Why AI Detection Cannot Be the Foundation of Academic Integrity
Generative AI changed what a submitted assignment means.
For decades, a paper or written response served as evidence that a student had done the thinking behind it. The system was never airtight. Students plagiarized, hired others, or skimmed a reading and still submitted passable work. Even so, the finished product usually carried enough instructional signal to support grading. That assumption is much weaker now, and most faculty already recognize it.
Students use AI for brainstorming, outlining, drafting, revision, source summaries, grammar, citations, and entire sections of text. Some uses violate course policy. Some are allowed. Others fall into a gray area shaped by the instructor, assignment, and course. The greatest concern for many faculty is simple: AI can make shallow understanding look like polished writing.
That uncertainty has made academic integrity conversations more tense. Faculty must evaluate work that may no longer show what a student understands. Institutions want to protect the value of learning without turning classrooms into systems of suspicion. Students face policies that vary by instructor, course, department, and tool.
AI detection entered this situation as an apparent solution. The promise was straightforward: analyze the text and determine whether AI wrote it. When so much feels uncertain, a number on a screen feels reassuring. Detection has limited value as a prompt for further review, but a probability score cannot bear the full weight of an academic integrity decision.
Detection scores are probability estimates
AI detectors do not know who wrote a sentence. They scan text for patterns associated with model-generated writing and estimate the likelihood of AI involvement. The result is a probability estimate, not a finding of fact.
OpenAI discontinued its own AI text classifier in 2023 because the tool did not perform well enough to continue. In one evaluation, OpenAI reported that the classifier identified only 26 percent of AI-written text as “likely AI-written” and incorrectly labeled human-written text as AI-written 9 percent of the time. The company also warned against using the classifier as a primary decision-making tool. (OpenAI)
If the company behind the model could not reliably identify output from its own system, then classroom decisions built around detection scores rest on weak evidence. A score might justify a closer look. The distance between a flag and a finding of academic misconduct is still substantial, and integrity decisions require more evidence than a probability estimate provides.
False positives sound small until institutions scale them
Vendors often highlight low false-positive rates. A 1 percent false-positive rate sounds minor in isolation. Across an institution, the numbers change quickly.
Jisc’s 2025 update on AI detection and assessment notes that a 1 percent false-positive rate across hundreds of thousands of annual assessments produces thousands of false accusations. Its example of an institution handling 480,000 assessments each year produces roughly 4,800 false positives annually. (Artificial intelligence) That means thousands of students may face unnecessary scrutiny, along with the staff time required for meetings, reviews, and appeals. The larger cost is damage to trust, which rarely appears in a spreadsheet.
Turnitin acknowledges that false positives occur. The company advises instructors to consider professional judgment, knowledge of the student, assignment context, and institutional policy when interpreting AI-writing data. Turnitin also advises faculty to plan for false positives and give students the benefit of the doubt when evidence is unclear. (turnitin.com)
That guidance is sensible, and it points to the central issue. A detector that requires instructor judgment, classroom context, and a fair appeal process is one input in a larger human process. The score should be treated accordingly.
Detection creates equity concerns
AI detectors rely on patterns such as predictability, sentence regularity, and word choice. Those features describe the statistical shape of writing. They do not establish authorship. A careful human writer, an AI system, or a combination of both may produce similar patterns.
Multilingual writers face the greatest risk. Stanford HAI reported that seven AI detectors classified more than half of TOEFL essays written by non-native English students as AI-generated. The researchers explained that detectors often rely on measures such as perplexity, which penalize writing that is simpler, more predictable, or less syntactically complex. (Stanford HAI) The underlying study in Patterns stated the problem plainly: GPT detectors frequently misclassify non-native English writing as AI-generated, raising fairness concerns in educational settings. (ScienceDirect)
A system intended to identify AI use is statistically more likely to suspect students whose English is still developing. An integrity process built around that signal directs more scrutiny toward students already working under more difficult conditions. Honest students deserve a process that protects them. No student should need to prove their humanity because their syntax matches a statistical pattern.
