Skip to content

Reviewed edition · qasb report 10/2026 · Topic report · Education

AI and education: mixed evidence on learning with AI

In several trials, AI support improved performance while AI was available. A later advantage without AI is not reliably shown, and an AI tutor at a US university was associated with less participation.

Published
October 27, 2026 · as of October 26
Reading time
12 min
Publisher
qasb Research
Method
qasb methodology v1

Key findings

  1. 0.90 SDlower participation on the learning platform in course sections with access to an AI tutor (SD: standard deviations).1Evidence: observed
  2. 2.56 pointsmore on mathematics homework with earlier AI access, but no significant difference on the midterm without AI.11Evidence: observed
  3. 48.6 minless preparation time a week for primary teachers in England who were encouraged to use the AI assistant Aila.8Evidence: observed
  4. 12 of 38European education systems covered had AI-related competences in lower secondary curricula in 2024/25.20Evidence: observed

1.0Summary

The picture is mixed. Positive effects appeared more often for purpose-designed applications and checked materials, sometimes especially for weaker students. Evidence on retention over weeks, early childhood education and students with special needs in real learning settings is largely missing.

  • Where testing later took place without AI, no clear advantage remained: on the midterm in a mathematics course, and after one week in an older study.
  • Only about 15% of students with access used an AI tutor at a US university, and lower grades appeared only when comparing sections of the same course.
  • Primary teachers in England who were encouraged to use the AI assistant Aila spent 76% of the comparison group's preparation time. In Peru a training package saved no time.
  • In Germany the conference of education ministers added AI competences to its framework on 16 October 2026. The decision sets neither teaching time nor examination rules.

2.0Starting point: common use, frameworks and curricula

According to surveys, many US school staff and participants in two Estonian programming courses used AI. For Europe, a Eurydice data collection records policy frameworks and curricula that address AI in the 2024/25 school year, and for Germany there is a new decision by the conference of education ministers. Such documents describe expectations, not teaching. The usage figures rest on self-reports.

Figure 1

AI in Europe's education systems

Education systems covered with a policy framework or AI competences in the curriculum, 2024/25 school year

Eurydice survey. A policy framework may also address other emerging technologies besides AI; the curriculum figure refers to lower secondary education. Frameworks and curricula describe expectations, not teaching.Evidence: observed · Sources 20

59% of surveyed US school staff (teachers, paraprofessionals and other classroom staff) said they had used AI tools for their work in the 2025/26 school year; among users, 64% said AI freed up time for the parts of the job that matter most to them, and 7% said it cost them time.6A1

About one third of the European education systems covered (12 of 38) had included AI-related competences in lower secondary curricula in the 2024/25 school year, mostly in western and southern Europe and mainly in subjects dedicated to digital education or informatics.20A2

Among US school staff who used AI, 64% said it freed up time for the parts of the job that matter most to them, and 7% said it cost them time.

On 16 October 2026 the German conference of education ministers (Bildungsministerkonferenz) added AI-related competences to the KMK framework "Education in the digital world": the six competence areas keep their structure and status and are extended, for example, by reflectively assessing AI-generated results and possible biases, transparently labelling AI-generated contributions and awareness of deepfakes.10A3

Further findings

  • In the 2024/25 school year, 24 of 38 European education systems covered had a policy framework on AI or other emerging technologies.20A4
  • In two Estonian programming courses for mostly adults, 75.9% and 82.4% had used generative AI at least once, according to the survey.15A5

3.0Learning: performance with AI, little on retention

Trials from the US, Germany and China ask what AI changes in learning: in homework, comprehension questions, revisions and practical training. They mostly show an advantage with AI or immediately afterwards. Where testing took place later without AI, there was no significant difference or the advantage was not statistically clear, and evidence on retention over weeks is largely missing.

Figure 2

Advantage on comprehension questions right after the task

IT apprentices with ChatGPT access compared with a control group without internet

Field experiment with 673 apprentices, working paper without peer review. Only immediate comprehension was measured. Excluding the 12.7% of the treatment group who may have used AI during the questions, the effect on the second task falls to 3.3 percentage points and is no longer significant.Evidence: observed · Sources 2

In a randomized trial with 153 students in a proof-based mathematics course, the group given earlier access to an AI chatbot and AI feedback scored 2.56 points higher on homework during that period; on the following midterm, taken without AI, it did not differ significantly from the group with later access.11A6

In the German field experiment with 673 IT apprentices, apprentices with ChatGPT access answered the immediately following comprehension questions, taken without aids, correctly more often than the control group: by 6.7 percentage points on the within-curriculum task and by 9.7 and 4.4 percentage points on the tasks beyond it.2A7

Older context study: in an experiment with 211 college students, those allowed to use generative AI while studying scored 7.2 percentage points higher on an immediate knowledge test taken without AI; one week later the advantage was 3.9 percentage points and not statistically clear (p = 0.068).7A8

Further findings

  • A 45-minute training session on using AI helped STEM students revise answers with AI more successfully, but nobody was tested without AI.5A9
  • With ChatGPT guidance, dental students averaged 77.25 of 100 points in the practical test, with instructor guidance 67.39.3A10

4.0Tutors and materials: mixed results

Six studies examine specific learning tools: tutors at universities and schools, a mathematics platform, checked study materials and a voice-based discussion partner. The results range from better exam and test results to less participation and, when comparing sections of the same course, lower grades.

In a compulsory first-year economics course at a UK university, access to AI-generated study materials that a named tutor had checked before release was associated with 2.34 additional marks on a 50-mark exam component; about three-quarters of this advantage came from the bottom fifth, and the share of marks below the 60% boundary fell by 24.7 percentage points.12A11

In a randomized trial at a large US public university, access to a course-integrated AI tutor lowered final grades by 0.37 standard deviations and participation in learning-platform activities by 0.90 standard deviations when comparing sections of the same course; the full sample showed no significant grade effect but a similarly large decline in participation.1A12

About three-quarters of the advantage associated with checked AI study materials came from the bottom fifth.

In four-week, individually randomized micro-trials in English secondary schools, students who revised with an AI tutoring platform scored higher on the post-test than students doing their usual self-directed revision (Hedges' g = 0.33; 2.56 points on a 35-point test).19A13

Further findings

  • In Arizona, regular users of an AI maths platform were 0.109 standard deviations ahead of comparison students, without random assignment.18A14
  • Older context study: Khan Academy with a rarely used AI tutor brought as much gain in maths as practice without AI would predict.24A15
  • After their second preparation with a voice-based AI discussion partner, MBA students later spoke up voluntarily in class more often.4A16

5.0Teachers: time and the quality of feedback

Two school-randomized trials reach different results on lesson preparation. Both measure time and teachers' work, not pupils' learning. In Peru time use is self-reported, and Aila was used for an average of only 7% of lessons. In feedback and pedagogical diagnosis, language models showed gaps in three studies, one of them with simulated dialogues.

