Learning with generative AI: tutors and coaches

Rigorous evidence on LLM tutors now runs in both directions, and the design of the tool decides the sign. A general chat assistant used during practice raised practice scores and lowered later unassisted exam scores in the one large high-school RCT (Bastani et al. 2025), and a course AI tutor left in its default answer-giving mode lowered university grades by 0.27 to 0.37 SD (Liu et al. 2026). Structured tutors that hold the answer key, withhold the answer and guide the student step by step either removed that harm or produced gains of roughly 0.2 to 0.4 SD in Nigeria, Ghana and English secondary schools, and larger short-run gains in a Harvard physics course (De Simone et al. 2025; Henkel et al. 2024; Harrison et al. 2026; Kestin et al. 2025). The largest and longest trial, of a guardrailed tutor (Khanmigo), found small gains that matched plain practice software, because students rarely talked to the tutor (Oreopoulos and Low 2026). For Avanti this means: ground every problem in a verified answer key, give hints after an attempt rather than answers on request, step in at the moment of a mistake, and measure learning only on tests the student takes without the AI.

Scope note. No completed RCT of an LLM tutor with Indian adolescents or with JEE/NEET-level content was found. Every effect below is from another population and should be read as a direction, not a forecast.

Evidence table

Only sources opened and read in the archived copy are listed.

