Does AI help students learn, or just help them do the work?

The short answer

What we found

We looked at 14 ways students use AI, from asking it for answers to practising a language, against 65 sources. These include randomised trials, reviews of many studies, UK regulators’ and government documents, and usage surveys. We judged each use on one question: did students learn more, measured when they no longer had the AI? Whether the work itself got better is shown separately.

Better work does not necessarily mean more learning.

• When AI gave feedback, drafted writing, supplied answers or wrote code, the work usually improved.
• Learning measured without AI usually did not improve, and in one trial it fell.
• This is not universal: in some university studies, students who had AI access while learning a new topic did better on later tests taken without it.
• More

How a student uses AI matters more than whether they use it.

• The same AI model harmed exam results when it gave answers, and did not harm them when it was set up to give hints.
• In one university course, encouraging students to use a course chatbot made no measurable difference. Showing them how to use it as a tutor raised exam results.
• (Our reading across studies.)
• More

Answer-giving is the clearest risk.

• In a peer-reviewed trial with nearly 1,000 secondary maths students, those who practised with an answer-giving version of GPT-4 scored 48% higher during practice and 17% lower on a later exam without AI.
• Other studies point the same way.
• It is still one main school trial, so we rate the evidence as Moderate, not settled.
• More

Structured AI tutoring is the most promising positive use, but the evidence is only moderately mature.

• Tutors that guide rather than tell improved learning in trials in the US, Nigeria and Sierra Leone.
• Two other scaffolded tutors produced no learning gain.
• Against good existing maths software, an AI tutor added little.
• More

Using a chatbot to explain new material looks positive, but it is early.

• University experiments found gains on tests taken without AI.
• Those studies have not yet completed peer review.
• In English secondary schools, using a chatbot instead of making notes led to worse recall.
• More

Before believing any AI-and-learning result, check whether AI was available during the test.

• Where a review of programming studies could tell, the learning benefit was close to zero when AI was barred from the test, and large when it was allowed.
• Many studies don’t say.
• More

Students can prefer, and rate as more helpful, the method that teaches them less.

• English 14–15-year-olds rated a chatbot as more helpful than taking their own notes, and remembered less with it.
• Maths students who used answer-giving AI didn’t realise they had learned less.
• This is a “can”, not a rule: in other studies students’ sense of what helped was accurate.
• More

No use of AI for learning reaches our highest level of evidence maturity.

• No positive or negative finding yet has two independent, peer-reviewed school-level trials pointing the same way, with learning measured without AI.
• Most of the evidence comes from universities and adults, and there is almost none from primary schools.
• More

Three questions to ask before believing an AI-and-learning claim

You will see headlines, school announcements and product claims saying AI improves learning, or damages it. Three questions sort most of them out.

1. Was AI available during the test?

• If students could use AI in the test, a higher score may show what the AI can do, not what the student learned.
• Many studies don’t say. Where one review of programming studies could tell, the apparent benefit disappeared when AI was barred from the test.

2. Compared with what?

• An AI tutor that beats “no extra lessons” is not the same as one that beats a good teacher or existing software.
• One AI maths tutor looked very different against no help than against good non-AI practice software.

3. When was learning measured?

• A test straight after a lesson shows what students can recall that day. A test days or weeks later is a better guide to what they learned.
• Most studies only test immediately.

Further down, the methodology adds four more checks: peer review, independence from the company that made the tool, study design, and whether the students studied resemble the ones you care about.

Which uses of AI help students learn? The matrix

Learn more? means learning measured when students no longer had the AI. Better work? means the quality of the work produced with AI’s help. AI in test? records whether students could use AI during the assessment in the main studies.

Two labels are kept separate on purpose. The assessment is what the evidence currently suggests. The evidence maturity is how developed and reliable that evidence is (High, Moderate or Early). Something can look risky on evidence that is still early. How we rate the evidence

What the AI does	Better work?	Learn more?	AI in test?	Assessment	Evidence maturity	Who was studied
Gives the answer or solves the problem	Yes	No: lower exam results once AI was removed, in the clearest trial	No	Poor fit / caution	Evidence maturity: Moderate	Secondary; university; adults
Tutors with hints and steps, within a course	Yes	Yes in structured programmes; small or none against good existing software	Mostly no; not stated in one key trial	Promising	Evidence maturity: Moderate	Secondary; university
Explains new material in open chat	Yes	Positive in university studies not yet peer reviewed; worse than making notes in a school trial	Mostly no; not stated in one	Emerging	Evidence maturity: Early	University; adults; secondary
Gives feedback on your work	Yes	Little or no difference	No	Weakly evidenced	Evidence maturity: Moderate	Secondary; university
Writes or drafts your work	Yes	No difference in knowledge or transfer	Varies	Poor fit / caution	Evidence maturity: Early	University
Summarises reading or notes	Not measured	No: own notes remembered and understood better	No	Poor fit / caution	Evidence maturity: Moderate	Secondary (England)
Replaces search and study materials	Easier; reasoning weaker	No: less recalled, shallower learning	No	Poor fit / caution	Evidence maturity: Moderate	University; adults
Writes practice questions	Quicker to make; easier, less accurate	Not tested	—	Weakly evidenced	Evidence maturity: Early	University (item quality only)
Checks your answers	Wrong answers often accepted	Not measured	—	Poor fit / caution	Evidence maturity: Early	Secondary; university (small studies)
Finds sources and references	Some references made up or wrong	Not measured	—	Poor fit / caution	Evidence maturity: Early	No students studied (audits of AI output)
Plans study, prompts reflection	—	No advantage found	—	Insufficient to classify	Evidence maturity: Early	Secondary; university
Writes or fixes code while you learn to program	Yes	No overall; about zero when AI was barred from the test	Key factor	Poor fit / caution	Evidence maturity: Moderate	Ages 10–17; university; professionals
Language practice	—	Moderate immediate gains, mostly with older chatbots	Not reported	Emerging	Evidence maturity: Early	Mixed ages (not separated)
Support for SEND	—	Not known	—	Insufficient to classify	Evidence maturity: Early	Secondary; university (very small studies)

Assessments are editorial judgements made against stated rules, not scientific scores. The counts depend on how the uses are divided.

