College Board lets AI help early, but not the final submission

The classroom note

A teacher running AP Seminar or AP Research works inside a rule that changes by stage of the project, as set out in the College Board's own guidance for artificial intelligence tools, current as retrieved on 16 September 2026. Generative AI is allowed for 'exploration of potential topics of inquiry, initial searches for sources' and for 'confirming their understanding of a complex text', language the policy uses for AP Capstone courses specifically. What is not allowed is using a generative tool to produce the finished, submitted piece of work. For a student, the same software can be a legitimate research aid in September and a disqualifying shortcut in the piece handed in for a grade in the spring.

What the evidence says

The policy is the College Board's own statement of its rules, not an independent study of how AP students actually use AI; it makes no claim about compliance rates. What it does specify, in detail, is enforcement mechanics for the two research courses: AP Seminar and AP Research require students to 'complete checkpoints with their teacher to demonstrate genuine engagement', with a teacher affirmation of authenticity attached to each one, and 'failure to complete the checkpoints will result in a score of 0 on the associated task'. The AP Seminar course page confirms these are 'Required Checkpoints' with a fixed deadline for teachers to submit their affirmations, a real administrative date attached to an otherwise judgement-based policy.

The implementation question

The checkpoint model shifts cost onto the school calendar rather than onto software: a teacher must schedule and hold a series of interim check-ins with every student across a course, and sign an affirmation for each one, before the associated task can score above zero. That is instructional time a purely detection-based policy would not require, but it avoids relying on a detector's accuracy for a stake as high as an AP score. The policy 'reserves the right to investigate submissions' after the fact, a backstop, but checkpoints are presented as the primary mechanism, so most verification happens during the project, not after submission.

What holds and what fails

What holds is a rule that names the stage rather than the tool: because the boundary is exploration versus final artefact, it survives a change of AI product without being rewritten. What is likely to fail, editorially, is a checkpoint system in a course where a teacher cannot realistically meet with every student individually at each interim stage; the policy assumes a level of contact time the document itself does not guarantee is scheduled or funded. A score of zero for a missed checkpoint is a serious consequence resting on a logistical assumption.

• How many checkpoint meetings has this teacher actually been able to hold with each student this term?
• Is the line between exploration and final submission clear for this specific assignment, or only in the abstract?
• What does the affirmation a teacher signs actually attest to, beyond attendance at a meeting?

A policy built on checkpoints trades detector uncertainty for calendar and staffing certainty, which is a reasonable trade only where the calendar and staffing exist to support it.