenting students and three arms: retrieval practice without AI, retrieval practice with Microsoft Copilot, and a placebo reading task. The registry listed no results when checked on October 27, 2026.
A study routine that fits the evidence
- Read or review the material first, then close it.
- Ask the tool for a set of questions on that material and answer each from memory before looking anything up.
- Have the tool mark the answers, then verify any disputed point against the textbook or class notes, since chatbots can state errors fluently.
- Repeat the questions a few days later and again after a week, instead of one long session.
- Record the questions missed and ask for new ones on those topics only.
- Check the school's or instructor's AI policy before using any tool on graded work.
This article summarizes research and does not replace guidance from teachers or learning support staff. Students with documented learning needs may need different strategies, and that decision belongs with those staff.
FAQ
Does using AI for studying hurt learning?
It can, depending on use. The Brazilian randomized trial reported lower retention after ChatGPT-assisted research, while the Malaysian study linked chatbot-generated practice questions to better exam results. Both have limits, so neither settles the question.
What is the testing effect?
It is the finding that retrieving information from memory improves later recall more than restudying the same material. Roediger and Karpicke showed it with students reading prose passages in 2006.
Is it safe to trust an AI study helper's answers?
Not without checking. The sources reviewed here do not measure accuracy, so every claim a tool makes should be compared with course materials before it is memorized.
Are there trials of AI retrieval practice with random assignment?
One was registered at Virginia Tech for spring 2026 but had no published results as of October 27, 2026.
Related reading
- Building a study schedule with an AI chatbot
- How AI is used in learning
- AI and English language learning
- Artificial intelligence in education
More in AI Basics.
Last verified: October 2026
Sources
- Roediger HL and Karpicke JD, Test-Enhanced Learning, Psychological Science, 2006
- Yusof IJ, ChatGPT-Assisted Retrieval Practice and Exam Scores: Does It Work?, Journal of Information Technology Education: Research, 2025
- Barcaui A, ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention, Social Sciences and Humanities Open, 2025, as summarized by PsyPost
- Investigating the Retrieval Effect with and without AI Support, AEA RCT Registry AEARCTR-0018528, 2026