What the evidence says, in four groups

The same 14 uses as the matrix, in plain English. Each appears once, in the group that best matches the evidence on learning, with the same two labels: the assessment and the evidence maturity.

1. The most promising way to use AI for learning

An AI tutor that gives hints and steps rather than answers, used within a course Promising Evidence maturity: Moderate

• Improved learning measured without AI in a US university physics course, in after-school sessions in Nigeria, and in schools in Sierra Leone.
• The conditions matter: structure, a teacher or programme around it, and students actually using it.
• Against good non-AI maths practice software, or where students rarely used it, the gain was small or zero.
• Do AI tutors work?

2. Can improve the work without improving learning

Feedback on your work Weakly evidenced Evidence maturity: Moderate

• AI feedback reliably helped students improve the piece of work in front of them.
• In the same studies it did not improve their performance on a new task or their knowledge.
• Does AI feedback improve your writing?

The same gap between better work and learning appears in several uses in the next group: answers, drafting and coding. There, the evidence also shows learning falling or failing to develop. In the matrix above, the “Better work?” and “Learn more?” columns show the gap row by row.

3. Can get in the way of learning

The evidence for these uses points to a risk to learning, or to the quality of the work, without good checks. It does not mean every use is harmful. Some of these ratings rest on early evidence and may change.

Getting AI to give you the answer Poor fit / caution Evidence maturity: Moderate

• Practice scores rose, and later exam results without AI fell, in the clearest trial.
• Is it bad to use AI for homework?

Getting AI to write or draft work you are meant to learn to produce Poor fit / caution Evidence maturity: Early

• The work improved while knowledge and transfer did not.
• In UK qualifications, submitting AI-produced work as your own is malpractice.
• Is it bad to use AI for homework?

Getting AI to summarise your reading instead of making notes Poor fit / caution Evidence maturity: Moderate

• In a trial in seven English schools, pupils who took their own notes remembered and understood more three days later.
• Should students use AI to summarise their notes or reading?

Using AI instead of searching and studying sources yourself Poor fit / caution Evidence maturity: Moderate

• Students learned less deeply and recalled less in several small studies.
• Is AI a good replacement for search?

Using AI to check your answers Poor fit / caution Evidence maturity: Early

• Students often accepted wrong AI judgements as correct.
• Can AI check your homework answers?

Asking AI to find sources and references Poor fit / caution Evidence maturity: Early

• Chatbots have made up references, though newer models do so less often.
• Can AI find sources for an essay?

Getting AI to write or fix code while you are learning to program Poor fit / caution Evidence maturity: Moderate

• Exercise scores rose, but learning did not when AI was taken away for the test.
• Does AI help you learn to code?

4. The evidence isn’t good enough yet

Early, but positive

Asking a chatbot to explain new material and answer your follow-up questions Emerging Evidence maturity: Early

• University experiments found gains on tests taken without AI, some still there a week later.
• None has completed peer review.
• One peer-reviewed school study found it worse than making your own notes.
• Does asking ChatGPT or another chatbot to explain things help you learn?

Language practice with a chatbot Emerging Evidence maturity: Early

• Reviews of chatbot language practice find moderate gains.
• Most of the chatbots studied came before today’s AI models.
• Can AI help you learn a language?

Untested or too little evidence

AI-generated practice questions and flashcards Weakly evidenced Evidence maturity: Early

• Testing yourself is one of the best-evidenced ways to learn.
• Whether AI-written questions work as well has not been tested, and they were easier and less accurate than human-written ones.
• Is AI good for revision questions?

AI to plan study or prompt you to reflect Insufficient to classify Evidence maturity: Early

• Can AI help you plan your studying?

AI for students with special educational needs and disabilities Insufficient to classify Evidence maturity: Early

• No robust outcome evidence yet.
• Does AI help students with SEND?

Where the evidence comes from: by age and stage

Most of the evidence about AI and learning comes from universities and adults. Very little comes from primary schools or further education colleges, and UK pupil-level evidence is thin.

What the AI does	Primary	Secondary (incl. sixth form)	FE college	University	Adults (not students)	UK study
Gives the answer	–None found	●At least one controlled study	–None found	●At least one controlled study	●At least one controlled study	–None found
Tutors with hints and steps	○Limited or indirect evidence only	●At least one controlled study	–None found	●At least one controlled study	–None found	○Limited or indirect evidence only
Explains in open chat	–None found	●At least one controlled study	–None found	●At least one controlled study	●At least one controlled study	●At least one controlled study
Feedback	–None found	●At least one controlled study	–None found	●At least one controlled study	–None found	–None found
Writes or drafts	–None found	–None found	–None found	●At least one controlled study	○Limited or indirect evidence only	–None found
Summarises	–None found	●At least one controlled study	–None found	–None found	–None found	●At least one controlled study
Replaces search	–None found	–None found	–None found	●At least one controlled study	●At least one controlled study	–None found
Practice questions	–None found	–None found	–None found	○Limited or indirect evidence only	–None found	–None found
Checks answers	–None found	○Limited or indirect evidence only	–None found	○Limited or indirect evidence only	–None found	–None found
Finds sources	–None found	–None found	–None found	–None found	–None found	–None found
Plans study	–None found	●At least one controlled study	–None found	○Limited or indirect evidence only	–None found	–None found
Coding while learning	–None found	●At least one controlled study	–None found	●At least one controlled study	●At least one controlled study	–None found
Language practice	Not separated	Not separated	Not separated	Not separated	Not separated	–None found
SEND support	–None found	○Limited or indirect evidence only	–None found	○Limited or indirect evidence only	–None found	○Limited or indirect evidence only

Key:

• ● at least one controlled study that measured learning;
• ○ only small, uncontrolled or indirect evidence (for example, AI helping human tutors rather than pupils, or a survey);
• – none found;
• “Not separated”: the reviews do not report results by age.