Figure 3

Feedback comments that name a misconception

3,041 exam answers from computer science courses at one Spanish university, academic year 2025/26

JorGPT dataset of Universidad Francisco de Vitoria. For weak answers (grades 0 to 2 out of 10), instructors named a misconception in 31.5% of comments, the models in 5 to 7%. This was detected with keyword rules, and newer models may behave differently.Evidence: observed · Sources 13

In a school-randomized trial in England, primary teachers encouraged to use the AI lesson assistant Aila spent 76% of the comparison group's time preparing lessons and resources (about 2.5 instead of 3.3 hours a week, 48.6 minutes less); the quality of the resources was no worse.8A17

In a school-randomized trial with 1,269 sixth-grade teachers in Lima, a package of one-day training, a Copilot licence and eight months of follow-up raised licence activation by 20.4 percentage points but saved no measurable time on any of seven core tasks; teachers did, however, write better prompts and were less likely to rate inadequate AI responses as good.9A18

In 3,041 real exam answers from computer science courses, instructors named misconceptions in 15.0% of their feedback comments and three commercial language models in only 4.8 to 7.4%; for weak answers (grades 0 to 2 out of 10) the gap was 31.5% versus 5 to 7%.13A19

Further findings

  • In Brussels, AI feedback on handwritten mathematics bonus tests showed no stable effect on linked exam questions.23A20
  • For simulated pupils with learning and behavioural difficulties, the best AI model met at least one criterion in 55.8% of open-ended items.16A21

6.0Judgement: checking sources, reading matches with care

Two experiments show how students handle AI output: in a lab experiment, those with AI alone checked less often, which also reflects the tools available, and in a course experiment a justification requirement reduced the adoption of flawed suggestions. Two further studies measure matches with AI solutions in programming submissions and detector-flagged AI text in dissertations.

Figure 4

Cross-checking information by tool

Lab experiment with 96 students at a Chinese vocational college, 32 per condition

In the AI-only condition, source checking was restricted by design to source information shown in the AI interface, so the comparison also reflects the tools available. One vocational college, convenience sample, explicit instruction to verify.Evidence: observed · Sources 22

In a randomized lab experiment with 96 students at a Chinese vocational college, participants allowed to use only generative AI checked information against other sources far less often (25.0% versus 68.8% with web search and 78.1% with AI plus web search) and copied AI output directly more often; on the immediately following similar task, the AI-plus-web-search group performed best.22A22

In a randomized experiment embedded in a course, students who had to justify in writing each decision to accept or reject an AI revision suggestion adopted flawed suggestions less often (41% instead of 65%) without accepting sound suggestions less often (86% versus 81%).21A23

In an introductory programming course, student submissions in 2025 matched retrospectively generated AI reference solutions 1.76 to 2.12 times as often as in 2021; because the models converge on few solution forms for tightly specified tasks and submissions from 2021 also produced matches, the study says a match does not prove AI use in an individual case.14A24

Further findings

  • According to the authors' measurement, 29% of US STEM dissertations filed in 2026 contained AI-generated passages, none before 2023.17A25

7.0Evidence

Not every statement in this report is equally well supported. The table shows the key statements with their evidence type, AI relationship and sources.

StatementEvidenceAI linkSources
59% of surveyed US school staff used AI tools in 2025/26observednot stated6
12 of 38: European education systems covered had AI-related competences in lower secondary curricula in 2024/25observednot stated20
2.56 points more on mathematics homework with earlier AI access, but no significant difference on the midterm without AIobservednot stated11
Studying with AI: 3.9 percentage points ahead a week later, not statistically clearobservednot stated7
Checked AI study materials: 2.34 more marks on a 50-mark exam componentobservednot stated12
0.90 SD lower participation on the learning platform in course sections with access to an AI tutor (SD: standard deviations)observednot stated1
48.6 min less preparation time a week for primary teachers in England who were encouraged to use the AI assistant Ailaobservednot stated8
Feedback naming a misconception: instructors 15.0%, language models 4.8–7.4%observednot stated13
Duty to justify: flawed AI suggestions adopted less often, 41% instead of 65%observednot stated21

The 25 reviewed statements from 24 original sources fall into chapters 2.0 (5), 3.0 (5), 4.0 (6), 5.0 (5) and 6.0 (4); all are classed as observed. “Not stated”: the edition does not assess the AI relationship of these statements separately. The scope and limitations of every statement are listed in the appendix (A1 to A25).

Original sources checked on October 26, 2026.

8.0What schools and universities can do now

  1. Test learning without AI as well: alongside performance with AI, plan later tests without AI. In the mathematics course, the midterm without AI showed no significant difference despite the advantage on homework.
  2. Choose tools deliberately: positive effects appeared more often for purpose-designed applications and for study materials that a tutor had checked before release. Part of this evidence comes from studies involving the vendor or without random assignment.
  3. Measure time savings for teachers rather than assume them: a planning assistant saved about 49 minutes a week in England, while a training package with an AI licence in Peru saved no measurable time. Neither trial measured pupils' learning.
  4. Try a justification requirement: in an ungraded exercise in which four of eight AI suggestions were deliberately flawed, students who had to justify their decisions adopted flawed suggestions less often (41% instead of 65%). Only adoption was measured.

9.0Method and limitations

Reporting period: October 1, 2026 up to the data cutoff of October 26, 2026 at 1:00 PM GMT+1 (Berlin time). This topic report is published before the end of the month; sources published after the cutoff follow in the next edition of the series. Compiled: October 27, 2026 at 12:23 AM GMT+1. This is the first baseline. The sources come from different years; their publication dates are not events in the reporting month. No change from an earlier edition was measured.

The report uses only reviewed public original sources. Customer conversations, private dossiers and personal customer assessments are not inputs. Observed, model estimate, scenario and open describe the evidence for each finding. Source values retain their original units and baselines; they are not combined into an overall score.

Editorial approval: qasb editorial team · Approved: October 27, 2026 at 12:25 AM GMT+1. The editorial presentation arranges the reviewed statements and contains no figure that is not in them. Presentation approved by the qasb editorial team · October 27, 2026 at 2:27 PM GMT+1.

Limits

  • This topic report covers one topic area: “Education”. It is not an exhaustive survey of all studies or incidents.
  • A narrow study finding is not a forecast for a particular customer, occupation or security system.
  • No qualified finding does not prove no effect; source-access gaps are disclosed separately.
  • This report contains no illustrative measurement series or invented confidence rating.

