Put an unassisted checkpoint inside every big task
Not a surveillance measure. A baseline: one small piece of the work, done in the room, that shows what the student can do alone.
Aug 17, 2026 · 1 min read
The risk with AI-assisted work isn't just misconduct; it's that assistance masks the learning gap until the exam reveals it. In the Bastani et al. (2025) field experiment, students with unrestricted AI looked great during practice and 17% worse when the tool was removed. An unassisted checkpoint, one paragraph, one solved problem, one sketch of the plan, done in class partway through a big task, surfaces that gap while there's still time to close it. (Manning's education-agent-skills library formalises this as the unassisted-evidence-checkpoint skill.)
In practice
- Schedule it mid-task, not at the end: ten in-class minutes producing a small artefact that the final submission must visibly grow from.
- Frame it as support, and mean it: "this is where I find out who needs help before the deadline," not "this is where I catch you."
- A big mismatch between checkpoint and final piece starts a conversation about process (see "Ask how they used AI, not whether"), with evidence instead of suspicion.
Quick takeaway: build one AI-free moment into the task, and the final grade stops being a guess.