Study	Design and population	n	Effect (unit)	What it means for the coach
Bastani et al. 2025 (PNAS)	Classroom-randomised RCT, one Turkish high school, grades 9-11 maths, four 90-min review sessions, Fall 2023, GPT-4. Arms: GPT Base (chat with the problem in the prompt, told to act as a tutor), GPT Tutor (teacher-written solutions and common mistakes in the prompt, told to give hints and not the answer), control (book and notes)	~1,000 students, ~50 classes, 2,848 student-session obs	Practice: Base +0.137, Tutor +0.361 on a 0-1 scale (+48%, +127% vs control mean 0.284). Unassisted exam: Base -0.054 (-17% of control mean 0.321; about -0.19 control SD, my calculation from the reported SD 0.277); Tutor -0.004, not significant	Assisted practice scores say nothing about learning. Answer-giving harms; a guardrailed tutor with the answer key removes the harm but did not add learning here. Students did not notice the harm
Liu et al. 2026 (EdWorkingPaper 26-1598)	RCT, instructors randomised within course blocks, large US public university, Fall 2025, GPT-4o tutor with RAG over course materials; most instructors left the default direct-instruction mode on	2,379 students, 30 instructors	Final grade -0.27 to -0.37 SD; LMS participation in set activities -0.90 SD (same-course sample); larger losses for first-generation students	An easy-to-reach answer tool can displace other study and contact with teachers. Default settings matter: only two instructors ended the term in tutor mode
Kestin et al. 2025 (Sci Rep)	Crossover RCT, Harvard intro physics, two lessons (surface tension, fluids). AI tutor at home vs in-class active learning. GPT-4 with step-by-step solutions in the prompt and a platform that walked students through each problem part in order	194 students (142 and 174 post-tests across weeks)	Post-test: 0.63 SD by OLS, 0.73 to 1.3 SD by quantile regression; median 49 min on task vs 60 min in class	Structured, answer-key-grounded tutoring works well for first exposure to new material. Immediate post-test only; elite US undergraduates
De Simone et al. 2025 (World Bank PRWP 11125)	Student-level RCT among volunteers, nine public schools, Benin City, Nigeria, Senior Secondary 1 (about 15 years), June-July 2024. Twelve 90-min after-school lab sessions with Microsoft Copilot (GPT-4), working in pairs, teacher-provided starting prompts written to make the AI tutor rather than answer, teachers circulating	1,328 randomised; 759 completed endline (422 T, 337 C)	Endline 0.31 SD (SE 0.068); English 0.24 SD; third-term school English exam 0.21 SD; about 0.031 SD per extra day attended; mean attendance 72%	Supervised lab sessions with scripted prompts can work on shared devices. Gains track attendance. Differential attrition (64% vs 50% completed) and no time-matched control limit the claim
Henkel et al. 2024 (Rori, arXiv)	School-randomised trial, 11 Rising Academies schools in Ghana, grades 3-9, two 30-min sessions a week in study hall on a phone via WhatsApp, 8 months. Rori is a scripted micro-lesson sequence with LLM-based language understanding; on a struggle the student gets a hint first, then a worked solution	477 students with both tests (637 at baseline)	Growth-score difference 3.01 points on a 35-item test; d = 0.36	Short, regular, timetabled sessions with a hint-then-worked-solution ladder can move scores cheaply. Preliminary: 11 clusters analysed with an unclustered t-test, authors include the provider
Oreopoulos and Low 2026 (EdWorkingPaper 26-1551)	Cluster RCT, grades randomised within 18 Tennessee middle schools, two school years, Khan Academy plus Khanmigo (configured to coach, not give answers) in daily remedial maths blocks	6,902 student-terms; 53 clusters	+1.3 national percentile ranks per term; 0.06 to 0.08 SD per year; implied 0.14 SD for a full year of participation. Gains similar to Khan Academy practice without AI	Access is not use: the median student messaged the tutor on a third of practice days and in 17% of sessions with a mistake; most messages were bare answers or suggested-prompt clicks
Oreopoulos, Liut, Sungu and Low 2026 (EdWorkingPaper 26-1552)	Individually randomised 2x2 field experiment, Hamilton County middle schools, March 2026, NUMI platform. AI tutor (guardrailed, no direct answers) vs no AI, crossed with a mastery rule (three correct in a row) vs free progression; delayed test one week later	More than 6,000 students	AI: fewer questions attempted, slower, more accurate; higher next-attempt correctness after a mistake. Mastery alone: no delayed-test gain. AI within mastery: about 3 percentage points on the practised delayed item, marginally significant	AI help is most useful right after a mistake, and a mastery rule gives students a reason to use the help to understand rather than skip
LearnLM Team and Eedi 2025 (arXiv)	Exploratory RCT, five UK secondary schools, Years 9-10 maths on Eedi, May-June 2025. Static hints vs chat tutoring; tutoring randomised to human tutor or LearnLM with a human tutor approving or editing every message. Model given the question, the wrong answer and the misconception note	165 students, 17 tutors	Tutors approved 76.4% of AI drafts with zero or minimal edits. Transfer to the next topic: 66.2% (AI+supervisor) vs 60.7% (human), +5.5 pp, 95% interval -1.4 to +12.4	A model given the student’s wrong answer and the misconception can write Socratic help that experts accept. Every message was human-checked; written by the model’s developer
Lehmann, Cornelius and Sting 2025 (arXiv)	Two pre-registered, incentivised lab experiments (learning Python) plus a field study of two Dutch graduate courses using ChatGPT outages	108 lab subjects in Study 1; 113 course students	No overall effect of LLM access in the labs. Exploratory: substituting (generating solutions) covers more topics with less understanding; complementing (asking for explanations) deepens understanding. LLM access helped high-prior-knowledge learners and hurt low-prior-knowledge learners	When free to choose, students substitute. Weaker students are the most exposed
Fan et al. 2024 (arXiv; “metacognitive laziness”)	Randomised lab study, essay revision task, four support arms (ChatGPT, human expert, writing checklist, none), Chinese university students	117	ChatGPT arm improved the essay score most; knowledge gain and transfer did not differ; no difference in intrinsic motivation; fewer self-regulation steps in the AI arm	Better product, same learning. The coach should ask the student to plan, check and evaluate, rather than doing it for them
Pardos and Bhandari 2024 (PLOS ONE)	Online RCT, 3 hint conditions x 4 maths areas, US Mechanical Turk adults	274	ChatGPT worked-solution hints gave significant gains vs no help; no significant difference vs human-written hints. 32% of ChatGPT hints failed quality checks before self-consistency checking	Worked solutions help after an attempt, but LLM-written solutions need verification. Adults, not adolescents
Wang et al. 2024/2026 (Tutor CoPilot, arXiv)	Pre-registered RCT of a tutor-facing AI that suggests moves to human tutors during live chat tutoring, US low-income K-12	700+ tutors, 1,000+ students	+4 pp topic mastery on exit tickets; +9 pp for students of lower-rated tutors; about USD 20 per tutor per year. No end-of-year score outcome	AI that assists a human can raise quality at low cost. Relevant to teacher briefs more than to the student coach
Harrison et al. 2026 (Medly, arXiv)	Individually randomised multisite micro-RCT, English secondary schools, GCSE Biology, Chemistry, Physics, four weekly 30-min activities with LLM marking of extended answers against exam-board criteria vs self-directed revision	929 at baseline, 644 at post-test	Hedges g = 0.33 (95% CI 0.18 to 0.48)	Exam-aligned practice with AI marking can help over four weeks. Funded by the vendor; 30.7% attrition; curriculum-aligned outcome
Huntington-Klein 2026 (arXiv)	Observational: regional “summer dip” in ChatGPT use as a proxy for school use, linked to US high-school test scores	US regions (aggregate)	-0.0096 SD (SE 0.0098) per SD of exposure; bounded no worse than about -0.04 SD	Unsupervised student use has not shown a large aggregate harm in US test scores. Weak identification; cannot rule out small harms
Kosmyna et al. 2025 (arXiv, not peer reviewed)	Lab study, essay writing with LLM, search engine or nothing, with EEG, Boston-area adults	54 (18 in session 4)	LLM group quoted their own essays less and reported lower ownership; lower measured connectivity	A small, unreviewed signal about effort and recall. Not evidence on exam learning