10.0Sources

  1. 1The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement: A Randomized University Trial (EdWorkingPaper 26-1598)Annenberg Institute at Brown University (EdWorkingPapers; auch SSRN und IZA DP 18989); Jing Liu, Tracy Sweet, Marty Haoyuan Chen, Jonathan Engelberg, Megan C. Masters, Martyn Clark, Axel Persaud, Alia Lancaster, Jeffrey K. Hollingsworth & Jennifer King Rice (University of Maryland)September 30, 2026edworkingpapers.com/ai26-1598
  2. 2Task Expansion with Generative AI: Experimental Evidence on Middle-Skilled, Early-Career WorkersStanford Digital Economy Lab (Working Paper, zugleich IZA Discussion Paper 18999); Hanna Brosch, Christina Langer, Philipp Lergetporer, Sandra Mueller, Harald Pfeifer & Nele TschöpeOctober 7, 2026digitaleconomy.stanford.edu/publication/task-expansion-with-generative-aiexperimental-evidence-on-middle-skilled-early-career-workers/
  3. 3Effects of Generative AI–Supported Virtual Reality Dental Skill Training on Learning Performance, Visual Behavior, and Cortical Activation Among Dental Students: Stratified Randomized Controlled TrialYuting Wen, Chang Wen, Siyu Huang, Yufei Chong & Dong Yang; Journal of Medical Internet Research (JMIR Publications)October 1, 2026www.jmir.org/2026/1/e90938
  4. 4How assigned AI use before class shapes active student engagement in classDan J. Wang, Neelam Modi Jain, Vanessa Burbano, Jorge Guzman, Daniel Keum, Soomi Kim, Bruce Kogut & Nataliya Wright (Columbia Business School); arXiv-Preprint 2610.10463October 7, 2026arxiv.org/abs/2610.10463
  5. 5From answer machine to collaborative partner: impact of structured generative AI training on scientific problem solving in STEM educationBingjie Huang, Li Xie, Yamei Liu u. a. bis Lei Bao; International Journal of STEM Education 13, Artikel 62 (Springer Nature)October 6, 2026link.springer.com/article/10.1186/s40594-026-00652-9
  6. 6Teachers Using AI Say It Frees Up Time for What Matters MostGallup, Inc. (Walton Family Foundation-Gallup-Befragung mit The Mattering Institute); Megan BrenanOctober 14, 2026news.gallup.com/poll/715364/teachers-using-say-frees-time-matters-mostost.aspx
  7. 7Experimental Evidence on the Learning Impact of Generative AIZara Contractor & Germán Reyes (Middlebury College); arXiv-Preprint 2607.08849 (v1 09.07.2026, v2 04.10.2026)July 9, 2026arxiv.org/abs/2607.08849v2
  8. 8Lesson planning using AI lesson assistant, Aila - A Teacher Choices TrialNational Foundation for Educational Research (NFER) im Auftrag der Education Endowment Foundation; Palak Roy, Helen Poet, Ruth Staunton, Katherine Aston, David Thomas & Ben StylesOctober 6, 2026www.nfer.ac.uk/publications/lesson-planning-using-ai-lesson-assistant-aila-a-teacher-choices-trial/
  9. 9Not Faster, but Better: Generative AI and Teacher Skill in Peru (Policy Research Working Paper 11476)Weltbank (Development Research Group und Education and Skills Global Department); Carla Z. Glave, Carolina Lopez, Ezequiel Molina & Román Andrés ZárateOctober 1, 2026documents.worldbank.org/curated/en/099309210012625906/pdf/IDU-c261aadb-6c32-48af-a07c-9cff5e86c716.pdf
  10. 10KMK-Kompetenzrahmen „Bildung in der digitalen Welt“; Aktualisierung im Kontext Künstlicher Intelligenz (2026)Sekretariat der Kultusministerkonferenz (Beschluss der Bildungsministerkonferenz vom 16.10.2026)October 16, 2026www.kmk.org/fileadmin/Dateien/veroeffentlichungen_beschluesse/2026/2026_10_16-KI-Kompetenzrahmen.pdf
  11. 11Supporting or Replacing Students’ Reasoning? Cognitive Offloading in AI-Assisted Proof LearningEason Chen, Sophia Judicke, Kayla Beigh u. a. bis Ken Koedinger (Carnegie Mellon University; University of Hong Kong); arXiv-Preprint 2610.31481October 22, 2026arxiv.org/abs/2610.31481
  12. 12Verified, not generated: expert-verified AI study materials and the distribution of learning gains in a university courseCanh Thien Dang & An Nguyen (King’s Business School, King’s College London); arXiv-Preprint 2610.07097October 5, 2026arxiv.org/abs/2610.07097
  13. 13Beyond Score Accuracy: Examining the Diagnostic Quality of LLM-Generated Structured Assessment in Higher EducationXi Zhao, Xinyue Jiao & Zhen Xu (Teachers College, Columbia University; New York University); Beitrag zur ICCE 2026, arXiv-Preprint 2610.09460October 7, 2026arxiv.org/abs/2610.09460
  14. 14Correctness, Convergence, and AI-Generated Code Detection: A Longitudinal Study of Student and Large Language Model Code in Introductory ProgrammingRunlong Ye, Jing Fan, Angela Zavaleta Bernuy, Oscar Karnalim, Paul Denny, Juho Leinonen & Michael Liut; Beitrag zu Koli Calling ’26, arXiv-Preprint 2610.00863October 1, 2026arxiv.org/abs/2610.00863
  15. 15Understanding Generative AI Use in Programming MOOCs: The Role of Course Context and Learner CharacteristicsMarina Lepp (University of Tartu); Beitrag zum SIGCSE TS 2027, arXiv-Preprint 2610.04447October 3, 2026arxiv.org/abs/2610.04447
  16. 16TutorCaseBench: A Case-Grounded Benchmark for Evaluating LLM Tutors in Supporting Students with Challenging BehaviorsYeon Su Park, Chanyoung Park, Tae Soo Kim u. a. bis Juho Kim (KAIST, Seoul National University, Google Cloud AI Research, Korea Institute for Curriculum and Evaluation, SkillBench); arXiv-Preprint 2610.30828October 22, 2026arxiv.org/abs/2610.30828
  17. 17How Heavily Are PhD Students Offloading Their Scientific Writing to AI? Evidence from US STEM PhD Dissertations (NBER Working Paper 35859)National Bureau of Economic Research; Daniel P. Gross, Dror Shvadron & Hansen ZhangOctober 5, 2026www.nber.org/papers/w35859
  18. 18Teacher-Mediated Multimodal Platform and Math Achievement: Evidence from Goblins' Yearlong District Implementation (EdWorkingPaper 26-1611)Annenberg Institute at Brown University (EdWorkingPapers); Anika Alam, Nicolas Dominguez-Mel & Jing Liu (University of Maryland)October 7, 2026edworkingpapers.com/ai26-1611
  19. 19Evaluating AI Tutoring at the Speed of Innovation: Practitioner-Led Micro-Randomised Trials of an AI Tutoring Platform in GCSE ScienceWayne Harrison, Rahil Khowaja, Emma Dobson, Germaine Uwimpuhwe & Steve Higgins (WhatWorked Education; Durham University); arXiv-Preprint 2609.14789September 13, 2026arxiv.org/abs/2609.14789
  20. 20Digital education at school in Europe 2026: Bridging gaps in access, teaching and learning (Eurydice-Bericht)Europäische Kommission / EACEA / Eurydice; Publications Office of the European UnionSeptember 22, 2026eurydice.eacea.ec.europa.eu/publications/digital-education-school-europe-2026-bridging-gaps-access-teaching-and-learning
  21. 21Think Before You Accept: Can Written Justification Reduce Uncritical Uptake of AI Writing Suggestions?Yan Tao, Jennifer Meyer & René F. Kizilcec (Cornell University; Universität Wien); arXiv-Preprint 2609.23936September 20, 2026arxiv.org/abs/2609.23936
  22. 22Responsible use of generative AI in Chinese vocational education: eye-tracking evidence on source verification and immediate near-transfer across three information-seeking conditionsJi Ma, Hongjie Yue & Xinyang Li; Frontiers in Psychology 17 (Frontiers Media)September 14, 2026www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2026.1946196/full
  23. 23Evaluating LLM-Generated Feedback on Handwritten Mathematics: Feedback Quality and Subsequent Exam PerformanceArne Vanhoyweghen, Vincent Holst, Melika Mobini u. a. bis Vincent Ginis (Vrije Universiteit Brussel); arXiv-Preprint 2610.27881October 20, 2026arxiv.org/abs/2610.27881