Detection weakens as AI use becomes ordinary
“AI use” no longer describes one behavior. One student asks AI to summarize a difficult reading. Another organizes study notes with it. A third uses it to check grammar in self-written work. A fourth asks for a more professional version of an awkward paragraph. A fifth generates an entire essay and submits it without reading it.
Those actions are academically different, yet a detector sees only the final text. It reduces every form of assistance to one signal because it has no access to the student’s process, judgment, or understanding.
This is why detection-centered policy keeps collapsing into one question: Did AI write this? That question no longer does enough academic work when AI assistance is woven into brainstorming, drafting, revision, and editing. A more useful question asks whether students can explain and defend their submissions, and whether the work shows evidence of their own thinking.
The cat-and-mouse problem is real
Detection operates within an arms race. Students seeking to avoid detection paraphrase AI output, use rewriting tools, change sentence structures, combine personal writing with model output, or prompt for less predictable prose. Jisc describes this as a cat-and-mouse dynamic and notes that paraphrasing and manual manipulation reliably reduce detection effectiveness. (Artificial intelligence)
Turnitin’s release notes show how quickly the tools shift. Its AI-writing model has been updated repeatedly to address recall, false positives, AI paraphrasing, bypasser tools, and language coverage in Spanish and Japanese. (guides.turnitin.com) A detector result reflects one tool at one moment against one version of model behavior. Student practices, available tools, and bypass strategies continue changing beneath it. Academic integrity decisions should not rest on evidence so unstable.
Detection-first teaching changes the classroom
When detection sits at the center of academic integrity, the emotional climate of a course changes. Faculty begin reading every paper for signs of misconduct rather than for evidence of learning. Students begin writing defensively. Strong prose feels risky. Plain prose feels risky too. Students start saving version histories, screen recordings, draft trees, and keystroke logs in case they later need to prove authorship.
The assignment becomes evidence in a case rather than a record of learning.
That tradeoff is difficult to justify. A healthy integrity system gives faculty meaningful evidence without requiring them to infer misconduct from instinct or a probability score. It also gives honest students a fair way to demonstrate what they know. The strongest work happens through assessment design, not detection scoring.
The stronger foundation is verification of understanding
Academic integrity in the AI era should rest on evidence of learning. Ask students to explain, defend, apply, and revise the work they submit. Treat the finished artifact as one source of evidence and the student’s explanation as another. Design assessments where AI assistance does not erase the assignment’s purpose because students still need to own the ideas in conversation.
The principle is familiar. Oral exams, thesis defenses, design critiques, and qualifying exams have long recognized that a finished artifact is incomplete on its own. Students at advanced levels are expected to explain their work because explanation reveals understanding in ways a polished document does not.
Scale has always been the challenge. A professor can conduct oral defenses with five graduate students without much difficulty. Doing so with 80 undergraduates requires more planning and administrative support. Yet when AI can produce the final product, a student’s ability to explain that product becomes the clearest academic signal. The scale problem deserves attention rather than avoidance.
What this means for faculty
This is not an argument against written work. Writing, research, drafting, and revision still contain much of the learning. The needed change is in the assessment layer around the writing.
A stronger AI-era assignment might require students to submit a paper and participate in a short, targeted conversation about it. The conversation can stay focused:
- Why did you select this source rather than another?
- Where did the hardest thinking happen?
- How did you use AI, if at all?
- What do you understand now that you did not understand at the beginning?
These questions reveal what students own intellectually. They also give faculty evidence that is harder to fake than polished prose alone.
What this means for institutions
For institutions, AI detection belongs within a broader integrity framework that includes policy, assessment design, faculty development, student AI literacy, and transparent appeal procedures. A serious institutional approach treats those elements as a connected system. Assessment methods that produce direct evidence of student understanding should carry the most weight.
Clear policies tell students what acceptable AI use looks like. That work matters, but policy does not establish what occurred in a particular submission. Detection sometimes offers a guess. Assessment design provides the stronger evidence: whether the learning an assignment was meant to produce is present in the student’s understanding.