What the evidence says

Answer-giving harms unassisted performance; guardrails remove most of the harm

The cleanest test is Bastani et al. (2025). Both AI arms raised scores on practice problems, by 48% (Base) and 127% (Tutor) relative to control. On the closed-book exam that followed, the Base arm scored 17% lower than control and the Tutor arm was no different from control. GPT Base was not raw ChatGPT: its prompt contained the problem and told GPT-4 to act as a tutor. The harm came anyway. The authors’ mechanism analysis points to copying: the most common Base message was a request for the answer, Base was correct only 51% of the time on these problems, and its errors moved practice scores but did not carry over to exam scores (Bastani et al. 2025). Students in both arms rated their own learning higher than it was.

Liu et al. (2026) found the same direction at university scale. With a course-grounded GPT-4o tutor left mostly in answer-giving mode, final grades fell by 0.27 to 0.37 SD and recorded participation in course activities fell by 0.90 SD. AI users reported asking instructors fewer questions. Lehmann et al. (2025) show why defaults matter: given free access, most learners substituted the AI for their own work, which raised coverage and lowered understanding.

Structured tutors that hold the answer can produce gains

Every positive trial gave the model the verified solution or the misconception, and constrained the path:

• Kestin et al. (2025): step-by-step solutions in the prompt, a platform that forces the sequence of problem parts, and a prompt built for active learning and cognitive load. Effect 0.63 SD or more on an immediate post-test, in less time than class.
• Bastani et al. (2025): GPT Tutor prompt carried teacher solutions and common mistakes and was told to give hints only.
• LearnLM Team and Eedi (2025): the model received the question, the student’s wrong answer and the misconception explanation, and was told to lead the student to find the mistake.
• Henkel et al. (2024): Rori runs scripted micro-lessons; a struggling student gets a hint, then a worked solution.

Effects in low-resource secondary settings are moderate and depend on attendance

De Simone et al. (2025) is the closest setting to a JNV centre: supervised after-school lab sessions on shared computers, about 15-year-olds, teachers present, scripted starting prompts. The effect was 0.31 SD on the endline and 0.21 SD on the regular school exam, with about 0.031 SD per extra day attended. Rori in Ghana reported d = 0.36 from one hour a week over eight months (Henkel et al. 2024). Medly in English GCSE science reported g = 0.33 over four weeks (Harrison et al. 2026). All three are early, and two have provider involvement.

Access is not use; engagement is the binding constraint

Khanmigo, configured not to give answers and placed in mandatory session time, produced 0.06 to 0.08 SD a year, similar to Khan Academy practice without AI (Oreopoulos and Low 2026). The median student messaged it in only 17% of exercise sessions with a mistake, and most messages were bare answers. In the NUMI experiment (Oreopoulos et al. 2026), the AI’s value showed up after mistakes: better next-attempt accuracy and fewer attempts to recover. A mastery rule alone did not improve delayed learning; AI inside a mastery rule gave the best delayed result, marginally significant.