Notes:

• The primary-school entry is a US study in which AI made suggestions to human tutors, not to pupils.
• The UK entries are:
  • a trial in seven English schools (explaining and summarising);
  • a small study in England, written by the companies involved, in which human tutors reviewed AI-drafted messages;
  • a survey of disabled university students.
• The “Secondary” column includes England, Germany, Estonia, the US, Turkey, Nigeria and Sierra Leone. School systems and ages differ.
• The secondary entry for planning study compares versions of AI, not AI against none.

Does AI help weaker students? Who benefits, and who may lose out

There is not yet good evidence that AI closes attainment gaps. In two tutoring trials that examined prior attainment, stronger students gained more.

• Nigeria (after-school sessions with Microsoft Copilot, guided by teachers; working paper):
  • Gains were positive at every level of prior attainment, and larger for students who had done better before.
  • The study also reported a larger effect for girls.
• Sierra Leone (Google’s Gemini tutor; vendor report, independently assessed):
  • Students who started with stronger maths skills gained more.
  • Differences by gender and reading level were not statistically significant.
• Two university studies of AI access while learning (both preprints) pointed the same way:
  • In one, gains were concentrated among higher-attaining students.
  • In a programming study, students with less prior knowledge understood less when given AI.

There are exceptions, and the picture is not one-directional:

• In a German university experiment, AI help seemed most useful to students who started with less knowledge, though the authors’ evidence for this was weak.
• In the Turkish maths trial, lower-attaining students gained most during practice with answer-giving AI. On the later exam, the harm did not differ by prior attainment.
• In the Harvard physics trial and the US Khanmigo trial, the effects did not clearly differ by prior attainment.

One finding to watch: at the University of Maryland, losses in course grades linked to a course AI assistant were more than twice as large for first-generation students as for others (working paper; grades, not an independent test).

Why it may happen (our interpretation, not tested): students who already know more may be better at asking useful questions and spotting wrong answers. Students who know less may be more likely to let AI do the work.

What are the UK rules on AI coursework? UK rules and evidence

Under the exam boards’ rules (JCQ), using AI to produce coursework or other assessed work that you submit as your own is malpractice, and AI-produced content that students acknowledge earns no marks (ST61). Ofqual, the exams regulator, has said that using AI in coursework deprives students of the intended learning experience, because the AI does the work instead of the student (ST60). Separately, the Department for Education sets design standards for AI tools used in schools in England (ST53). These rules point in the same direction as some of the evidence on this page, but they were not derived from it, and evidence on UK pupils’ learning remains thin.

Does ChatGPT help students learn? The evidence, use by use

It depends on what students use it for. In the clearest school trial, secondary maths students who practised with a ChatGPT-like version of GPT-4 scored 17% lower on a later exam taken without AI, while a version set up to give hints rather than answers avoided that harm. In university experiments, students who used AI chatbots such as GPT-4o to learn new material did better on later tests without AI, but those studies have not yet completed peer review. In English secondary schools, using a chatbot instead of making notes led to worse recall. The sections below take each use in turn.

Is it bad to use AI for homework?

Whether using AI for homework harms your learning depends on what the AI does. When it gives you the answer or produces the work you are meant to learn to produce, the evidence points to a risk: the homework improves and the learning doesn’t. In the clearest trial, secondary maths students who practised with an answer-giving version of GPT-4 did worse on a later exam taken without AI.

In a trial at a Turkish high school, nearly 1,000 students practised maths over four sessions:

• with a ChatGPT-like interface;
• with a tutor version told to give hints rather than answers;
• or with their textbook and notes.

Practice scores rose by 48% with the ChatGPT-like version and by 127% with the tutor version. On a later exam without AI:

• the ChatGPT-like group scored 17% lower than the textbook group;
• the tutor group scored about the same as the textbook group.

The authors: “These negative learning effects are largely mitigated by the safeguards in GPT Tutor.” (ST07)

Other studies point the same way, with weaker designs:

• In university programming experiments, students who asked an AI for solutions understood less than those who asked for explanations. This was an exploratory analysis.
• At the University of Maryland, a course AI assistant, mostly used to get information, explanations or solutions, coincided with lower final grades. This was a working paper, and grades are not an independent test.

For written work, AI drafting improved essays without improving what students knew or could do on a new task. In UK qualifications, AI-produced work submitted as your own is malpractice (see the UK box).

This does not mean every use of AI for homework harms learning: see using a chatbot to explain.

Study cards

Bastani et al. (2025), PNAS · ST07

• Who: nearly 1,000 students, grades 9–11, one high school in Turkey; maths.
• Compared with: textbook and notes.
• AI during test: no; practice was with AI, the exam was without.
• Measured: practice sessions, then a later exam.
• Status: peer reviewed; independent.
• AI: GPT-4, autumn 2023. Its own answers were correct only 51% of the time.
• Found: practice +48% (ChatGPT-like) and +127% (tutor version). Exam −17% (ChatGPT-like); tutor version no different from control.
• Limits: one school; four sessions; the tutor version’s instructions contained the solutions.

Lehmann, Cornelius & Sting, “AI Meets the Classroom” · ST33

• Who: university students, Germany and the Netherlands; programming; lab and field studies.
• Compared with: no AI.
• AI during test: no.
• Status: preprint; independent.
• AI: GPT-3.5, 2023–24.
• Found: no overall effect on understanding. Solution-seeking was associated with lower understanding, explanation-seeking with higher. Students with less prior knowledge understood less with AI.
• Limits: the use-type comparison is exploratory, not randomised.

