AI-generated code can be part of the learning process, but decide which skills the teenager must still demonstrate without generation. Do not make either total prohibition or unrestricted generation the default for every activity.
Stanford’s November 10, 2026 evidence review finds a mixed picture: using AI to understand material appears more promising than using it to offload work, but neither use consistently improves learning. That argues for inspecting the activity and evidence rather than assuming a tool label determines educational quality. (scale.stanford.edu)
Our suggested agreement divides work into three categories:
- Core learning task: the learner first attempts the concept or component being taught.
- Assisted work: AI may suggest code, but the learner identifies its purpose, tests it and explains accepted changes.
- Scaffolding outside scope: the mentor may provide boilerplate, clearly marked as supplied rather than mastered.
For example, a teenager learning evaluation need not independently construct an entire web interface. But they should own the test cases and explain what the results do—and do not—show.
Avoid relying on an authorship detector as the renewal decision. Our preferred evidence is a live explanation, a small changed requirement and a transparent contribution record.
Sources cited on this page
- The Evidence Base on AI in K-12: A November 2026 Review ↗scale.stanford.edu