AI and learning in numbers: statistics with their sources
The most quoted number is this: students who practised maths with a plain chatbot scored 17% lower on the exam than students with no AI (Bastani and others, PNAS, 2025, about 1,000 students, peer reviewed). This page lists that figure and sixteen more. Each has its source, year, sample size and status. Facts were last checked on 13 November 2026.
How to quote these
Take the whole row, not just the number. The glossary explains each status.
Experiments on learning
| Number | What it is | Source and date | Sample | Status |
|---|---|---|---|---|
| 17% lower | Exam scores after practising with plain GPT-4, against students with no AI | Bastani and others, June 2025 | About 1,000 high school students | Peer reviewed |
| 48% and 127% higher | Practice scores with plain GPT-4 and with a hint-only tutor, against no AI | Bastani and others, June 2025 | About 1,000 high school students | Peer reviewed |
| 57% against 74% | Problems solved alone after 10 to 15 minutes of AI help, against no AI (experiment 1) | Liu, Christian and others, 2026 | 1,222 adults across 3 experiments | Peer reviewed (COLM 2026) |
| 20% against 11% | Problems skipped after AI help, against no AI (experiment 1) | Liu, Christian and others, 2026 | 1,222 adults across 3 experiments | Peer reviewed (COLM 2026) |
| More than double | Median learning gain with a designed AI tutor, against an active class | Kestin and others, June 2025 | 194 university physics students | Peer reviewed |
| 0.37 standard deviations lower | Final grades with access to a course AI tutor | Liu and others, University of Maryland, October 2026 | 2,379 undergraduates | Working paper |
| 15 of 18 | People who wrote with ChatGPT and could not correctly quote their own essay, in one session | Kosmyna and others, June 2025, revised December 2025 | 54 adults | Preprint |
Work, coding and creativity
| Number | What it is | Source and date | Sample | Status |
|---|---|---|---|---|
| 19% slower, felt 20% faster | Experienced developers on real tasks with early-2025 AI tools | Becker and others (METR), 2025 | 16 developers, 246 tasks | Preprint |
| About 50% against 67% | Quiz scores after a coding task with AI, against without | Shen and Tamkin, January 2026 | 52 developers | Preprint |
| 8.1% more novel | Short stories written with five AI ideas, which were also more similar to each other | Doshi and Hauser, July 2024 | 293 writers, 600 judges | Peer reviewed |
Records and surveys, which cannot show cause
| Number | What it is | Source and date | Sample | Status |
|---|---|---|---|---|
| 18% up, 20% down | Homework scores and monthly exam scores after students took up AI, as reported | Stromberg, Lei and Wu, June 2026 | 26,811 students over 30 months | Working paper, school records |
| 28.4% to 22.4% | Doctors' detection rate in colonoscopies without AI, before and after AI was introduced | Budzyn and others, August 2025 | 1,443 colonoscopies, 4 centres | Peer reviewed, observational |
| r = -0.68 | Link between AI use and critical thinking scores | Gerlich, January 2025, corrected September 2025 | 666 people in the UK | Peer reviewed, survey |
| About 60% | Work examples (555 of 936) where workers reported some critical thinking when using AI | Lee and others, 2025 | 319 knowledge workers | Peer reviewed, survey |
| About 47% | University student conversations with one chatbot that were direct requests for answers | Anthropic Education Report, April 2025 | 574,740 conversations | Company report |
| 54% | US teenagers aged 13 to 17 who have used chatbots for help with schoolwork | Pew Research Center, February 2026 | 1,458 US teens | Survey |
The number that is zero
| Number | What it is | Source and date |
|---|---|---|
| 0 | Studies we found that measured IQ before and after people started using chatbots | Our own search, last checked 13 November 2026 |
The nearest thing is a preprint with 30 people (Benitez and others, 2025 to 2026). It found no change on four subtests of an adult intelligence scale after four weeks of ChatGPT use. That is too few people to prove anything either way. See the full answer on IQ.
What these numbers do not show
- No long-term effect from an experiment. The experiments here last from minutes to weeks. The only 30-month source is school records, which cannot prove cause.
- No harm from AI as such. One tutor here more than doubled learning gains. Another went with lower grades.
- No final word. Six of the seventeen rows are preprints, working papers or a company report. Their numbers can change.
One warning. The 19% developer figure describes 16 people in 2025. The same group's 2026 data points to a possible speedup, and its own authors call it unreliable.
A fair one-line summary. "In experiments, AI that hands over answers tends to go with weaker results once the AI is gone, and AI that guides tends to do better. Results are mixed, and no study has measured IQ."
What to do with this
Quote any row with its study, year, sample size and status. For yourself, the pattern is short: try first, then ask for hints.
Questions
Why are some well-known figures missing?
We list figures we could trace to a paper, its official summary or the publisher's page. The one exception is the school records row, which is marked "as reported" because we have not yet read those figures in the paper itself. Other numbers we saw only in news stories are left out.