Liu et al. (2026), University of Maryland · ST09

• Who: 2,379 students, 30 instructors; mixed courses.
• Compared with: the same courses without the course AI assistant. Students in both groups could use ChatGPT or Gemini.
• AI during test: course grades; not controlled.
• Status: working paper; the tool was built by the university.
• AI: GPT-4o, autumn 2025.
• Found: final grades −0.37 SD in the preferred sample (−0.27 SD, not significant, in the full sample). 73.8% of requests sought information, explanations or solutions.
• Limits: only about 15% used it; grades rather than an independent test.

Fan et al. (2025), British Journal of Educational Technology · ST22

• Who: 117 university students learning English; essay revision.
• Compared with: a human expert, a checklist, and no support.
• AI during test: no; knowledge and transfer were tested after support was withdrawn.
• Status: peer reviewed.
• Found: essay scores improved most with ChatGPT. There was no difference in knowledge gained or in transfer to a new task.
• Limits: single session; short tests.

Do AI tutors work?

AI tutors sometimes improve learning, under the right conditions. AI tutors that guide students through problems, rather than giving answers, improved learning measured without AI in several trials. This was usually as part of a structured course with a teacher or programme around it.

The gains were small or absent:

• against good non-AI practice software;
• when students rarely used the tutor;
• in two scaffolded tutors that produced no learning gain at all.

Promising Evidence maturity: Moderate

Where it worked:

• Harvard physics: students who used an AI tutor built by the course team scored higher on an immediate test than after an in-class active-learning lesson (about 0.63 standard deviations). They also reported more engagement. The paper does not say whether AI was available during that test.
• Nigeria: six weeks of after-school sessions using Microsoft Copilot, guided by teachers, raised English results by about 0.31 SD against students who had no sessions. The extra time and supervision may account for part of that.
• Sierra Leone: Google’s Gemini tutor raised maths results by about 0.26 SD. The trial was pre-registered and the outcome independently assessed, but the report is Google’s own and not peer reviewed.

Where it didn’t:

• Khanmigo (Tennessee): over two years, Khan Academy’s AI tutor added about 0.06–0.08 SD a year compared with existing non-AI maths software, similar to Khan Academy practice without AI. The authors: “The binding constraint appears to be engagement: realizing the promise of AI tutoring will require getting students to use it, not just giving them access.” (ST10)
• NUMI: a guardrailed maths tutor improved students’ next attempt after a mistake. A week later there was at most a marginal gain on practised questions and none on new ones.
• Iris (Germany): in a university programming course, neither a hint-first tutor nor ChatGPT improved learning, though both raised exercise scores. The authors found generative AI “acted primarily as a performance aid rather than a learning enhancer”. (ST17)

Guardrails can prevent harm: in the Turkish trial, the hint-giving tutor avoided the exam penalty seen with answer-giving AI. But guardrails alone did not produce a learning gain there.

Study cards

Kestin et al. (2025), Scientific Reports · ST08

• Who: 194 Harvard physics students.
• Compared with: an in-class active-learning lesson.
• AI during test: not stated.
• Measured: immediately after.
• Status: peer reviewed; the tutor was built by the authors.
• AI: GPT-4, autumn 2023.
• Found: +0.63 SD. Higher engagement and motivation.
• Limits: elite sample; two lessons; no delayed test.

De Simone et al. (2025), World Bank · ST12

• Who: 1,328 first-year senior secondary students randomised (759 completed), Nigeria; English.
• Compared with: no sessions.
• AI during test: no (paper test).
• Status: working paper; Microsoft Copilot used, no Microsoft funding found.
• Found: +0.31 SD overall. Gains at every level of prior attainment, larger for higher attainers.
• Limits: extra time and supervision were not separated from AI; drop-out differed between groups (64% vs 50%).

Google LearnLM & Fab AI (2026), Sierra Leone · ST13

• Who: 1,763 students, grades 7–8; maths.
• Compared with: usual provision.
• AI during test: no; blind independent assessment.
• Status: vendor report (Google), pre-registered, not peer reviewed.
• Found: +0.26 SD. Larger gains for students with stronger starting maths.
• Limits: run by the developer.

Oreopoulos & Low (2026), Khanmigo · ST10

• Who: 18 Tennessee schools, grades 6–8; remedial maths.
• Compared with: existing non-AI practice software.
• AI during test: no (standardised MAP test).
• Status: working paper; Khan Academy facilitated the trial.
• Found: about 0.06–0.08 SD a year.
• Limits: low use; Khanmigo cannot be separated from the wider Khan Academy package.

Oreopoulos, Liut, Sungu & Low (2026), NUMI · ST11

• Who: 6,997 students, grades 6–8; maths.
• Compared with: the same platform without the tutor.
• Status: working paper.
• Found: a better next attempt after mistakes. One week later, +3.2 percentage points on practised items (marginal) and no gain on new items.
• Limits: single session.

Bassner et al. (2026) · ST17

• Who: 275 first-year computing students, Technical University of Munich.
• Compared with: no AI, ChatGPT, and a hint-first tutor.
• Status: peer reviewed; the tutor was built by the authors’ group.
• Found: both AI groups raised exercise scores; neither improved knowledge or code comprehension.
• Limits: we read the abstract only.

Also relevant:

• Wang et al., Tutor CoPilot (ST15, working paper): AI suggestions to human tutors raised lesson mastery but not end-of-year test results.
• LearnLM/Eedi (ST14, preprint, written by the companies involved): in England, Years 9–10, human tutors reviewing AI-drafted messages were followed by better success on the next topic than human tutors alone.

Does asking ChatGPT or another chatbot to explain things help you learn?

Using an AI chatbot to explain new material may help you learn, especially if you are trying to understand rather than to get the answer, but the evidence is early.

• In university experiments, students given AI while learning a new topic did better on later tests taken without it.
• None of these studies has yet completed peer review.
• The peer-reviewed evidence shows something narrower: explanations beat bare answers.
• In English secondary schools, using a chatbot instead of making your own notes led to worse recall.