  24. 24One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment (EdWorkingPaper 26-1551)Annenberg Institute at Brown University (EdWorkingPapers); Philip Oreopoulos (University of Toronto) & Nina Low (Charles River Associates)August 17, 2026edworkingpapers.com/ai26-1551
Authors, reuse basis and rights records
  1. 1Annenberg Institute at Brown University (EdWorkingPapers; auch SSRN und IZA DP 18989); Jing Liu, Tracy Sweet, Marty Haoyuan Chen, Jonathan Engelberg, Megan C. Masters, Martyn Clark, Axel Persaud, Alia Lancaster, Jeffrey K. Hollingsworth & Jennifer King Rice (University of Maryland) · Reuse basis: No licence stated; copyright remains with the authors; citation is provided for in the EdWorkingPapers FAQ; our own factual paraphrase and link only, no tables or figures · Rights and license record
  2. 2Stanford Digital Economy Lab (Working Paper, zugleich IZA Discussion Paper 18999); Hanna Brosch, Christina Langer, Philipp Lergetporer, Sandra Mueller, Harald Pfeifer & Nele Tschöpe · Reuse basis: All rights reserved (© Stanford University; terms allow downloads for personal, non-commercial use only); our own factual paraphrase and link only, no figures · Rights and license record
  3. 3Yuting Wen, Chang Wen, Siyu Huang, Yufei Chong & Dong Yang; Journal of Medical Internet Research (JMIR Publications) · Reuse basis: CC BY 4.0; first publication in the Journal of Medical Internet Research is credited; our own paraphrase and link only, no figures · Rights and license record
  4. 4Dan J. Wang, Neelam Modi Jain, Vanessa Burbano, Jorge Guzman, Daniel Keum, Soomi Kim, Bruce Kogut & Nataliya Wright (Columbia Business School); arXiv-Preprint 2610.10463 · Reuse basis: CC BY 4.0; our own paraphrase and link only, no figures · Rights and license record
  5. 5Bingjie Huang, Li Xie, Yamei Liu u. a. bis Lei Bao; International Journal of STEM Education 13, Artikel 62 (Springer Nature) · Reuse basis: CC BY 4.0; our own paraphrase and link only, no figures · Rights and license record
  6. 6Gallup, Inc. (Walton Family Foundation-Gallup-Befragung mit The Mattering Institute); Megan Brenan · Reuse basis: All rights reserved (© Gallup, Inc.); Gallup's permissions page provides for citing individual findings with attribution and a link; our own factual paraphrase and link only, no graphics or text reproduced · Rights and license record
  7. 7Zara Contractor & Germán Reyes (Middlebury College); arXiv-Preprint 2607.08849 (v1 09.07.2026, v2 04.10.2026) · Reuse basis: CC BY 4.0; our own paraphrase and link only, no figures · Rights and license record
  8. 8National Foundation for Educational Research (NFER) im Auftrag der Education Endowment Foundation; Palak Roy, Helen Poet, Ruth Staunton, Katherine Aston, David Thomas & Ben Styles · Reuse basis: Copyright NFER; under NFER's copyright page, freely available NFER material may be reproduced only for research, private study or internal circulation with acknowledgement, other uses need written permission; our own factual paraphrase and link only · Rights and license record
  9. 9Weltbank (Development Research Group und Education and Skills Global Department); Carla Z. Glave, Carolina Lopez, Ezequiel Molina & Román Andrés Zárate · Reuse basis: World Bank terms and conditions: non-commercial use with attribution, no derivative works without consent (the item licence in the Open Knowledge Repository could not be reached at retrieval); our own factual paraphrase and link only, no tables · Rights and license record
  10. 10Sekretariat der Kultusministerkonferenz (Beschluss der Bildungsministerkonferenz vom 16.10.2026) · Reuse basis: Protected by copyright under the KMK terms of use (any use beyond statutory exceptions needs consent); our own factual paraphrase and link only · Rights and license record
  11. 11Eason Chen, Sophia Judicke, Kayla Beigh u. a. bis Ken Koedinger (Carnegie Mellon University; University of Hong Kong); arXiv-Preprint 2610.31481 · Reuse basis: arXiv default licence (non-exclusive distribution licence granted to arXiv); copyright remains with the authors; our own factual paraphrase and link only · Rights and license record
  12. 12Canh Thien Dang & An Nguyen (King’s Business School, King’s College London); arXiv-Preprint 2610.07097 · Reuse basis: CC BY 4.0; our own paraphrase and link only, no figures · Rights and license record
  13. 13Xi Zhao, Xinyue Jiao & Zhen Xu (Teachers College, Columbia University; New York University); Beitrag zur ICCE 2026, arXiv-Preprint 2610.09460 · Reuse basis: CC BY 4.0; our own paraphrase and link only, no figures · Rights and license record
  14. 14Runlong Ye, Jing Fan, Angela Zavaleta Bernuy, Oscar Karnalim, Paul Denny, Juho Leinonen & Michael Liut; Beitrag zu Koli Calling ’26, arXiv-Preprint 2610.00863 · Reuse basis: CC BY-NC-ND 4.0; our own paraphrase and link only, no adaptations or figures · Rights and license record
  15. 15Marina Lepp (University of Tartu); Beitrag zum SIGCSE TS 2027, arXiv-Preprint 2610.04447 · Reuse basis: CC BY 4.0; our own paraphrase and link only, no figures · Rights and license record
  16. 16Yeon Su Park, Chanyoung Park, Tae Soo Kim u. a. bis Juho Kim (KAIST, Seoul National University, Google Cloud AI Research, Korea Institute for Curriculum and Evaluation, SkillBench); arXiv-Preprint 2610.30828 · Reuse basis: CC BY 4.0; our own paraphrase and link only, no figures · Rights and license record
  17. 17National Bureau of Economic Research; Daniel P. Gross, Dror Shvadron & Hansen Zhang · Reuse basis: All rights reserved (© National Bureau of Economic Research); our own factual paraphrase and link only · Rights and license record
  18. 18Annenberg Institute at Brown University (EdWorkingPapers); Anika Alam, Nicolas Dominguez-Mel & Jing Liu (University of Maryland) · Reuse basis: No licence stated; copyright remains with the authors; citation is provided for in the EdWorkingPapers FAQ; our own factual paraphrase and link only, no tables or figures · Rights and license record
  19. 19Wayne Harrison, Rahil Khowaja, Emma Dobson, Germaine Uwimpuhwe & Steve Higgins (WhatWorked Education; Durham University); arXiv-Preprint 2609.14789 · Reuse basis: CC BY 4.0; our own paraphrase and link only, no figures · Rights and license record
  20. 20Europäische Kommission / EACEA / Eurydice; Publications Office of the European Union · Reuse basis: CC BY 4.0 (Commission reuse policy, Decision 2011/833/EU); our own paraphrase and link only, no figures or third-party material · Rights and license record
  21. 21Yan Tao, Jennifer Meyer & René F. Kizilcec (Cornell University; Universität Wien); arXiv-Preprint 2609.23936 · Reuse basis: arXiv default licence (non-exclusive distribution licence granted to arXiv); copyright remains with the authors; our own factual paraphrase and link only · Rights and license record
  22. 22Ji Ma, Hongjie Yue & Xinyang Li; Frontiers in Psychology 17 (Frontiers Media) · Reuse basis: CC BY 4.0; our own paraphrase and link only, no figures · Rights and license record
  23. 23Arne Vanhoyweghen, Vincent Holst, Melika Mobini u. a. bis Vincent Ginis (Vrije Universiteit Brussel); arXiv-Preprint 2610.27881 · Reuse basis: CC BY-NC-ND 4.0; our own paraphrase and link only, no adaptations or figures · Rights and license record
  24. 24Annenberg Institute at Brown University (EdWorkingPapers); Philip Oreopoulos (University of Toronto) & Nina Low (Charles River Associates) · Reuse basis: No licence stated; copyright remains with the authors; citation is provided for in the EdWorkingPapers FAQ; our own factual paraphrase and link only, no tables or figures · Rights and license record

