The Playbook

AI-supported project-based learning: a guide for teachers

Start with what each pupil should learn. Give pupils a meaningful problem that requires that knowledge, teach the difficult parts and check their understanding along the way. AI can offer questions, feedback or ideas. Teach pupils how to check what it produces and give them opportunities to show what they understand independently. ^[1][2][3]

A group project shows what pupils can accomplish with tools. A short individual task or conversation can help you see what each pupil understands. Use both kinds of evidence when they fit the learning goal. ^[3][4]

What the evidence says

Results vary across projects

Chen and colleagues' 2026 review of 22 controlled studies reported positive averages for learning, thinking, motivation and collaboration. The estimate fell substantially when the authors adjusted for small-study effects. They call the evidence preliminary. The review combines problem-based and project-based learning across education levels; the full paper was unavailable for this summary. ^[5]

In a non-randomised study of 61 vocational students, Lestari and Setyawati (2026) reported higher statistical literacy and critical thinking scores with AI-supported PBL. This summary draws on the abstract. In a separate non-randomised comparison of two high-school classes studying climate change, the class using ChatGPT scored higher, but the difference in critical thinking was not statistically significant. That leaves the effect uncertain. ^[6][7]

In a high-school app-design project, pupils reported gains in some professional skills. Scores on AI concepts changed only slightly, with no statistically significant difference. Pupils also reported overreliance and avoidance. The study measured pupils before and after the project, so it cannot tell us what AI alone contributed. ^[8]

Check what pupils can do without help

In a randomised study at one Turkish high school, both groups using GPT-4 did better during maths practice than the group using ordinary resources. On a later exam without books or devices, pupils who used a general GPT-4 interface scored worse than the comparison group. Pupils using a tutor with teacher-informed safeguards showed no clear difference from the comparison group. This study identifies a risk in one setting. A prompt alone cannot guarantee that the tool will support learning. ^[3]

In Nigeria, a randomised study found higher English scores after a six-week, teacher-guided after-school programme using a GPT-4 assistant. The programme added time, guidance and resources, so it cannot isolate the chatbot's effect. It did not study PBL. Both studies underline the need to check learning against the goal, with assistance removed when pupils need to show independent understanding. ^[9][3]

Teach pupils to check claims

A secondary-school case study describes pupils using ChatGPT to generate ideas, organise information and refine language. It does not establish that these uses improved learning. AI can produce convincing errors. Treat its citations as leads to check. Ask pupils to find the original source or data, examine it, compare it with other relevant evidence and explain why they accept, revise or reject a claim. A second AI answer is not an independent source. This is a suggested classroom routine. ^[10][11][4]

Keep design decisions with teachers

Teachers in Ravi and colleagues' 2025 co-design study wanted AI tools that supported their work and professional growth. The study examined prototypes and teachers' views. It does not establish time savings or better pupil outcomes. ^[13]

An approved tool can suggest project ideas, rubric criteria, possible misconceptions or feedback prompts. Check the facts, curriculum fit, accessibility and omissions before using its output. Follow your school's rules on age, access and privacy. Keep identifiable pupil data and confidential materials out of unapproved tools, and offer an equivalent route without AI when access is unsuitable or unavailable. ^[12]

Plan the learning and the assessment

• Choose the subject goal first. Decide what each pupil should know or be able to do. Bundle related goals into one investigation and teach the foundations explicitly where needed. A short explanation can still be useful practice. ^[1][14][15]
• Give pupils a problem that requires the knowledge. A real setting can make the task meaningful. Specify what pupils need to learn, how you will support them and what evidence will show their reasoning. AI can generate a plausible response even to a complex project brief. ^[1][2][3]
• Decide where AI helps. Say when pupils may use AI and when they must try, explain or decide for themselves. Teach them to compare suggestions, check evidence and justify a decision. Make permitted use and disclosure rules clear before the task. ^[3][4][12]
• Collect evidence that fits each goal. Check individual subject understanding, the group's reasoning and evidence, and AI judgement if it is a learning goal. A polished product can hide uneven contributions or the work of a tool. An individual checkpoint still needs clear conditions. Do not penalise pupils simply for using permitted help. ^[1][2][3]

Worked example: making the schoolyard more comfortable

This is a proposed design to adapt to your pupils and school. It has not been tested as an intervention and does not require pupils to have personal AI accounts. ^[1][3]

Challenge: Investigate how to make the schoolyard more comfortable on warm days. Compare at least two changes and recommend one feasible option to school leaders. Mathematics helps pupils compare temperature data. Geography or science helps them explain the microclimate and the limits of their measurements. Language helps them justify a recommendation to an audience.