Emerging Evidence maturity: Early

The university studies:

• Middlebury College (US): students given GPT-4o while learning unfamiliar topics scored 0.27 SD higher on an unaided test. A week later, about three-quarters of that gain remained. The authors: “AI access raises immediate test scores by 0.27 standard deviations. These gains persist one week later.” (ST19)
• Technical University of Berlin: students given unrestricted AI while studying a textbook scored 0.34 SD higher on an unaided test. Unrestricted access also did better than access delayed until after 10 minutes’ reading, by 0.21 SD (p = 0.066).
• A university in Chile: encouraging econometrics students to use a course chatbot made no measurable difference to exam results. Showing them how to use it as a tutor raised final-exam scores by 0.22 SD. The author: “The findings suggest that learning gains arise less from adoption than from guiding how students use course-specific AI assistants.” (ST20) Whether AI was available in the exams is not stated.

The peer-reviewed studies:

• In two pre-registered online experiments with 1,818 US adults, AI explanations of maths problems led to better unaided performance than seeing only the answers, especially when people tried the problem first.
• In another adult study, AI-written hints helped at least as much as human-written ones on an immediate test.
• In the school trial in England, using a chatbot alone led to worse recall and understanding than taking notes. Using it alongside note-taking did about as well as notes alone (see below).

What we don’t know:

• whether these gains hold for school pupils, over longer periods, or with less motivated students;
• why unrestricted AI helped in some university studies and harmed in the Turkish school trial.

The difference in tasks may matter: learning new material with a textbook to hand, versus practising problems the AI can simply solve. This is the most important open question on this page.

Study cards

Contractor & Reyes (2026) · ST19

• Who: 211 students (204 completed), Middlebury College; unfamiliar topics plus an essay.
• Compared with: no AI.
• AI during test: no, for both tests and the essay a week later.
• Status: preprint (also an IZA discussion paper); independent.
• AI: GPT-4o, spring 2025.
• Found: +0.27 SD immediately; +5.1 percentage points a week later; unaided essays improved. Gains were concentrated in higher-attaining students.
• Limits: one selective college; one-week window. The split by type of use is descriptive, not causal.

Fischer, Rau & Rilke (2025) · ST18

• Who: 334 students, TU Berlin; economics textbook material.
• Compared with: textbook only, and AI available only after 10 minutes’ reading.
• AI during test: no.
• Status: working paper; tool from acemate.
• Found: +0.34 SD vs textbook only.
• Limits: single lab session; immediate test.

Gallegos (2026), “Guidance Over Adoption” · ST20

• Who: about 300 econometrics students in each of two experiments, Chile.
• Compared with: no encouragement; no guidance.
• AI during test: not stated.
• Status: working paper, submitted.
• Found: encouragement +0.05 SD (not significant); guidance on tutor-style use +0.22 SD on the final exam, with the pass rate up 12.8 percentage points.

Kumar, Rothschild, Goldstein & Hofman (2025), AIED 2025 · ST16

• Who: 1,818 US adults online, two pre-registered experiments; high-school maths.
• Compared with: seeing only correct answers.
• AI during test: no.
• Status: peer reviewed (conference proceedings); Microsoft Research authors.
• Limits: adults, not students in courses; we read the abstract only.

Pardos & Bhandari (2024), PLOS ONE · ST21

• Who: 274 US adults online; algebra and statistics.
• Found: immediate gains of +17 percentage points with ChatGPT hints, +11.6 with human hints and +1.9 with none.
• Limits: immediate test on the same items; 32% of raw AI hints failed quality checks.

Does AI feedback improve your writing?

AI feedback improves the piece you are revising, but there is little evidence that it improves your writing on the next piece. In two trials, AI feedback raised the quality of the revised work but made no measurable difference to performance on a new task, to knowledge, or to transfer.

Weakly evidenced Evidence maturity: Moderate

• German Grade 10 English learners: GPT-3.5 feedback improved revisions slightly (d = 0.19). On a new writing task the difference was smaller and not statistically significant. Students rated the feedback as very useful, far more than its measured effect.
• University English learners: ChatGPT support raised essay scores more than a human expert or a checklist, with no difference in knowledge or transfer.
• Human versus AI feedback: in a small study, AI and human tutor feedback produced similar writing gains. In another, trained humans’ feedback was rated higher on four of five criteria (no learning measured).

Study cards

Meyer et al. (2024), Computers and Education: AI · ST23

• Who: 459 Grade 10 students, Germany; English writing.
• Compared with: no feedback.
• AI during test: no (new task).
• Status: peer reviewed.
• AI: GPT-3.5, 2023.
• Found: revision d = 0.19; new task d = 0.13 (not significant); perceived usefulness d = 1.19.
• Limits: the control group had no feedback at all.

Fan et al. (2025) · ST22. See above.

Escalante, Pack & Barrett (2023) · ST24

• Who: 48 university students; English writing.
• Compared with: human tutor feedback.
• Status: peer reviewed.
• Found: no significant difference.
• Limits: small; not randomised.

Steiss et al. (2024), Learning and Instruction · ST25

• What: 200 essays from grades 6–12, US.
• Found: trained human feedback was rated higher on four of five criteria.
• Limits: no learning outcome.

Should students use AI to summarise their notes or reading?

The only UK trial we found suggests not, if the AI summary replaces making your own notes. In seven English schools, 14–15-year-olds who took their own notes remembered and understood more three days later than those who used a chatbot alone. Using the chatbot and taking notes did about as well as notes alone.

Poor fit / caution Evidence maturity: Moderate

This rests on one strong trial with an older model, plus long-standing learning research showing that summarising and rereading are among the less effective ways to study (ST39).

What students preferred. Pupils were asked in the first session, before some were excluded from the analysis:

• In the first comparison (212 pupils), 42% preferred the chatbot and 27% their own notes. The rest had no preference or weren’t sure.
• In the second (198 pupils), 51% preferred chatbot-plus-notes, the more effective of the two, and 16% the chatbot alone.

The authors report that pupils “perceived it as more helpful”. (ST28) Their recall says otherwise.