Practice performance vs exam performance

Three opened studies separate the two. Bastani et al. (2025): large practice gains, zero or negative exam effect. Liu et al. (2026): homework effects null or negative, grade effect negative. Fan et al. (2024): a better essay with no extra knowledge gain or transfer. The pattern is consistent: with AI present, the task score rises; what the student can do alone does not necessarily follow.

Persistence and motivation

Direct evidence on persistence over months is thin. Kestin et al. (2025) report higher self-rated engagement and motivation after one lesson. Fan et al. (2024) found no difference in intrinsic motivation. Liu et al. (2026) found lower participation in course activities. Bastani et al. (2025) found that GPT Tutor users sent more messages over time and asked fewer superficial questions as sessions went on, while Base users moved the other way. De Simone et al. (2025) report 72% attendance in a voluntary after-school programme. No opened study measures whether students keep studying through difficulty longer with an AI coach than without.

Who benefits

The direction differs by study. Lehmann et al. (2025) found LLM access helped stronger and hurt weaker learners; Liu et al. (2026) found larger losses for first-generation students; De Simone et al. (2025) found larger gains for higher-performing and female students. Tutor CoPilot helped most where the human tutor was weakest (Wang et al. 2024). Students starting from the bottom deserve separate monitoring.

Human-in-the-loop

Two opened trials support AI that works through or under a human: Tutor CoPilot (+4 pp mastery; Wang et al. 2024) and supervised LearnLM (at least as good as human tutors; LearnLM Team and Eedi 2025). Neither compares supervised AI to unsupervised AI, so neither shows that the hybrid beats an autonomous tutor.

Where the evidence is weak or contested

• No Indian adolescent or JEE/NEET trial. The nearest are Nigerian SS1 English and Ghanaian primary-to-lower-secondary maths. JEE problems are harder than any outcome tested. Language mixing (Hindi, Marathi, English) is untested.
• Short horizons. Kestin is two lessons with an immediate post-test; Bastani four sessions; Medly four weeks; NUMI one session with a one-week delayed test. Only Khanmigo (two years) and Rori (eight months) run long, and Khanmigo is the weakest result.
• The models move faster than the trials. Bastani used GPT-4 in Fall 2023 and found 51% correctness on high-school problems. Current models are more accurate, which weakens the error mechanism but not the copying mechanism, which the authors judged to be the main one.
• Provider involvement. LearnLM is written by the model’s developer; Medly funded its evaluation; Rori’s authors include the provider’s staff.
• Design flaws in positive trials. De Simone: volunteers only, student-level randomisation, some control students attended, differential attrition (64% vs 50% completed endline), and a control group with no extra study time. Rori: 11 schools and no clustered inference reported.
• The aggregate picture is less alarming than the trials. Huntington-Klein (2026) finds near-zero effects of school-related ChatGPT use on US high-school test scores, bounded at about -0.04 SD, with weak identification. One reading: harm happens where AI is used during practice that feeds a later unassisted test, and averages dilute it. A second reading: the trials overstate harm. Opened evidence does not settle this. A Chinese secondary-school panel study (Strömberg, Lei and Wu 2026) is reported to find large exam penalties; it could not be opened and is listed as a lead.
• Arguing the other side. One could argue the harms are 2023-era artefacts and a 2026 frontier model with good instructions will self-limit. The 2025-26 studies cut against this: Liu (GPT-4o, Fall 2025) found harm in default mode, and Khanmigo (built to coach) found students simply did not engage. Prompts alone did not fix student behaviour in either. The stronger position is that structure (when help appears, what it may contain, what counts as progress) does the work, not the model’s good intentions.