How to cite

qasb Research (2026). AI and education: mixed evidence on learning with AI. Topic report 10/2026, data cutoff October 26, 2026. https://www.qasb.ai/en/reports/2026-10-bildung

Appendix

For every statement: evidence type, change status, scope and limitations, and the observation period, source value, comparison and uncertainty where the source gives them. The texts are the approved entries of the edition, as is its summary.

Scope and limitations of all 25 statements (A1–A25)

Edition summary

The evidence from October and September 2026 is mixed: in several randomized studies AI support improved performance while the AI was available but did not reliably improve later performance without AI, and a course-integrated tutor at a US university went with less participation and lower grades; positive effects appeared more often for purpose-designed applications and checked materials, sometimes especially for weaker students. For teachers, a planning assistant in England saved about 49 minutes a week, while a training programme in Peru saved no time; in Germany the conference of education ministers added AI-related competences to its competence framework on 16 October 2026. Evidence on retention over weeks, early childhood education and students with special educational needs in real learning settings is largely missing.

  1. A1Back to chapter 2.0

    Evidence: observed · baseline · Source 6

    Scope
    Online survey of 1,952 K-12 educators from the probability-based Gallup Panel, 28 July to 14 August 2026; weighted to the 2024 American Community Survey.
    Limitations
    Self-reports on use and time effects; the publication checked gives no frequency, tools or learning effects. Captures individual use, not institutional adoption. Funded by the Walton Family Foundation.
    Observation period
    School year 2025/26 (survey 28 July to 14 August 2026)
    Source value
    59 Percent of surveyed K-12 educators using AI for work
    Baseline
    All surveyed K-12 educators (n = 1,952)
    Comparison
    School year 2025/26
    Uncertainty
    Margin of sampling error ±2.7 percentage points (95%)

  2. A2Back to chapter 2.0

    Evidence: observed · baseline · Source 20

    Scope
    Eurydice report "Digital education at school in Europe 2026", curriculum analysis for lower secondary education; the competences cover, for example, understanding AI-driven technologies, benefits and risks, use and design, and ethical issues.
    Limitations
    Curriculum data say nothing about actual teaching time or students' competences; reference year 2024/25. The KMK decision of 16 October 2026 is not yet reflected.
    Observation period
    School year 2024/25
    Source value
    12 Education systems with AI-related competences in the lower secondary curriculum
    Baseline
    38 education systems covered
    Comparison
    School year 2024/25
    Uncertainty
    Full coverage; no sampling uncertainty

  3. A3Back to chapter 2.0

    Evidence: observed · baseline · Source 10

    Scope
    Joint decision of the German Länder for schools in Germany; according to the text it follows the final 2026 version of the OECD and European Commission AI literacy framework and adopts its four domains as reference terms.
    Limitations
    The decision describes expected competences; it sets no teaching time, subject allocation or examination rules, and implementation lies with the Länder. It says nothing about implementation status or effects in schools.
    Observation period
    Decision of 16 October 2026

  4. A4Back to chapter 2.0

    Evidence: observed · baseline · Source 20

    Scope
    Eurydice data collection in 38 education systems (EU Member States and nine other countries), lower secondary education; binding and non-binding documents included; some countries adopted further frameworks after the reference period.
    Limitations
    Covers policy documents, not their implementation in schools or effects; reference year 2024/25, manuscript completed in June 2026. The count says nothing about the quality or binding force of the frameworks.
    Observation period
    School year 2024/25
    Source value
    24 Education systems with a top-level policy framework on AI or emerging technologies
    Baseline
    38 education systems covered
    Comparison
    School year 2024/25
    Uncertainty
    Full coverage through the Eurydice network; no sampling uncertainty