What each pupil should learn: Read and compare a temperature graph, calculate or interpret a difference and explain one limit of the evidence.

1. Teach and practise. Model how to read a graph and interpret a difference. Give pupils a short example to try before AI enters the task. ^[1][15]
2. Investigate. Use teacher-provided data or collect suitable measurements with school approval. Record where and when they were taken. Separate observations from explanations.
3. Ask for help with the reasoning. In an approved tool, share non-identifying observations and ask: “Ask us one question at a time to help us compare two options. Don't choose for us. Flag claims we should verify.” Check the response; the prompt cannot enforce the rule.
4. Check and choose. Pupils examine the data and reliable sources themselves. They identify which suggestions the evidence supports, compare benefits and practical limits, then make their decision.
5. Present and assess. The group recommends a change with evidence. Each pupil interprets a new graph and explains one limitation. If AI literacy is a goal, each pupil also describes a suggestion they checked and what they changed or rejected.

If pupils cannot use an approved tool, show a response you checked in advance, use a teacher-written example or run the investigation without AI. Ask pupils to disclose permitted help in line with school policy. ^[12]

Limits of the evidence

Direct comparisons of AI-supported PBL and well-designed PBL without AI in secondary schools are still limited. Some studies are small, non-randomised, short, or available only as abstracts. Some measure pupils' impressions or the quality of work with help. Reviews of school studies also find many short-term and quasi-experimental designs. Use the example as a starting point and check what your pupils learn. ^{[16][11][6][7]}

References

1. Barron et al., 1998. Doing with Understanding: Lessons from Research on Problem- and Project-Based Learning. The Journal of the Learning Sciences.
2. Kokotsaki et al., 2016. Project-based learning: A review of the literature. Improving Schools.
3. Bastani et al., 2025. Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences of the United States of America.
4. Hamsa Bastani, 2025. Generative AI without guardrails can harm learning: Evidence from high school mathematics - PMC.
5. Chen et al., 2026. Problem-based and project-based learning as promising frameworks for generative AI-supported education: Emerging evidence from a systematic review and three-level meta-analysis. Acta Psychologica.
6. Lestari & Setyawati, 2026. The Effect of Generative Artificial Intelligence-Assisted Project-Based Learning on Vocational High School Students' Statistical Literacy and Critical Thinking Skills. Desimal Jurnal Matematika.
7. Jannah & Sudrajat, 2026. The effect of project-based learning (PjBL) assisted by ChatGPT on critical thinking skills in climate change material. Jurnal Edukasi Biologi.
8. Hur et al., 2026. Cultivating AI literacy among high school students through generative AI as a collaborative partner. Discover Education.
9. Simone et al., 2025. From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria.
10. Xue et al., 2025. ChatGPT as an Intelligent Scaffolding Tool in Project-Based Learning for Junior Secondary EFL Students. ICAIE.
11. Zhang et al., 2025. Generative artificial intelligence in K-12 education: A systematic review. Research and Practice in Technology Enhanced Learning.
12. Miao, 2023. Guidance for generative AI in education and research.
13. Ravi et al., 2025. Co-designing Large Language Model Tools for Project-Based Learning with K12 Educators. International Conference on Human Factors in Computing Systems.
14. OECD Digital Education Outlook 2026 | OECD, 2026.
15. Lazonder & Harmsen, 2016. Meta-Analysis of Inquiry-Based Learning. Review of Educational Research.
16. Alfarwan, 2025. Generative AI use in K-12 education: a systematic review. Frontiers in Education.