Design rules for the Avanti coach

1. Ground every problem in a verified key, held by the model and not shown to the student. The connector should return the solution steps and the known misconceptions for the item to Claude, as Bastani’s GPT Tutor, Kestin’s tutor and LearnLM did. Never let the coach solve a JEE item from scratch and present it as correct (Bastani et al. 2025; Kestin et al. 2025; LearnLM Team and Eedi 2025; Pardos and Bhandari 2024).
2. Attempt first, then a hint ladder, then the worked solution. Hints come only after the student has logged an attempt. The worked solution comes only after a second failed attempt or an explicit give-up, and is followed at once by an isomorphic item the student solves without help (Henkel et al. 2024; Oreopoulos et al. 2026; Lehmann et al. 2025).
3. Stage the reveal through the connector, not through instructions alone. Because regular Claude will help whoever asks, the tool that returns a full solution should require a logged attempt first. Instructions in a response are a nudge; the order of tool calls is a control (Bastani et al. 2025; Liu et al. 2026).
4. Step in at the mistake. When a logged answer is wrong, the coach should open with a diagnostic question about that error, rather than waiting to be asked (Oreopoulos and Low 2026; Oreopoulos et al. 2026).
5. Couple AI help with a mastery rule that copying cannot satisfy. Progress on a chapter needs several consecutive correct answers without help (Oreopoulos et al. 2026).
6. Measure learning only without the AI. Chapter levels, the mastery gate and pilot success metrics should come from attempts on the test platform with Claude closed (AIATs, mocks, Friday quizzes on Lakshya or Sashakt). Accuracy inside Claude sessions is a usage metric, never a learning metric (Bastani et al. 2025; Liu et al. 2026; Fan et al. 2024).
7. Do not use student self-reports as evidence the coach works. Students in both AI arms over-rated their learning (Bastani et al. 2025).
8. Use the coach for structured first exposure. Conceptual priming (Framework Step 1) is where a sequenced, key-grounded AI tutor has the strongest result (Kestin et al. 2025).
9. Timetable the sessions and track dosage. Gains followed attendance in Nigeria; Rori ran in timetabled study hall; voluntary dialogue stayed rare even in mandatory time (De Simone et al. 2025; Henkel et al. 2024; Oreopoulos and Low 2026).
10. Watch displacement. Track whether coached students still book teacher one-on-ones and ask teachers questions; the coach should route students to teachers, not absorb them (Liu et al. 2026).
11. Ask the student to do the metacognitive work. Make the student state the plan, predict the answer, and check it, rather than receiving a finished plan or answer (Fan et al. 2024; Lehmann et al. 2025).
12. Track the superficial-request rate. Count messages that paste the question or ask for the answer, as Bastani did. A rising rate is an early warning (Bastani et al. 2025).
13. Report results by starting level. Check L0-L1 students separately; the harm and the benefit are not evenly spread (Lehmann et al. 2025; Liu et al. 2026; De Simone et al. 2025).
14. Keep a human check on generated content at the start. Teachers review generated practice sets and briefs before release in the pilot (LearnLM Team and Eedi 2025; Wang et al. 2024).
15. Randomise features within the pilot. Effects range from -0.37 to +0.6 SD across opened trials. Small within-centre randomisations of single features (hint ladder vs free chat; mastery rule on vs off) are feasible and informative (Harrison et al. 2026; Oreopoulos et al. 2026).