  5. A5Back to chapter 2.0

    Evidence: observed · baseline · Source 15

    Scope
    End-of-course survey in the 2025/26 academic year at the University of Tartu: 187 participants of the four-week course "About Programming" and 182 of the eight-week course "Introduction to Programming" (about 80% of registered participants); more frequent use in the longer course, mainly for debugging and idea generation.
    Limitations
    Self-reports; respondents may differ from non-respondents; two courses in one country. No analysis of actual AI interactions and no evidence of learning effects. Conference paper (SIGCSE TS 2027).
    Observation period
    Academic year 2025/26
    Source value
    82.4 Percent of respondents using AI at least once (eight-week course, 150 of 182)
    Baseline
    Four-week course: 75.9% (142 of 187)
    Comparison
    Eight-week course with a higher workload
    Uncertainty
    Difference between courses not significant (χ² = 2.35; p = 0.113)

  6. A6Back to chapter 3.0

    Evidence: observed · baseline · Source 11

    Scope
    Course on logic, sets, relations and proofs at Carnegie Mellon University (USA); 155 of 308 enrolled students consented and 153 were analysed (75 earlier, 78 later access); individual random assignment after the first midterm; a chatbot and a feedback interface that required a proof attempt first were released together.
    Limitations
    One course; consenting students may differ from the rest; both tools were assigned together, so their separate effects are not identified. The midterm is not a purpose-built transfer or delayed-retention test. The association between more chatbot messages and lower later exam scores rests on self-selected use and is not causal. Preprint without peer review.
    Observation period
    One semester up to the third midterm (year not stated in the version checked)
    Source value
    2.56 Homework points during the early-access-only period (adjusted)
    Baseline
    Group with later access
    Comparison
    Group with earlier access; on the midterm without AI: −2.47 points (95% CI −5.60 to 0.66; p = 0.122)
    Uncertainty
    Standard error 1.22; 95% CI 0.17 to 4.94; p = 0.036

  7. A7Back to chapter 3.0

    Evidence: observed · baseline · Source 2

    Scope
    Twelve vocational schools in six German states; four multiple-choice questions directly after each task, to be answered without AI or other aids; no-internet control group; estimates with controls.
    Limitations
    Immediate comprehension only, no measurement after days or weeks. The treatment group kept internet access during the questions; 12.7% of the treatment group may have used AI then according to self-reports or logs; excluding them, the effects on the within-curriculum task and the first beyond-curriculum task remain, while the effect on the second falls to 3.3 percentage points and is no longer significant. Comparison with a no-internet condition; working paper without peer review. The same study supports a different finding (task performance) in the work area.
    Observation period
    October 2025 to January 2026
    Source value
    6.7 Percentage points of comprehension questions answered correctly (within-curriculum task)
    Baseline
    No-internet control group: 65.5% correct
    Comparison
    Group with ChatGPT access (intention-to-treat, with controls)
    Uncertainty
    Standard error 2.2 percentage points; p < 0.01

  8. A8Back to chapter 3.0

    Evidence: observed · baseline · Source 7

    Scope
    Middlebury College (USA), spring 2025; two in-person sessions about one week apart; up to 35 minutes of study time on unfamiliar technical topics with or without AI; tests and essays in the second session without AI or other aids.
    Limitations
    First published in July 2026 (revised version 4 October 2026), therefore context only. One college, a short study phase with fixed study time. According to the authors, rule violations in the first session could account for 17 to 22% of the immediate effect. Preprint without peer review.
    Observation period
    Spring 2025
    Source value
    3.9 Percentage points on the knowledge test one week later (without AI)
    Baseline
    Control group: 50.0% correct
    Comparison
    Group allowed to use AI during the study phase
    Uncertainty
    0.21 standard deviations; p = 0.068

  9. A9Back to chapter 3.0

    Evidence: observed · baseline · Source 5

    Scope
    Second-year STEM students in a calculus-based introductory physics course at a Chinese university (116 recruited, 95 analysed: 50 training, 45 control); one session; items from a conceptual test on electricity and magnetism, answered first without AI and then revised with the chatbot Doubao.
    Limitations
    Only AI-assisted revision of the same items in the same session was measured; no test without AI and no retention or transfer measure; the authors state the design cannot support inferences about durable learning. The control group received no AI guidance; one site.
    Observation period
    Single session (data-collection date not stated in the version checked)
    Source value
    0.814 Cohen's d (performance in AI-assisted revision)
    Baseline
    Control group with problem-solving training without AI guidance
    Comparison
    Group with structured AI training
    Uncertainty
    t(93) = 3.964; p < 0.001; no confidence interval in the text checked

  10. A10Back to chapter 3.0

    Evidence: observed · baseline · Source 3

    Scope
    Preclinical tooth-preparation training at Wuhan University; one hour a day for seven days; 62 students with ChatGPT-4.0 guidance (outputs checked for errors by research assistants), 61 with on-site instructor guidance; 30-minute practical test scored by two blinded experts.
    Limitations
    Immediate performance after one week only; no measurement of retention, transfer or clinical performance. Participants and assistants were not blinded; AI outputs were checked and therefore do not reflect unsupervised use. One university and one basic skill; a preprint version had been public since April 2026.
    Observation period
    Seven training days (exact data-collection period not stated in the version checked)
    Source value
    9.86 Points in the practical test (0–100 scale)
    Baseline
    Instructor guidance: 67.39 points
    Comparison
    ChatGPT guidance: 77.25 points
    Uncertainty
    95% confidence interval 4.81–14.91; P < .001

  11. A11Back to chapter 4.0

    Evidence: observed · baseline · Source 12

    Scope
    170 students with 340 exam marks; difference-in-differences comparison of the two course halves over two cohorts (2023/24 without, 2024/25 with materials for one half); podcasts, FAQs and quiz-based study guides from a source-grounded model; plus interviews and feedback from 36 students.
    Limitations
    No random assignment; with only one earlier cohort, parallel trends cannot be tested, and cohort differences and shifts of study time between course halves cannot be ruled out. The average effect does not survive dropping the lowest earlier-cohort marks, while the threshold effect does. One course; preprint without peer review.
    Observation period
    Academic years 2023/24 and 2024/25
    Source value
    2.34 Marks on a 50-mark exam component (difference-in-differences)
    Baseline
    Counterfactual from the earlier cohort and the course half without materials (parallel trends)
    Comparison
    Course half with checked AI study materials in 2024/25
    Uncertainty
    Standard error 1.169 (p = 0.045), or 0.947 when clustered by student (p = 0.013)

