The first failure mode is abstraction. A policy that says faculty must address AI use in their syllabi gives no faculty member anything actionable on Monday morning. The second failure mode is scope creep. A document that tries to cover undergraduate essays, graduate dissertations, lab notebooks, and creative writing in one paragraph ends up saying nothing useful about any of them. The third failure mode is the absent author. When a policy reads as if it were written by a committee that has not taught a section in five years, faculty notice, and compliance drops accordingly.
The short version
Policies faculty follow are short, specific to assignment types, written with faculty rather than at them, revisited on a known cadence, and connected to existing academic integrity infrastructure. Mandates that ignore any of those five conditions tend to die quietly in inboxes. The Harvard FAS guidance, the Massachusetts DESE K-12 document, and the TeachAI policy library all converge on roughly the same principles, which suggests the pattern is real.
What follows is a breakdown of why most AI policies do not survive contact with actual teaching, the five principles that distinguish the ones that do, and three sample paragraphs your department can lift, edit, and adopt. We cite the exemplars we think are worth reading at the end.
Why Most AI Policies Don't Get Followed
There is a fourth failure mode worth naming, which is the punitive register. Policies that lead with sanctions before they explain expectations encourage faculty to either over-report, under-report, or quietly stop reporting at all. The Harvard FAS generative AI guidance is instructive on this point: it spends its first several paragraphs on pedagogical framing before it gets to anything resembling enforcement language, and that ordering is doing real work.
The worst AI policy is the one a committee writes in isolation. The second worst is the one a committee copies without adaptation. The best is the one a department writes together and revises annually.
Working Educators editorial position
The Five Principles
First, be short. A policy that runs longer than two pages will not be read. The Massachusetts DESE K-12 AI guidance is a useful counterexample in that it runs long, but it is a state-level reference document rather than a classroom policy. The document a faculty member needs on a Tuesday afternoon should fit on a single sheet.
Second, be specific to assignment type. A blanket rule about AI use is functionally a blanket rule about nothing. The kinds of assistance appropriate for a take-home essay differ from those appropriate for a timed lab report or an oral examination, and the policy should name those differences explicitly. The TeachAI policy resource library includes several district-level documents that handle this granularity well.
Third, write with faculty, not at them. A policy drafted by a small subcommittee and circulated for a two-week comment period will produce something faculty can defend in a hearing. A policy delivered as a fait accompli will produce a document faculty actively work around. The AASA superintendents' guide makes this point explicitly in its governance section.
Fourth, name a review date. Any AI policy written in 2026 will need substantive revision by 2027. Building the review cadence into the policy itself signals that the document is alive rather than inscribed in stone, which lowers the political cost of objecting to a clause that turns out not to work.
Fifth, hook the policy into existing academic integrity infrastructure rather than building a parallel one. Faculty already know how to file an integrity case, run a conduct conversation, and document a concern. A new AI policy that requires a different reporting pathway will not be used. The CoSN and ISTE joint resources emphasize this integration repeatedly.
Sample Policy Paragraph: Acceptable Use
Here is a paragraph any department can adapt. Strike what does not apply, add what does, and run it past two faculty who teach the relevant courses before adopting it.
In this course, generative AI tools may be used for brainstorming, outlining, and revising prose you have already drafted. They may not be used to produce first drafts of work you submit for credit, to write your reflective or analytical responses, or to generate citations. If you are uncertain whether a specific use is acceptable, ask before submitting. Disclosed AI use is an academic conversation; undisclosed AI use is an academic integrity matter and will be handled under the existing integrity policy at the institution's standing procedures.
Two notes on this paragraph. It distinguishes between revision and generation, which is the distinction faculty most often want and most often fail to articulate. It also collapses the disclosure question into a single sentence: ask first. Policies that try to enumerate every permissible use case become unreadable; policies that establish a default of inquiry stay usable.
Sample Policy Paragraph: Disclosure Expectations
Disclosure is the policy lever that does the most work for the least friction. Here is sample language.
Any use of generative AI tools in work submitted for credit must be disclosed in a short note at the end of the submission. The note should identify the tool, describe what it was used for, and indicate what portions of the work, if any, contain text generated or substantially rewritten by the tool. Disclosure does not by itself constitute permission; the acceptable use clause above governs what is permitted. Failure to disclose is treated as a separate matter from the underlying use.