Where the current spec conflicts with the evidence

• Spec section 2 and 3, “regular Claude … nudge, don’t control”. Bastani’s harmful arm already carried a tutor instruction, and Liu’s harmful tutor was course-grounded; both left the answer one request away. Connector coaching cues are of the same kind. The evidence supports controlling the order of events (attempt before hint, hint before solution) through the connector’s tools, even inside regular Claude (rules 2-3).
• Spec section 3, “worked answers are given, but at the right moment”. Consistent with the evidence only if “the right moment” is defined as after a logged attempt and followed by an unassisted isomorph (rule 2). The spec should state it.
• Spec section 8c, “mastery gate … Claude runs and scores it”, and section 10 success metric “accuracy change on target chapters”. If measured inside Claude sessions, these are assisted scores, which rose 48-127% in Bastani’s trial while exam scores did not. Gate and metric should run on the test platform without the AI (rule 6). The handoff concept’s “levels move only on assessed performance” is the right principle and should be written into the spec.
• Spec section 4 item 2 and section 7 “Generate: practice sets and quizzes”. LLM-written items and solutions need a verified key; 32% of ChatGPT hints failed quality checks in Pardos and Bhandari (2024) and GPT-4 solved only 51% of Bastani’s high-school problems in 2023. Prefer selecting items from Avanti’s own question bank with known answers (rule 1).
• Spec section 6, student sessions “bulk of time on targeted practice”. Consistent with the evidence, provided the coach intervenes at mistakes (rule 4). The spec assumes students will ask Claude for help; Khanmigo shows they mostly will not.
• Spec section 9, risk “shallow learning … measure accuracy change, not time spent”. Right direction, but “accuracy” must be unassisted accuracy.
• Pritam (Oct 2025), “hybrid human-AI workflows … outperform fully autonomous tutors”. No opened trial compares supervised and autonomous AI tutors head to head; the claim rests on a review that was not checked here. Treat it as a hypothesis.
• Spec section 8c, “0.15-0.20 SD vs matched control is a realistic pilot target”. Consistent with the range of opened trials (0.06-0.08 SD per year for Khanmigo, 0.21-0.36 SD in Nigeria, Ghana and GCSE science). The sign is not guaranteed: two opened RCTs found harm when answers were easy to get. The pilot should be able to detect a negative effect, not only a positive one.
• Spec section 5, teachers own “the conversation after a bad mock”. Consistent; Liu et al. (2026) suggest the coach should actively send students back to teachers, since AI access reduced student-instructor contact.

References

All opened 2026-10-26. Local files are in the strand archive folder scratchpad/sources/ai_tutors/, each with a .txt extraction beside it.

• Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., and Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences 122(26): e2422633122. https://doi.org/10.1073/pnas.2422633122 (full text via Europe PMC PMC12232635). Opened 2026-10-26. File: bastani2025_pnas.xml (Europe PMC full-text XML), bastani2025_pnas.txt.
• De Simone, M., Tiberti, F., Barron Rodriguez, M., Manolio, F., Mosuro, W., and Dikoru, E. J. (2025). From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria. World Bank Policy Research Working Paper 11125. https://openknowledge.worldbank.org/entities/publication/15e1ff08-15ae-4f7a-b2a8-d146e6c113ee Opened 2026-10-26. File: desimone2025_nigeria.pdf.
• Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., and Gašević, D. (2024). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. arXiv:2412.09315. https://arxiv.org/abs/2412.09315 Opened 2026-10-26. File: fan2024_metacog_laziness.pdf.
• Harrison, W., Khowaja, R., Dobson, E., Uwimpuhwe, G., and Higgins, S. (2026). Evaluating AI tutoring at the speed of innovation: Practitioner-led micro-randomised trials of an AI tutoring platform in GCSE science. arXiv:2609.14789. https://arxiv.org/abs/2609.14789 Opened 2026-10-26. File: gcse2026_microrandomised.pdf.
• Henkel, O., Horne-Robinson, H., Kozhakhmetova, N., and Lee, A. (2024). Effective and scalable math support: Experimental evidence on the impact of an AI-math tutor in Ghana. arXiv:2402.09809. https://arxiv.org/abs/2402.09809 Opened 2026-10-26. File: henkel2024_rori_ghana.pdf.
• Huntington-Klein, N. (2026). Little impact of ChatGPT on high school test scores. arXiv:2605.08812 (v3). https://arxiv.org/abs/2605.08812 Opened 2026-10-26. File: highschool2026_little_impact.pdf.
• Kestin, G., Miller, K., Klales, A., Milbourne, T., and Ponti, G. (2025). AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports 15: 17458. https://doi.org/10.1038/s41598-025-97652-6 Opened 2026-10-26. File: kestin2025_physics_tutor.pdf.
• Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., and Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv:2506.08872. https://arxiv.org/abs/2506.08872 Opened 2026-10-26. File: kosmyna2025_brain_chatgpt.pdf.
• LearnLM Team, Google, and Eedi (2025). AI tutoring can safely and effectively support students: An exploratory RCT in UK classrooms. arXiv:2512.23633. https://arxiv.org/abs/2512.23633 Opened 2026-10-26. File: learnlm2025_eedi_rct.pdf.
• Lehmann, M., Cornelius, P. B., and Sting, F. J. (2025). AI meets the classroom: When do large language models harm learning? arXiv:2409.09047 (v2, March 2025). https://arxiv.org/abs/2409.09047 Opened 2026-10-26. File: lehmann2024_llm_harm.pdf.
• Liu, J., Sweet, T., Chen, M. H., Engelberg, J., Masters, M. C., Clark, M., Persaud, A., Lancaster, A., Hollingsworth, J. K., and Rice, J. K. (2026). The effects of course-integrated AI tutoring on student performance and engagement: A randomized university trial. EdWorkingPaper 26-1598. https://doi.org/10.26300/y3f8-vh05 Opened 2026-10-26. File: courseai2026_undergrad.pdf.
• Oreopoulos, P., and Low, N. (2026). One click away: AI tutoring with Khanmigo in a two-year school experiment. EdWorkingPaper 26-1551. https://doi.org/10.26300/kner-hv33 Opened 2026-10-26. File: khanmigo2026_twoyear.pdf.
• Oreopoulos, P., Liut, M., Sungu, A., and Low, N. (2026). Making AI tutoring productive: Evidence from a mastery-based math practice experiment. EdWorkingPaper 26-1552 (also NBER WP 35621). https://doi.org/10.26300/01qv-6c22 Opened 2026-10-26. File: oreopoulos2026_numi_mastery.pdf.
• Pardos, Z. A., and Bhandari, S. (2024). ChatGPT-generated help produces learning gains equivalent to human tutor-authored help on mathematics skills. PLOS ONE 19(5): e0304013. https://doi.org/10.1371/journal.pone.0304013 Opened 2026-10-26. File: pardos2024_chatgpt_hints.pdf.
• Wang, R. E., Ribeiro, A. T., Robinson, C. D., Loeb, S., and Demszky, D. (2024; v3 2026). Tutor CoPilot: A human-AI approach for scaling real-time expertise. arXiv:2410.03017. https://arxiv.org/abs/2410.03017 Opened 2026-10-26. File: wang2024_tutorcopilot.pdf.

