{"resourceId":"oecd-pisa-2025-ai-chatbot-use-science-20260908","versions":[{"version":"external-5a0766f61c1a6546016cea32f259023e2223c0fcae80e53622d6ec224e92ef71","resource":{"id":"oecd-pisa-2025-ai-chatbot-use-science-20260908","title":"OECD PISA 2025 Insights: students who use AI chatbots for schoolwork score lower in science on average, with moderate learning use and AI-evaluation teaching the exceptions","organization":"OECD Programme for International Student Assessment (PISA); Insights and Interpretations by Andreas Schleicher","sector":"K-12 secondary (15-year-olds) / international large-scale assessment","geography":"International (91 countries and economies, OECD average); United States included with OECD caution flag on questionnaire results","publishedAt":"2026-09-08","publicationDate":"2026-09-08","eventDate":null,"sourceName":"OECD","sourceLabel":"PISA 2025 Insights and Interpretations (PDF, Volume I support material)","sourceUrl":"https://www.oecd.org/content/dam/oecd/en/publications/support-materials/2026/09/pisa-2025-results-volume-i_5265bfb1/PISA%202025%20Insights%20and%20Interpretations.pdf","evidenceClass":"independent-research","outcomeClass":"cautionary","topics":["knowledge-work","governance-procurement","accessibility-workforce"],"finding":"PISA 2025 (760,000+ 15-year-olds, 91 systems) asked for the first time how often students use AI chatbots to summarise texts, do preliminary research, draft writing and help them learn. Across the OECD, students who do not use AI for summarising, research or drafting outperform users by around 20 science points, which the OECD equates to roughly a year of schooling; gaps reach 29 (Canada) and 36 (New Zealand) points for regular drafting. Weekly use 'to help me learn' performs similarly to non-use. Frequent learning users who are taught at school to assess AI output score slightly higher, but disadvantaged students are less likely to get that teaching. The OECD stresses the data are not causal and advises caution until stronger evidence exists.","sledRelevance":"This is the first global, representative dataset tying student chatbot use for schoolwork to measured achievement. It gives K-12 boards an evidence anchor for the assisted-output versus retained-learning distinction, and it is newly relevant because the 1 Oct WSJ investigation of Google classroom AI pulled these PISA findings into the U.S. district debate while the K-12 archive had no record of them.","evidence":"Cross-sectional international assessment with student questionnaire index (AIUSESCH, four chatbot-use items) and science scores; some estimates adjusted for socio-economic status (Table I.B1.4.61). System-level scatter R-squared 0.16. Correlational only; OECD names reverse causation (struggling students relying on AI) as a plausible explanation. U.S. estimates flagged for unmet sampling standards and particular caution on questionnaire results.","architectureImplications":"Interpretation: general-purpose chatbots used for drafting and summarising are the configurations associated with the largest gaps, while purposeful learning use is not. Districts weighing student-facing tools should favour designs that require student reasoning and output checking over answer or draft generation, and should log use type, not just minutes.","governanceImplications":"Supports policy that distinguishes AI use types (drafting versus help-to-learn) and requires explicit instruction in assessing AI-generated information. The OECD calls for guidelines on pedagogical value, age-appropriateness, integrity, privacy, transparency and bias, plus sustained teacher training.","securityPrivacyImplications":"PISA does not measure privacy or data exposure. It names data privacy as one area guidelines must cover; no tool-level privacy evidence is provided.","caveats":"Correlational, not causal; self-reported AI use; OECD-average figures mask reversals (Thailand and Macao drafting users scored higher). U.S. questionnaire results carry an OECD caution flag. Insights document is interpretive commentary summarising Volume I tables; full chapter pages returned HTTP 403 to the research client and were not inspected.","streamIds":["k12"],"roles":{"sales":"Frame this as a measurement problem, not a ban argument. PISA shows chatbot drafting and summarising track with lower science scores, while moderate learning use does not, and AI-evaluation teaching helps. A district buyer needs help separating use types and teaching students to check AI output. Offer a bounded policy and measurement review. Do not claim any product raises PISA-type outcomes or that the association is causal.","engineering":"The technical takeaway is that use type matters more than access. Any student-facing deployment should let the district distinguish drafting, summarising, research and tutoring modes, restrict or log answer-generation modes, and build in source-checking prompts. Validate locally: compare achievement for students by use type with socio-economic controls, since PISA itself cannot rule out reverse causation and flags U.S. questionnaire data.","delivery":"Delivery starts with an AI-use-type policy: which modes are allowed for which grades and tasks, and explicit lessons on assessing AI-generated information delivered equitably across schools. Owners are teaching-and-learning and assessment. Track whether disadvantaged schools receive AI-evaluation instruction at the same rate, because PISA shows they currently get less. Treat local outcome data as the decision gate, not international averages."