The mechanism here is that failure to disclose becomes its own offense, independent of whether the underlying use would have been acceptable. This shifts the incentive from concealment to transparency without requiring faculty to adjudicate ambiguous cases. We discuss the limits of this approach in our piece on contested AI accusations, which is worth reading alongside any disclosure policy.
Sample Policy Paragraph: Reviewing New Tools
The third paragraph addresses what most policies omit, which is who decides whether a new tool is sanctioned. Sample:
New AI tools introduced into the course context, including detection tools, paraphrasing tools, and writing assistants, will be reviewed by the department's curriculum committee before being recommended, required, or prohibited. Faculty members proposing a tool should submit a brief written rationale describing its intended use, its data handling practices, and its known limitations. Tools will not be added to required reading lists or assignment workflows without committee review.
This paragraph does two things. It puts detection tools on the same footing as generative tools, which is appropriate given the empirical track record of both, and it prevents the all-too-common pattern where one enthusiastic faculty member mandates a tool that the rest of the department later has to defend. Our analysis of GPTZero accuracy and our review of Copyleaks are both relevant background reading for any committee evaluating a detection tool.
Where Existing Models Help
Several documents are worth reading before drafting. The Harvard FAS guidance is the best example we have seen of an institutional document that treats faculty as professionals rather than risks to be managed. The Massachusetts DESE document is the most thorough K-12 framing, and it includes useful language on equity considerations that most institutional policies omit.
The TeachAI policy library aggregates district and institutional policies from a wide range of contexts, which is helpful for finding language adjacent to your situation. The AASA guide is governance-focused and useful for superintendents and provosts. The CoSN and ISTE materials are the most operationally detailed, with checklists and review templates that translate principles into procedures.
None of these documents are a substitute for the conversation a department needs to have about its own values and its own students. They are useful as scaffolding. Our editorial view is that the worst AI policy is the one a committee writes in isolation; the second worst is the one a committee copies from another institution without adaptation; the best is the one a department writes together, revises annually, and treats as a working document rather than a settled question.
Frequently Asked Questions
How long should an AI policy actually be?▼
For a classroom or course-level policy, one page. For a department or institutional policy, two to three pages with an appendix for examples. Anything longer becomes a reference document rather than a working policy and will not be read by the faculty who most need to apply it.
Should the policy mention specific tools by name?▼
In the body of the policy, no. Tool names date the document and invite obsolescence. In an appendix or living addendum, yes, because faculty need concrete examples of what counts as a generative tool, a humanizer, or a detector. Keep the appendix updatable without reopening the main policy.
What should the policy say about AI detection tools?▼
Our editorial view is that detection tools should be treated under the same review process as any other classroom technology, with attention to known false positive rates and equity implications. The Stanford TOEFL study and other public research on detector accuracy are relevant context for any committee evaluating these tools.
How do we handle faculty who refuse to enforce the policy?▼
This is a management question rather than a policy question, and treating it as a policy question usually makes it worse. The policy should establish expectations; the chair or dean handles individual cases. If a significant fraction of faculty are refusing to enforce, the policy itself probably needs revision.
How often should the policy be reviewed?▼
Annually at minimum, with a named review date written into the policy itself. The field is moving quickly enough that any document more than eighteen months old should be considered provisional until it has been re-examined.
The Bottom Line
A policy that gets followed is one that faculty had a hand in writing, that fits on a desk reference card, that distinguishes between assignment types, that names a review date, and that hooks into the academic integrity infrastructure already in place. None of that is novel. All of it is harder than it sounds, because the easy thing for any committee to do is to write a long document, distribute it, and move on.
Our editorial position is that the document is the smallest part of the work. The conversation that produces the document is what determines whether it survives the semester. If the only people in the room are administrators and a token faculty representative, the policy will be a document. If the room contains the faculty who will be applying it, the policy will be a practice.
The five principles above are not original to us. They show up across the Harvard, DESE, TeachAI, AASA, and CoSN materials in slightly different language. That convergence is the strongest evidence we have that they are right.