Leads, not opened

These were found but not read in full. Their findings are not stated as fact above.

• Strömberg, D., Lei, V., and Wu, Y. (2026). The generative AI learning penalty: Evidence from Chinese secondary education. SSRN 6868618 / CEPR DP 21577. SSRN and the VoxEU and VoxDev columns all returned bot-check pages. Most relevant lead: exam-oriented secondary students, homework vs exam outcomes.
• Kumar, H., Rothschild, D. M., Goldstein, D. G., and Hofman, J. M. (2023). Math education with large language models: Peril or promise? SSRN 4641653. No open copy found (arXiv search returned a different paper; Microsoft Research record had no PDF link). Reported to test explanations before vs after an attempt.
• Fütterer, T., Bardach, L., Kuhn, J., Keller, S. D., and Gerjets, P. (2026). Enhancing school students’ self-regulated learning through generative AI support: A randomized controlled trial. Educational Psychology Review 38. https://doi.org/10.1007/s10648-026-10133-8 Springer returned a bot-check page.
• Barcaui, A. (2025). ChatGPT as a cognitive crutch (RCT, 120 Brazilian undergraduates, 45-day retention test). Social Sciences & Humanities Open. Found only in search snippets.
• Sungu, A., Lira, B., and Duckworth, A. (2026). Generative AI can harm teaching. Cited by Liu et al. (2026); not retrieved. Belongs to the teacher-AI strand.
• Darvishi, A., et al. (2024). Impact of AI assistance on student agency. Computers & Education. Paywalled; not retrieved.
• Khan Academy’s reported Uttar Pradesh trial (0.45 effect size claimed on a blog). No primary report found.
• Registered Indian trials with no results yet: a school-randomised custom-GPT revision trial in Rajasthan government schools (ended March 2026) and a Gemini Guided Learning evaluation in junior-school science (ends February 2027), both on the AEA RCT Registry.
• Wang, J., and Fan, W. (2025). Meta-analysis of ChatGPT and learning performance. Cited by Liu et al. (2026); not retrieved.