If your school won't say yes or no to AI on assignments, how do you decide what's fair anyway?

This explores how a teacher can decide what to do about generative AI in assessment when no institutional rule tells them what's allowed. One caveat first: the collection has no study of teachers acting without permission. What it offers instead is a set of findings that change how the decision looks. The most useful reframe comes from interviews with 20 university teachers, which found that the GenAI assessment challenge meets all ten criteria of a 'wicked problem' Does GenAI assessment challenge fit wicked problem theory?. It has no agreed definition, no point at which it's solved, and only better or worse responses rather than right answers. That matters for a teacher waiting on permission. If the problem is wicked, a definitive institutional policy may never arrive, and the realistic job is to make judgments you can defend and keep revising them.

There's also evidence that rules may matter less than people assume. In a randomized experiment on peer review at ICML 2026, banning LLM use and allowing limited use produced almost the same scores, decisions and reviewer confidence. Meanwhile, a large share of reviewers broke whichever rule they were given Does banning LLM use in peer review change review outcomes?. Peer review isn't a classroom, but the lesson carries over: a ban or permission on paper doesn't settle what actually happens. That moves attention away from 'am I allowed?' and toward assessment design. Work on the GPQA benchmark shows that carefully written expert questions can stay hard even for skilled people with full web access Can expert-written questions resist web-assisted non-expert answering?. That suggests it's possible to design tasks that are hard to answer just by looking things up, rather than relying on detection or prohibition.

A less obvious finding concerns openness. Interviews with knowledge workers show that people often hide signs of GenAI use, partly to avoid stigma but also to signal their own expertise. That hiding cuts off the informal peer learning organizations need to adapt Why do knowledge workers hide signs of using GenAI?. For a teacher acting without clear permission, the temptation is to experiment quietly. The research suggests that sharing your reasoning with colleagues may be what helps a department work out its norms when policy hasn't caught up.

If you're thinking of using AI on your own side of assessment, the evidence splits. For writing questions, ChatGPT-generated practice items matched published textbook questions on difficulty and on how well they told stronger and weaker students apart Can AI generate assessment questions as good as human experts?, so this is a relatively low-risk place to start. Grading is riskier. AI evaluators give higher scores to answers with fake references or polished formatting regardless of content Can LLM judges be tricked without accessing their internals?. They also tend to prefer text they recognize as their own Do LLMs favor their own text because they recognize it?, and they forgive rule-breaking when they think a human wrote the work Do authorship labels change how AI judges evaluate rule violations?. In practice: using AI to help draft questions is easy to defend, but letting it grade students on its own is hard to justify, whatever your institution has or hasn't approved.