  12. A12Back to chapter 4.0

    Evidence: observed · baseline · Source 1

    Scope
    2,379 students and 30 instructors across disciplines at the University of Maryland, fall semester 2025; random assignment of instructors within blocks of identical or similar courses; a GPT-4o-based study assistant drawing on instructor-selected course materials; both groups had access to other university AI tools.
    Limitations
    Only about 15% of students with access used the tutor at least once, so the mechanism remains unclear. The analysis plan was registered after data collection but before analysis. One university and one configuration: mostly the direct-answer mode, with only two instructors using the guided tutor mode at the end of the term; the university's in-house tool. Working paper without peer review; the manuscript is dated 30 September 2026, and EdWorkingPapers and IZA list it as the October 2026 version.
    Observation period
    Fall semester 2025 (academic year 2025/26)
    Source value
    -0.37 Standard deviations of the final grade (sections of the same course)
    Baseline
    Sections without tutor access
    Comparison
    Sections with tutor access (intention-to-treat; −4.12 grade points)
    Uncertainty
    Standard error 1.80 grade points (clustered by course section, 34 clusters); robust to randomization inference according to the authors; no confidence interval reported

  13. A13Back to chapter 4.0

    Evidence: observed · baseline · Source 19

    Scope
    Year 9 and 10 students in GCSE biology, chemistry and physics; 929 with a pre-test, 644 with a post-test; trials run by teachers via the WhatWorked platform; the content had already been taught and revision was compared; no evidence of a different effect by Pupil Premium status (an indicator of disadvantage).
    Limitations
    High attrition (30.7%; control group 33.2%, intervention group 27.7%); curriculum-aligned rather than standardized tests, pre-marked by AI and checked and entered by the teachers; short duration with no delayed test. Funded by the vendor Medly, which commissioned the evaluation and provided access to schools already using its platform. Preprint without peer review; the authors treat the results as preliminary.
    Observation period
    Four-week revision period (dates not stated in the version checked)
    Source value
    0.33 Hedges' g on the post-test (intention-to-treat)
    Baseline
    Usual self-directed revision: adjusted mean 11.39 points
    Comparison
    AI tutoring platform: adjusted mean 13.95 points
    Uncertainty
    95% CI 0.18 to 0.48

  14. A14Back to chapter 4.0

    Evidence: observed · baseline · Source 18

    Scope
    Deer Valley Unified School District, 2025/26 school year: 6,065 students in grades 6 to 8; 72.6% of maths teachers activated their licence and 50.9% of students solved at least two problems a week on average; weighted propensity-score comparison of users and comparison students; the study was pre-registered, but the analysis departs from the registered plan (effect among users instead of intention-to-treat); test: NWEA MAP mathematics.
    Limitations
    No random assignment: students determined their own use, and according to the authors the estimates remain vulnerable to bias from unobserved characteristics. The analysis departs from the pre-registered intention-to-treat plan, which the authors justify with highly variable usage. One district, one product; conducted in partnership with the vendor and funded by the nonprofit Accelerate. Working paper without peer review.
    Observation period
    2025/26 school year
    Source value
    0.109 Standard deviations on the district-standardized MAP mathematics test (spring)
    Baseline
    Weighted comparison students with no or low use (fewer than two problems solved a week)
    Comparison
    Students solving at least two problems a week
    Uncertainty
    Standard error 0.021 (clustered by teacher); p < 0.001

  15. A15Back to chapter 4.0

    Evidence: observed · baseline · Source 24

    Scope
    Hamilton County (Tennessee), school years 2024/25 and 2025/26; in each school one or two of three grades were randomly assigned; below-grade-level students in daily remedial blocks of 25 to 40 minutes; control grades with usual remediation, partly with other practice software without an AI tutor; pre-registered.
    Limitations
    First published in August 2026, therefore context only. The comparison does not isolate the AI tutor because practice volume and software also differed. 96% tried Khanmigo, but the median student messaged it on only a third of practice days. Working paper without peer review.
    Observation period
    School years 2024/25 and 2025/26
    Source value
    1.26 National percentile ranks on the MAP mathematics test per testing term (intention-to-treat)
    Baseline
    Control grades with usual remediation
    Comparison
    Grades using Khan Academy with Khanmigo
    Uncertainty
    Standard error 0.60; according to the study about 0.06 to 0.08 standard deviations per school year

  16. A16Back to chapter 4.0

    Evidence: observed · baseline · Source 4

    Scope
    Ten sections of a required first-year MBA strategy course at a US business school, fall 2025; randomization determined before which two of ten case discussions the tool could be used; contributions were recorded by teaching assistants blind to assignment.
    Limitations
    Outcomes are spoken contributions and self-rated learning, not exam performance. All students eventually received two uses; the comparison is earlier versus later use, with no human discussion-partner comparison. A purpose-built tool at one school; preprint without peer review.
    Observation period
    Fall semester 2025
    Source value
    0.112 Additional voluntary contributions per later session (intention-to-treat, effect of the second use)
    Baseline
    Sample mean: 0.73 voluntary contributions per student and session
    Comparison
    After the second assigned use; for use of at least ten minutes, the study puts the effect at about 31% of the mean
    Uncertainty
    Standard error 0.046; p < 0.05

  17. A17Back to chapter 5.0

    Evidence: observed · baseline · Source 8

    Scope
    464 Key Stage 2 teachers (Years 3 to 6) in 108 schools in England, ten weeks in the autumn term 2025; evaluated by NFER for the Education Endowment Foundation; the comparison group was asked not to use Aila, and both groups could use other generative AI tools.
    Limitations
    Aila was used for an average of 7% of lessons, and frequency and amount of use declined over the ten weeks; many teachers preferred other AI tools and often perceived the outputs as needing adaptation before classroom use. No measure of pupil attainment. The full EEF report with methods and uncertainty was blocked at retrieval and at the final check on 26 October 2026 (HTTP 403); the evaluator NFER's results page was checked instead. Tool of Oak National Academy.
    Observation period
    Autumn term 2025 (ten weeks)
    Source value
    48.6 Minutes less preparation time per week
    Baseline
    Group without Aila: about 3.3 hours a week
    Comparison
    Aila group: about 2.5 hours a week (76% of the comparison group's time)
    Uncertainty
    No uncertainty given on the results page checked

  18. A18Back to chapter 5.0

    Evidence: observed · baseline · Source 9

    Scope
    390 public schools in Lima (196 treatment, 194 control), 2025 school year; time data from an endline survey in December (n = 607), prompts graded blind; control schools kept access to freely available AI tools; prompt-quality gains were largest for teachers above the median age of 54.
    Limitations
    Time use is self-reported (minutes the last time a task was done); just under half of the teachers answered the endline survey. The study measures teachers' AI skill, not pupils' learning. Co-funded by the Microsoft AI Economy Institute, whose product Copilot was used. World Bank working paper without external peer review.
    Observation period
    2025 school year (training at the start, endline survey in December 2025)
    Source value
    1.653 Minutes difference in the duration of the last lesson planning (treatment effect)
    Baseline
    Control group: mean of 71.2 minutes
    Comparison
    Treatment group with training, licence and follow-up
    Uncertainty
    Standard error 6.24 minutes; not statistically significant

