We have now read a few hundred syllabus AI policies, partly out of professional interest and partly because instructors keep sending them to us to ask whether they are any good. Most of them are not, and they fail in three specific ways that are easy to fix.
Failure one: the policy is a mood, not a rule
"Use AI responsibly and with academic integrity." A nineteen year old at eleven at night cannot act on that sentence. Responsible according to whom, at what point, for which part of the task?
The fix is to write the policy as a list of activities rather than a principle. Brainstorming, outlining, grammar correction, translation, code debugging, generating a first draft, generating the final text. Say yes or no to each, per assignment type if they differ. The vagueness feels generous when you write it and reads as a trap when a student is trying to comply.
Failure two: the policy has no enforcement story
A rule nobody can check is a suggestion, and students work this out immediately. If your policy says AI-generated prose is not permitted and your only check is a detector you privately do not trust, the policy is decorative.
This is uncomfortable, and it is worth writing down honestly. Either the rule is checkable, or it is a norm you are asking people to hold to voluntarily, which is a fine thing to ask for as long as you do not pretend it is enforcement.
A policy that cannot be checked is not a policy. It is a hope with a font.
Failure three: the policy punishes disclosure
Plenty of policies say "you must disclose any AI use" and also "undisclosed AI use is misconduct", without saying what happens to a student who does disclose. If honest disclosure and getting caught lead to a similar-looking conversation, you have taught your students that the safe move is silence. Say explicitly what disclosed use costs, and make it cost less than concealment.
Wording you can adapt
Take this as a starting point rather than legal text, and check it against your institution's rules, which override anything here.
- Permitted without disclosure: spelling and grammar correction, translation for comprehension, and asking a model to explain course material to you.
- Permitted with a one-line note at the end of your submission: brainstorming, outlining, generating example code you then modify, and using a model as a study partner. The note says what you used it for. It does not affect your grade.
- Not permitted: submitting model-generated prose, code, or analysis as your own, or using a model during an assessed conversation or exam.
- How this is checked: you may be asked to talk through any submitted work. Being able to explain your own choices is the standard, and it applies to every student equally rather than only to work that raised a flag.
That last line does more work than the rest combined. It moves the burden from proving what a tool did to demonstrating what you understand, which is both fairer and closer to what you wanted to grade in the first place.
If you want the standard we hold ourselves to when a conversation is used this way, it is written up at the Viva standard.