Study card

Kreijkes et al. (2026), Computers & Education · ST28

• Who: 344 Year 10 pupils analysed (405 recruited), seven English schools; history reading.
• Compared with: own note-taking; chatbot plus notes.
• AI during test: no; tested three days later.
• Status: peer reviewed; pre-registered; Microsoft Research co-authors, with Cambridge University Press & Assessment.
• AI: GPT-3.5, limited to 20 prompts.
• Found: notes vs chatbot alone: retention d = 0.44, comprehension d = 0.38, free recall d = 0.21. Chatbot plus notes was close to notes.
• Limits: older model; one reading task.

Is AI a good replacement for search and study materials?

Studies so far suggest that using AI instead of searching and studying sources yourself makes the work easier but the learning shallower. Students who used ChatGPT instead of conventional resources gave weaker reasoning and recalled less, in several small studies.

Poor fit / caution Evidence maturity: Moderate

• Brazil: university students who used GPT-4 instead of notes, library and web search scored 57.5% vs 68.5% on a surprise test 45 days later. Nearly a third of participants dropped out.
• Germany: students researching a science question with ChatGPT rather than Google found it easier but gave weaker justifications.
• Australia: first-year programming students using ChatGPT scored higher on the task (89% vs 69%) but recalled less of it 48 hours later (39% vs 52%) than those using web search. This is a preprint.
• Adults: in seven experiments with more than 10,000 adults, people who learned about a topic from an AI chatbot rather than web search reported shallower learning. No knowledge test was used.

Study cards

• Barcaui (2025) · ST30
  • 120 randomised, 85 analysed; Brazil; business.
  • AI during test: no (surprise, proctored).
  • Peer reviewed.
  • Limits: 29% drop-out; one institution.
• Stadler, Bannert & Sailer (2024) · ST29
  • 91 university students; Germany; compared with Google.
  • Peer reviewed.
  • Limits: no learning test afterwards.
• Bergh, Tag, Vassar & Renzella (2026) · ST32
  • 55 students; UNSW; ChatGPT-4.5.
  • Recall tested without access to the code.
  • Preprint.
• Melumad & Yun (2025), PNAS Nexus · ST31
  • 10,462 adults.
  • Peer reviewed.
  • Limits: self-reported learning only.

Is AI good for revision questions and flashcards?

Testing yourself works; whether AI-written practice questions work as well hasn’t been tested. Practice testing is one of the best-evidenced ways to learn. No randomised study has compared learning from AI-written and human-written questions. In quality checks, AI-written questions were easier, contained more factual errors and were less good at separating students who knew the material from those who didn’t.

Weakly evidenced Evidence maturity: Early

Study cards

• Rowland (2014) · ST38: across 159 effect sizes, testing yourself beat restudying (g = 0.50). Not about AI.
• Law et al. (2025) · ST40: GPT-4o medical questions were easier and had more factual errors (6% vs 4%) than expert-written ones. No learning outcome.
• Al-Najafi et al. (2026) · ST41: AI-written medical questions discriminated less well than student-written ones (0.19 vs 0.25), but took far less time to write.

Can AI check your homework answers?

Small studies suggest AI can’t be relied on to check your answers. Students often accepted AI’s judgement even when it was wrong, and current models often fail to find the first mistake in a student’s working.

Poor fit / caution Evidence maturity: Early

• In Estonia, 36 final-year school students accepted 52% of deliberately wrong AI suggestions on maths puzzles.
• University physics students judged nearly half of wrong ChatGPT-aided solutions to be correct.
• In tests on collections of student solutions, AI models found the first error in 20% to 64% of cases, depending on the collection.

Study cards

• Puppart & Aru (2025) · ST42 (preprint).
• Krupp et al. · ST43 (peer reviewed, 39 students).
• Srivatsa, Maurya & Kochmar (2025) · ST44 (peer reviewed; no students involved).

Can AI find sources for an essay?

AI chatbots can’t be relied on to find real sources. In one audit, GPT-3.5 made up 55% of the references it supplied and GPT-4 18%. Many real references contained errors. Newer models may do better, but no audit we found shows the problem has gone.

Poor fit / caution Evidence maturity: Early

This is an audit of AI output, not a study of students’ learning.

Study card

• Walters & Wilder (2023), Scientific Reports · ST45: made-up references 55% (GPT-3.5) and 18% (GPT-4). Substantive errors in real references 43% and 24%. 2023 models.

Can AI help you plan your studying?

There isn’t enough evidence yet to say whether AI helps students plan their studying. The one randomised trial found no advantage from building reflection or study-strategy prompts into ChatGPT for German secondary pupils. It compared versions of AI, not AI against none. A small university study improved some self-regulation measures but not results.

Insufficient to classify Evidence maturity: Early

Study cards

• Fütterer et al. (2026) · ST46: 371 pupils, grades 7–9, Germany. Peer reviewed. 35% drop-out.
• Xu et al. (2025) · ST47: 68 university students, China. Peer reviewed.

Does AI help you learn to code?

AI coding help lets you finish more exercises, but the evidence that it helps you learn to program is weak.

• Across ten studies of programming education, the overall learning effect was not statistically significant.
• Where AI was barred from the test, the effect was close to zero. Where AI was allowed, it was large.

Poor fit / caution Evidence maturity: Moderate

• Review of studies: pooled learning effect g = 0.14 (not significant). With AI barred from the assessment, g = −0.06 (seven studies); with AI allowed, g = 0.76 (three studies). The authors: “When participants were not allowed to use AI for the assessment, the effect was non-significant and even pointed slightly in a negative direction”. (ST04) This review is a preprint.
• Germany: AI raised exercise scores but not knowledge (Iris trial, above).
• Ages 10–17: an early AI code generator helped learners complete more tasks. One-week retention was not significantly different.
• Professional developers learning a new code library: those with AI scored 17% lower on a quiz afterwards. This is a small preprint by authors at an AI company.
• University programming experiments: students with less prior knowledge understood less when given AI.