  19. A19Back to chapter 5.0

    Evidence: observed · baseline · Source 13

    Scope
    Public JorGPT dataset: answers to 50 open-ended questions from regular assessments in the 2025/26 academic year at Universidad Francisco de Vitoria (Spain), each graded and commented on by instructors and by DeepSeek-chat-V3.2, Qwen-flash and Gemini-2.5-flash-lite (zero-shot prompting).
    Limitations
    One institution, software engineering courses; the role of feedback comments was detected with keyword rules and may miss subtler phrasing. Three models of a specific generation; newer models may behave differently. Conference paper (ICCE 2026).
    Observation period
    Academic year 2025/26 (exam answers)
    Source value
    15 Percent of feedback comments naming a misconception (instructors)
    Baseline
    Instructors: 15.0%
    Comparison
    Language models: 4.8 to 7.4%; for grades 0 to 2: instructors 31.5%, models 5 to 7%
    Uncertainty
    No uncertainty given in the source

  20. A20Back to chapter 5.0

    Evidence: observed · baseline · Source 23

    Scope
    English- and Dutch-taught cohorts of the same bachelor course (343 and 236 students with bonus-test data); on tests 1 and 3 one half received feedback and the other served as control; reviewers rated the generated practice exercises strong in 95.5% of cases but diagnostic specificity in only 41.1%.
    Limitations
    Small comparison groups (31 to 52 per arm) of students who sat both crossover tests; one-off feedback on a single topic, months before the exam. Students knew the feedback was AI-generated. The association between more feedback and better exam results cannot be separated from attendance and practice. Preprint without peer review.
    Observation period
    One semester with six weekly bonus tests and a final exam (year not stated in the version checked)

  21. A21Back to chapter 5.0

    Evidence: observed · baseline · Source 16

    Scope
    TutorCaseBench: scenarios built from published case studies in pedagogy, educational psychology and special education; five types of challenging behaviour; 1,560 open-ended and 2,537 multiple-choice items; scored by a language model acting as judge.
    Limitations
    Simulated dialogues instead of real students; construction and scoring rely mainly on Gemini models, and the authors tested possible bias with other models; human validation covered only grounding in the sources. No evidence on learning effects for students with special educational needs. One author affiliation with Google Cloud AI Research; preprint without peer review.
    Observation period
    Benchmark evaluation 2026
    Source value
    55.8 Percent of open-ended items meeting at least one rubric criterion (best model)
    Baseline
    1,560 open-ended items
    Comparison
    Items on longer-term support: at most 43.3%
    Uncertainty
    No uncertainty given in the source

  22. A22Back to chapter 6.0

    Evidence: observed · baseline · Source 22

    Scope
    Liaoning Economy Vocational and Technical College, recruitment 9 May to 9 June 2026; 32 participants in each of three conditions (ChatGPT Plus with GPT-5.3 without a search engine or external webpages, Bing web search, both), stratified assignment; work-related research task (20 minutes) and a similar follow-up task (10 minutes); eye-tracking and screen logs.
    Limitations
    In the AI-only condition, source checking was restricted by design to source information shown in the AI interface, so the comparison of checking rates also reflects the tools available. The follow-up task was solved with the assigned tools and does not measure later or tool-free transfer. One vocational college, convenience sample, explicit instruction to verify.
    Observation period
    9 May to 9 June 2026 (recruitment and experiment)
    Source value
    25 Percent of participants attempting cross-verification (AI only)
    Baseline
    Web search: 68.8%
    Comparison
    AI plus web search: 78.1%
    Uncertainty
    χ² = 21.03; p < 0.001; Cramér's V = 0.468

  23. A23Back to chapter 6.0

    Evidence: observed · baseline · Source 21

    Scope
    129 students in a data science course at a US university (64 with a justification requirement, 65 selection only); eight personalized GPT-4.1 suggestions per proposal draft, four of them deliberately flawed; ungraded single-session activity.
    Limitations
    Only adoption decisions were measured, not the quality of revised texts or learning; four flawed suggestions out of eight are far more than in everyday use. Justifications were often superficial. One course; preprint without peer review.
    Observation period
    One course session (date not stated in the version checked)
    Source value
    24 Percentage points fewer flawed suggestions adopted
    Baseline
    Selection only: 65% adopted
    Comparison
    With justification requirement: 41% adopted
    Uncertainty
    Condition-by-quality interaction: OR = 0.180; p < 0.001 (mixed logistic model)

  24. A24Back to chapter 6.0

    Evidence: observed · baseline · Source 14

    Scope
    29,970 submissions from ten Python labs in the 2021, 2023 and 2025 cohorts; 90,000 solution attempts by GPT-5.5, Gemini 3.1 Pro and Claude Opus 4.8, compared with MOSS and syntax-tree forms; at a 30-line minimum match length, 2.9 to 3.5% of 2021 submissions still had a match.
    Limitations
    Observational cohort comparison without verified labels of which submissions were produced with AI; the study can neither validate individual detection nor estimate the prevalence of AI use. Changing tutors, cohorts and permitted support; one course. Conference paper (Koli Calling 2026).
    Observation period
    Course cohorts 2021, 2023 and 2025
    Source value
    1.76 Multiple of the 2021 match incidence (lower bound across three models)
    Baseline
    2021 cohort (before ChatGPT)
    Comparison
    2025: 1.76 to 2.12 times; 2023: 1.06 to 1.19 times
    Uncertainty
    BH-adjusted q < 0.001; range across three models

  25. A25Back to chapter 6.0

    Evidence: observed · baseline · Source 17

    Scope
    US STEM doctoral dissertations from the 2019 to 2026 graduation cohorts, linked to graduates' later employment; AI-generated text measured with a detector (according to the acknowledgements, with API credits from the vendor Pangram).
    Limitations
    The abstract and details on the NBER page were checked; the PDF was blocked at retrieval and at the final check on 26 October 2026 (HTTP 403), so the detector's method, thresholds and error rates could not be checked. Detectors can misattribute text; AI-generated text does not prove a rule violation. Working paper without peer review.
    Observation period
    Graduation cohorts 2019 to 2026
    Source value
    29 Percent of dissertations filed in 2026 containing AI-generated text
    Baseline
    Before 2023: no AI-generated text
    Comparison
    Filing year 2026
    Uncertainty
    No uncertainty given in the abstract checked

And in your case? Check it with sources.

Start an analysisFor organizations
All reports