The Language Problem in Academic Integrity Policy
The biggest issue with today’s academic integrity policy — and arguably the biggest issue that hasn’t been addressed — is the lack of specificity in the terminology. The lack of specificity in the terminology is creating results that don’t stand up to scrutiny.
I want to be clear about this, since I see a lot of resistance when I talk about this line between AI-generated content, AI-assisted content, and AI-influenced content. The line is not too thin. It matters.
Three types of use that policy conflates:
We generally refer to the use of AI as one of three possible actions (each action has different implications for ethics, education, and policy):
AI-generated Content — Text generated by a language model that the student submitted as their own work with no meaningful transformation or intellectual contribution; the model created the argument and the student provided it. This is the most obvious case of academic dishonesty and the case that academic integrity policy is intended to handle.
AI-Assisted Content — Work completed with the aid of AI tools during the creation process. However, the student maintained significant intellectual property rights to the argument and selected evidence. There are many examples of AI-assistance including, but not limited to: getting a synonym from a chatbot; using a chatbot to create an outline that the student then rewrote extensively.
AI-Influenced Content — Work completed using AI. However, there was no direct text generation. An example could include a student using ChatGPT to describe a topic prior to writing about it from their own knowledge base. Similarly, a student using a model to provide feedback on a draft and subsequently rewriting based upon that feedback. Most universities do not intend to capture this category through their policies. Nevertheless, this is often captured by the language used in university policies.
Harm caused by Conflation
Most institutional AI-policies prohibit “the use of AI” without stating which category(s) are prohibited. As such, a student that uses AI for brainstorming purposes and then writes his/her own arguments may find themselves in the same policy position as a student that submits completely AI-created text due to the fact that the policy does not differentiate between the two scenarios.
The conflation causes two major issues. Firstly, it makes the policy overly broad. In other words, it captures behaviors that institutions likely did not intend to ban and that are not clearly understood regarding their educational impact. Secondly, it renders the policy unenforceable relative to its present formulation. Detention tools cannot reliably discern between these categories. For instance, a detention tool that flags text as potentially generated by an AI cannot tell you whether the student wrote the entire piece, wrote a draft and significantly modified it, or developed his/her ideas via AI-assisted brainstorming and then wrote independently.
When both students receive the same outcome — i.e., an academic integrity proceeding — the policy is not calibrated to what it is attempting to prevent.
Students experience this ambiguity as frustrating and debilitating. Many students are simply unsure whether they violated policy. They used a chatbot to better understand a concept. They asked a model for feedback on a paragraph. They ran their outline by an AI and then rejected virtually all of what it suggested. Did they violate? The policy does not indicate. And that uncertainty is not neutral in terms of pedagogy. If students are uncertain whether they are violating policy they cannot fairly attempt to comply.
There is another problem with consistency that poses risks to institutions. When policy language is sufficiently vague to allow individual faculty members to interpret policy in whatever manner they choose, the application of the policy becomes a function of individual interpretation rather than the actual content of the policy. Thus, one instructor finds fault with a student for using AI for brainstorming purposes. Conversely, another instructor in the exact same department with similar students and factual circumstances does not take similar action against a student for the same reason. Both instructors are applying the exact same policy. This is not an imaginary situation. It represents what occurs in large-scale applications of vague policy language. Students challenged to demonstrate that they were treated unfairly in these proceedings have strong grounds to argue about inconsistencies in the fairness and consistency of the process and institutions will face considerable challenges defending it.
In addition to consistency problems, instructors bear a heavy burden in enforcing these policies. Instructors are being called upon to evaluate and enforce behaviors that detection tools cannot reliably identify; they must judge whether the process used by a particular student to complete an assignment is reasonable, and they must do so within a framework established by a policy that provides insufficient guidance. This is an unfair burden on individuals, and it generates exactly the inconsistency it seeks to eliminate.
Good Policy Language
Policy that can be supported by merit specifies what behavior it intends to restrict with sufficient precision that students know where to draw boundaries for compliance and reviewers can follow those boundaries. A policy that states students shall not submit AI-generated content as their own work represents a far narrower and therefore more defensible statement than one prohibiting “the use of AI”. The former addresses the quality of the work product. The latter focuses on a method of production that detection tools cannot accurately replicate.
Universities also have an obligation to inform students regarding these distinctions, not just administrative personnel. The empirical evidence demonstrates students who are informed as to what specific protections their institutions seek to afford regarding their academic integrity policies and why; comply with those policies at much greater levels than students who are simply presented lists of prohibitive practices.
A number of educators have identified one potential viable solution worthy of serious consideration; namely assignment level disclosure. Instead of implementing general course-wide prohibitions, instructors state what uses of AI are permissible for each assignment — and students disclose what tools were utilized for each assignment. A student disclosing he/she used an AI model to outline an assignment argument and then wrote every section independently is not equivalent to a student submitting entirely AI-produced text. The disclosure establishes documentation. It changes the pertinent question from “Were you utilizing AI?” to “Did you appropriately represent your work?” The latter question is something that academic integrity proceedings can ultimately answer. Furthermore, disclosure mirrors evolving standards in professional settings — journalism, law, medicine — where AI support is rapidly increasing, and disclosure is increasingly becoming commonplace.
Over the last two years, I have reviewed numerous institutional policies addressing AI usage. The majority of these policies contain wording that if uniformly applied would either prohibit practices institutions do not actually plan to prohibit or would prohibit practices they do plan to permit. The inconsistent nature of these policies is not an inconsequential drafting error. It invites arbitrary enforcement.
Detection Tool Limitations
Current detection technology cannot reliably distinguish between AI-generated content and AI-assisted content. Proofademic’s sentence-level analysis allows a reviewer to identify which particular passages contained statistically identifiable features consistent with AI-based generation — but identifying those features does not tell you whether the student authored that passage, whether someone else authored it for them, or whether they wrote it after developing extensive ideas through AI-assisted drafts.
Institutions need to do the harder work first: deciding what behavior they actually want to protect, and why. Do they fear students will not develop writing skills? Then policy should be oriented around demonstrated competency, not tool prohibition. Do they fear deceptive submission? Then they should define deception precisely and calibrate sanctions proportionately. Detection supports enforcement. It does not substitute for the underlying policy clarity that enforcement requires.