Study cards

• Maier et al. (2026) · ST04
  • 10 studies, 1,069 participants.
  • Preprint.
  • Limits: the “AI allowed” group is three studies.
• Kazemitabaar et al. (2023), CHI · ST34
  • 69 learners aged 10–17; OpenAI Codex, 2022.
  • Peer reviewed.
• Nie et al. (2025) · ST35
  • 5,831 learners on a free online course.
  • An offer of GPT-4 reduced exam participation by 4.3 percentage points.
  • Peer reviewed.
• Shen & Tamkin (2026) · ST36
  • 52 professional developers.
  • Preprint; authors work at Anthropic.
• Prather et al. (2024) · ST37
  • 21 students.
  • Struggling students came away believing they understood more than they did.
  • Qualitative.
• Lehmann et al. · ST33 and Bergh et al. · ST32: see above.

Can AI help you learn a language?

Chatbot conversation practice may help with learning a language: reviews find moderate gains, but most of those chatbots came before today’s AI. No delayed effects were reported, and studies of current models are small and not randomised.

Emerging Evidence maturity: Early

Study cards

• Wang, Cheung, Neitzel & Chai (2025), Review of Educational Research · ST48: 28 studies; g = 0.48.
• Hou & Min (2026), ReCALL · ST49: 16 studies of speaking; g = 0.61.
• Both are peer reviewed. We read the abstracts only.

Does AI help students with SEND?

There is no robust evidence yet on whether AI helps students with special educational needs and disabilities (SEND).

• The only trial we found lasted two days in one Hong Kong school. Engagement rose, but dyslexic pupils’ writing scores fell and copying was high.
• UK evidence is a survey of disabled university students with a response rate of about 1.4%. Most said they used AI and found it helpful.

Insufficient to classify Evidence maturity: Early

Study cards

• Fung et al. (2026) · ST50: 101 pupils. Peer reviewed. The design description is inconsistent.
• Zhao, Cox & Chen (2025) · ST51: 124 UK disabled students. Survey.

Why can it feel as if you’re learning when you aren’t?

Because AI makes the work go well, which is easy to notice, while what you will remember later is not. In several studies, students’ sense of how much they had learned did not match what they could do without AI.

• In the Turkish maths trial, the authors write: “While students in the GPT Base arm performed worse on the exam (relative to the control arm), they did not perceive that they performed worse or learned less.” (ST07)
• English 14–15-year-olds rated the chatbot as more helpful than their own notes, and remembered less (above).
• German students rated AI feedback as very useful, far more than its measured effect (above).
• Adults who used AI on reasoning questions overestimated their scores (16.5 estimated vs 13.0 actual) (ST52).

But students are not always wrong:

• At Harvard, students who found the AI tutor more engaging and motivating also scored higher.
• At Middlebury, students’ estimates of their gains were close to the measured gains.
• In the English trial, the largest group in the second comparison preferred the more effective method.

Our interpretation (not tested): perceptions seem to mislead most when AI replaces effort, and to be more accurate when AI supports it.

Why do some big reviews say AI improves learning so much?

Because many of the studies they combine don’t separate better work from more learning.

• The most-cited positive review (Wang & Fan, 2025) was retracted in April 2026. The journal’s notice: “Taken together, the identified issues undermine the Editor’s confidence in the validity of the analysis and the conclusions drawn from it.” (ST01)
• Another large review reported a large effect on academic performance. Its authors acknowledge an alternative explanation: “the higher quality of work produced with ChatGPT’s assistance, which might be misconstrued as genuine improvement in student learning”. (ST02) Of its studies, 9 allowed ChatGPT in the post-test and 33 did not say.
• An independent audit of 19 comparisons used in such reviews found that only 4 (21%) clearly described the treatment, used a meaningful comparison group and measured learning validly.
• Older evidence on computer tutoring shows the same pattern. In one review, effects were 0.86 on tests the researchers wrote themselves but 0.16 on standardised tests. A newer review of 18 US school experiments found the same pattern (0.48 vs 0.20).

Study cards

• Wang & Fan (2025) · ST01: retracted on 22 April 2026. It had reported g = 0.867.
• Deng et al. (2025), Computers & Education · ST02: 69 studies reviewed. Academic performance g = 0.71. Publication bias detected.
• Weidlich et al. (2025), Journal of Computer Assisted Learning · ST03: audit; 4 of 19 comparisons met all three criteria.
• Kulik & Fletcher (2016) · ST05: 50 evaluations of intelligent tutors, before generative AI.
• Leite et al. · ST06: 18 US school experiments. Conference paper or preprint.

How we assessed the evidence

What we included

We started from the ways students most often use AI, and for each one looked for published evidence of results:

• randomised trials and other controlled studies;
• reviews that combine many studies;
• UK government, regulator and exam-board documents;
• usage surveys;
• tests of AI output.

We included positive, null and negative findings, and reports by companies that make AI tools, clearly labelled.

We re-opened every source we rely on and checked each figure against the original document where we could reach it. Where we could read only the abstract, we say so. Direct quotations were each extracted twice, independently, and used only where both extractions matched word for word.

The question we asked

Every assessment is judged on learning measured when students no longer had the AI. “Better work?” is shown separately, because the two often diverge. For every study, we record whether AI was available during the test: yes, no, or not stated.

The assessment ladder

Assessment	Rule
Established	At least two independent, peer-reviewed causal studies in relevant student groups, pointing the same way, with learning measured without AI. No use currently meets this.
Promising	Peer-reviewed causal evidence that learning measured without AI improved; or consistent evidence from independent studies, at least one of them peer reviewed and relevant. Working papers and preprints may contribute, but do not normally support Promising on their own.
Emerging	Positive evidence that is credible but mostly not yet peer reviewed, too new, too narrow or not yet repeated.
Weakly evidenced	Plausible, but no convincing evidence that learning improves. This includes uses where only the work improves.
Poor fit / caution	Evidence that learning is harmed, or does not improve while the work looks better; or a high risk of confident error. May draw on credible studies not yet peer reviewed. Where it does, we say the evidence “points to” a risk, not that the harm is settled.