},"roleDetails":{"sales":{"customerProblem":"Boards face pressure to adopt or ban student chatbots without data on which uses relate to learning; PISA links drafting and summarising use to lower science scores.","stakeholders":["superintendent","chief academic officer","assessment and accountability","board","curriculum leads","equity office"],"discoveryQuestions":["Which student AI use types are currently allowed, by grade?","Do students receive explicit instruction in checking AI output?","Is that instruction equally available in high-poverty schools?","What local achievement data could separate use types?"],"valueHypothesis":"A use-type policy plus AI-evaluation instruction lets a district keep purposeful learning uses PISA associates with neutral or slightly positive scores while limiting the drafting uses associated with lower scores.","potentialEngagement":"Four-week AI-use-type policy and measurement review: inventory allowed modes, draft grade-band rules, design a local outcome comparison with socio-economic controls.","unsupportedClaims":["That chatbot use causes lower science scores","That any vendor tool improves PISA-type outcomes","That OECD averages predict a specific U.S. district's results"]},"engineering":{"fit":"Relevant to configuring student-facing AI: evidence favours learning-support modes over draft or summary generation.","architectureIntegration":"Mode-level controls and logging in the student AI tool, exported to the district data warehouse alongside assessment results and socio-economic indicators.","prerequisites":["tool supports mode or feature restrictions","use logs exportable by student","assessment linkage approved by data governance"],"constraints":["PISA is correlational and self-reported","U.S. questionnaire results carry OECD caution flag","country averages mask reversals"],"security":"Use logs are student records; limit retention, restrict access to research and instructional staff, and exclude prompts from vendor training.","proofOfValue":["Compare achievement by AI use type with socio-economic controls","Test whether AI-evaluation lessons change source-checking behaviour","Check equity of access to AI-evaluation instruction"]},"delivery":{"work":"Adopt grade-band AI use-type rules and an AI-output evaluation curriculum, then monitor outcomes by use type.","dependencies":["board policy revision","teacher PD on assessing AI output","tool configuration by IT","assessment office analysis capacity"],"ownership":"Chief academic officer owns use-type policy and curriculum; assessment office owns outcome monitoring; IT owns tool configuration.","skillsAdoption":"Teachers need to teach students to verify AI-generated information and to design tasks where drafting by AI is not the default.","governanceCheckpoints":["use-type policy approval","mid-year equity check on AI-evaluation instruction","end-of-year outcome review by use type"],"acceptanceCriteria":["Proposed: every secondary school delivers documented AI-output evaluation lessons","Proposed: local analysis reports achievement by use type with controls","Proposed: no expansion of drafting modes without local evidence"],"risks":["over-reading correlational data as causal","widening divide if AI-evaluation teaching reaches advantaged schools first","blanket bans removing neutral learning uses"]}},"retrievedAt":"2026-10-04T22:11:01-05:00","enrichedAt":"2026-10-04T22:11:01-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"PISA flags a potential AI divide: disadvantaged students are less likely to be taught to assess AI-generated information, so access-only rollouts may widen gaps.","procurementImplications":"RFPs should require mode-level controls and use-type logging so districts can restrict draft generation and measure outcomes, rather than buying undifferentiated chatbot access.","operatingModelImplications":"Operate student AI as a governed set of permitted uses tied to instruction in evaluating AI output, reviewed against local outcomes, not as a single on/off switch.","sourceVerification":{"openedUrl":"https://www.oecd.org/content/dam/oecd/en/publications/support-materials/2026/09/pisa-2025-results-volume-i_5265bfb1/PISA%202025%20Insights%20and%20Interpretations.pdf","referenceExcerpt":"Until there is stronger evidence that AI is routinely resulting in positive outcomes, teachers and students should use with caution.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}