Real-world learning in the AI era

Capability is not output

Generative AI can make an essay, image, plan, or block of code appear in seconds. That changes the cost of producing an output. It does not remove the work of building knowledge, deciding what matters, testing an idea, collaborating with other people, or explaining why a choice is sound.

The distinction matters because a learner can submit something plausible without being able to reconstruct the reasoning behind it. The OECD's 2026 outlook makes the same boundary visible: successful performance with a general-purpose AI tool does not automatically become a learning gain, and educational use should not replace cognitive effort.¹

Two experiments show the gap directly. In a randomised field experiment at one Turkish high school, nearly 1,000 maths students who practised with an unrestricted GPT-4 chat interface scored 48% higher than a no-AI group during assisted practice, then 17% lower on a later exam taken without AI. A version with teacher-designed, hint-only guardrails avoided that drop, but its exam results were not significantly different from the no-AI group.² In a separate lab experiment, 117 university students writing in a second language were randomly assigned different kinds of revision support. ChatGPT improved their essays more than any other condition, including a human writing expert, yet knowledge gain and transfer did not differ across groups.³

This is not an argument for pretending AI does not exist. It is an argument for designing learning around capability: what a learner can understand, do, explain, revise, and transfer when the shortcut is unavailable.

What real-world learning asks of a learner

Real-world learning connects disciplinary knowledge to a problem, audience, decision, or consequence beyond completing an assignment. The point is not that every lesson must simulate a workplace. The point is that students have reasons to inquire, make choices, encounter constraints, and improve work that someone can question.

Four qualities are especially valuable when polished output is cheap:

1. Authenticity. The work has a credible purpose, audience, context, or impact.
2. Inquiry. Students must frame questions, find and evaluate information, and make defensible choices.
3. Collaboration. Other people introduce perspectives, expertise, disagreement, and accountability that a solitary prompt cannot reproduce.
4. Reflection. Learners explain what changed, where they struggled, what evidence altered their view, and what they would do next.

PBLWorks names authenticity, sustained inquiry, and reflection among the seven design elements of its Gold Standard PBL model, alongside a challenging problem or question, student voice and choice, critique and revision, and a public product. It treats collaboration differently: as one of the success skills projects should help students develop, rather than as a design element.⁴

What the research can—and cannot—establish

Project-based learning is not one uniform intervention. A meta-analysis of 30 published studies found a positive average effect of PBL on academic achievement compared with conventional instruction, with results that varied by subject area, school location, and instructional hours.⁵ That pooled estimate needs a caution. Across 645 education evaluations, published studies and small trials reported effect sizes roughly twice as large as others,⁶ and the meta-analysis included only published articles, so the pooled estimate is likely to overstate the effect a school should expect. The individual evaluations below also suggest that implementation demands and school conditions matter, though these are harder to isolate. Two prominent studies show why both encouragement and caution belong in the same account.

An Education Endowment Foundation trial of a particular PBL programme found no clear impact on literacy, engagement, or attendance overall, and a statistically significant negative literacy impact for pupils eligible for free school meals, though not for lower-attaining pupils overall. The evaluation reported substantial attrition and crossover between groups, rated the findings as low security, and noted the organisational demands of implementation. Its process evaluation also reported perceived gains in oracy, communication, teamwork, and self-directed study skills, but those observations are not equivalent to a secure causal effect.⁷

A separate cluster-randomized study compared project-based and lecture-based approaches in two Advanced Placement courses. It found positive AP exam effects in both courses studied, including among students from lower- and higher-income households. The authors bound their conclusion to those courses, participating U.S. districts with open-enrollment AP policies, and settings supportive of the approach.⁸

Taken together, these studies do not support “PBL always works” or “PBL does not work.” They support a more useful question: what design, support, subject context, and implementation conditions help a particular group of learners develop particular capabilities?

Support is part of that question, not a detail. Kirschner, Sweller, and Clark argued that minimally guided inquiry works poorly for novices.⁹ Hmelo-Silver, Duncan, and Chinn replied that problem-based and inquiry learning are highly scaffolded rather than minimally guided.¹⁰ One way to read the exchange: both sides agree that novices need guidance, and they disagree about whether problem-based and inquiry approaches provide enough of it.

Design implications for educators

Make the process inspectable

Ask students to preserve the path, not just the final product: early questions, source choices, rejected ideas, drafts, feedback, and revisions. A process portfolio can give teachers evidence of developing understanding and give students material for reflection.

Build in moments that require judgment

Use oral explanations, design critiques, demonstrations, field observations, peer review, and decisions under constraint. These are not “AI-proof” assessment tricks. They are opportunities to reveal how knowledge is being used.

There is also a learning reason to ask students to explain why. A meta-analysis of 64 reports found that prompting learners to explain material to themselves tended to support learning, and did so more than giving them an explanation provided by a teacher or the materials.¹¹ That is evidence about how generating an explanation helps learning, not evidence that an oral assessment is valid or resistant to AI misuse.

Treat AI assistance as a design choice

Name when AI is unavailable, when it may be used, what must be disclosed, and what the learner remains responsible for. If a tool supplies a suggestion, the student should be able to test it, explain whether it belongs, and show what changed because of it.

Protect the teacher's role

Teachers connect disciplinary goals to the students in front of them. They notice misconceptions, motivation, context, and social dynamics that a generic system cannot fully know. Technology may help prepare options or scaffold inquiry, but educational judgment remains a human responsibility. Read our companion resource on teacher-led educational AI, the practical educator and school pathways, and our responsible AI principles.

Questions a school can ask now

• What should students know and be able to do without AI?
• Where can AI extend inquiry without replacing the thinking the task is meant to develop?
• What evidence of process will students and teachers keep?
• When will learners explain, demonstrate, defend, or revise their work with other people?
• What support do teachers need to plan and sustain this kind of learning well?

Real-world learning is not valuable because it sounds modern. It is valuable when it gives learners disciplined practice in using knowledge with agency, judgment, other people, and consequences. In an AI-rich world, that practice becomes easier to overlook—and more important to design deliberately.