Assessment states outside the ladder:

State	Rule
Mixed	Credible evidence points in both directions and cannot yet be reconciled. Not currently used.
Insufficient to classify	Too little evidence of any kind.

These states are not steps between other levels.

Evidence maturity

Separately from the assessment, each use carries an evidence-maturity label. It describes how developed the evidence is, not which way it points.

Maturity	Rule
High	At least two independent, peer-reviewed causal studies in relevant student groups on the main outcome, consistent in direction, with no major gap in relevance. No use currently meets this.
Moderate	At least one relevant peer-reviewed causal study, or several small or short peer-reviewed studies, plus supporting evidence.
Early	Mostly working papers, preprints or vendor reports; single small studies; tests of AI output only; or a major gap in relevance (adults only, or AI tools that pre-date current models).

We weigh five things:

• Peer review: has the deciding study been through independent review?
• Replication: do other studies point the same way?
• Independence: was the study free of the company that made the tool?
• Design: was it randomised or otherwise controlled, and was learning measured without AI?
• Relevance: are the students the right age and stage, is the AI current, and was learning measured later as well as immediately?

Rules we apply

• Every assessment has a “because” naming the evidence it rests on.
• Null and negative results are reported as prominently as positive ones.
• Every gain states what it was compared with.
• Students’ own views of what helped never establish an assessment on their own.
• Tests of AI output can support caution, never a positive assessment.
• Reports by companies that make the tools are kept and labelled, and never relied on alone for a positive assessment.
• Every assessment carries a last-checked date.

We review the page at least annually, and sooner when a major study appears. The Education Endowment Foundation’s research and the DfE tutoring pilot are both expected to report.

What we don’t do

• We don’t combine studies into a single average effect.
• We don’t give numerical scores.
• We don’t rank or recommend products.
• We don’t offer general study tips.

Assessments are editorial judgements against stated rules, not scientific scores. The interpretations labelled “our interpretation” or “our summary” are SuperIntelligenceGuide’s own, drawn across studies.

Download the data

The full evidence behind this page is available as two spreadsheet files (CSV, which opens in Excel, Google Sheets or Numbers):

• Assessments (ai-student-learning-assessments.csv): the 14 uses of AI, each with its assessment, evidence maturity and the reasons for both. Also includes what happened to the work and to learning, whether AI was available in the test, who was studied, limitations and key sources.
• Sources (ai-student-learning-sources.csv): 65 sources (ST01–ST65), each with authors or issuing body, year, full title, publication, status and link. Also: who was studied, the comparison, the AI tool and date, whether AI was available in the test, what was found, developer or vendor involvement, limitations, and how each source and any quotation was checked.
• Read me (README.txt): what the files contain, the licence and attribution.

The source table also contains further evidence reviewed for this page, including sources that did not contribute directly to its conclusions. The Sources list below shows only those cited on this page.

Last checked: 4 October 2026. Version 1.0.

Reusing the data

Our licence. SuperIntelligenceGuide’s compilation, classifications, assessments and original analytical text are licensed under Creative Commons Attribution 4.0 (CC BY 4.0). Please credit “SuperIntelligenceGuide.co.uk, Does AI help students learn, or just help them do the work?”, with a link to this page.

What the licence does not cover. Source titles and other third-party material remain the copyright of their owners. The data files do not reproduce quotations from sources; the short quotations on this page are used under the quotation exception in UK copyright law.

Attribution statement: Contains public sector information licensed under the Open Government Licence v3.0.

Data dictionary

Assessments file

Column	Meaning
row_id	Use of AI, R1–R14
what_ai_does	What the AI does for the student
assessment_kind	“Tier” (on the ladder) or “Assessment state (not a tier)”
assessment	Established, Promising, Emerging, Weakly evidenced, Poor fit / caution; or Mixed, Insufficient to classify
confidence_note	Any qualification on the assessment
maturity	Evidence maturity: High, Moderate or Early
maturity_basis	Why that maturity level applies
because	Our justification, naming the evidence
task_quality	What happened to the work produced with AI (“Better work?”)
learning_without_ai	Learning measured without AI (“Learn more?”)
later_recall	Results from tests taken days or weeks later
transfer	Performance on new or different tasks
confidence_vs_measured	Whether students’ sense of learning matched measured learning
levels_with_evidence	Ages and stages studied
subjects	Subjects studied
comparisons	What AI use was compared with
ai_at_assessment	Whether AI was available during the test
evidence_types	Study designs relied on
independence	Developer or vendor involvement
models_dates	AI models and when they were used
limitations	Main weaknesses of the evidence
key_sources	Source IDs (ST01–ST65)
last_checked	Date the row was last checked

Sources file

Column	Meaning
source_id	ST01–ST65
citation	Short citation used on this page
authors	Authors, or the issuing body for official documents
year	Year of publication
title	Full title, taken from the source itself
title_check	Whether the title was confirmed against the source
publication	Journal, series, publisher or issuing body
status	Peer reviewed, working paper, preprint, vendor report, retracted, or the type of official document
doi	DOI, where one exists
url	Where we read it
cited_on_page	Whether the source is cited on this page
code	Study design or document type
vendor_or_developer	Involvement of the company or team that made the tool
level	Age or stage
country	Where the study took place
n	Number of participants or studies
subject	Subject or task
comparison	What AI use was compared with
ai_tool_date	AI tool or model and when it was used
ai_at_assessment	Whether AI was available during the test
task_quality	Effect on the work produced
learning_without_ai	Effect on learning measured without AI
later_recall	Delayed results
transfer	Performance on new tasks
confidence_vs_measured	Perceived versus measured learning
heterogeneity	Differences between groups of students
limitations	Main weaknesses
access	What we could read (full text, abstract only)
quoted_on_page	Whether this page quotes the source directly
quote_check	Result of our two-pass check on any quotation we considered
rows	Which uses of AI the source informs