Can Universities Detect AI-Written Work?

Not reliably enough to prove academic misconduct on their own. Universities can use AI detectors to flag writing that may deserve review, but a detector cannot establish who wrote an assignment or how a student used AI. A fair decision needs other evidence: the course rules, the student’s work process, drafts and sources, and sometimes a conversation about the work itself.

What an AI detector actually detects

AI-writing detectors do not identify a hidden watermark or retrieve the history of a document. They look for statistical patterns in prose that resemble text produced by language models.

That can be useful as a signal. But a signal is not proof of authorship.

Turnitin, the most widely used tool in this area, says its AI Writing Report can misidentify human-written, AI-generated, and AI-paraphrased text. It explicitly says the score should not be the sole basis for adverse action against a student. Read Turnitin’s guidance.

The tool itself reflects that uncertainty: Turnitin does not show precise AI scores between 1% and 19%, because false positives are more likely in that range.

Why a score is not a verdict

A detector may flag writing that is entirely human-written. It may also miss AI use, especially when a student has substantially rewritten a draft, used AI only for planning or editing, or worked in a format the tool does not assess well.

The problem is not just technical. Research has found that several AI detectors were much more likely to flag TOEFL essays written by non-native English speakers as AI-generated. That creates an obvious fairness risk: predictable or formal English can begin to look suspicious simply because it differs from the detector’s idea of natural writing. Read the Stanford HAI summary.

A high score may justify a closer look. It does not tell a university whether the student broke a rule, used AI in a permitted way, or simply wrote in a style the system misunderstood.

What universities can actually investigate

If a university suspects inappropriate AI use, the useful question is not “What percentage did the detector produce?” It is: “What evidence do we have about this student’s process and understanding?”

That evidence can include:

• the assignment rules and what AI use was permitted;
• drafts, notes, citations, and version history;
• whether the sources and claims in the work hold up;
• whether the student can explain key choices, evidence, and reasoning;
• whether the submitted work matches other observed work in a meaningful way.

A short discussion or oral defense can help test current understanding. It does not prove that AI was never used, and it must be handled carefully: anxiety, language, disability, or a poorly designed conversation can all affect performance.

The aim should be a fair academic-integrity process, not a technological guessing game. Jisc’s current guidance makes the same distinction: detection tools may inform a wider inquiry, but decisions should never rest on a detector alone. Read Jisc’s 2025 update.

What students should do

The practical rule is simple: follow the policy for that specific course, disclose AI use when required, keep your research and drafting trail, and be able to explain the work you submit.

That is not just defensive record-keeping. If AI helped with an outline, feedback, coding, or language editing, understanding the final argument still matters more than producing a polished document.

AI Caver take

Universities can sometimes detect suspicious patterns. They cannot reliably turn those patterns into proof that a student cheated.

The deeper problem is that a polished essay, report, or piece of code no longer proves as much as it once did about the person who submitted it. The durable answer is not better AI policing. It is assessment that combines open work with credible, fair ways to verify understanding.

Read the full research

For the wider question of what student work still proves when AI can help produce it, read The Breaking Point for Higher Education: Student Work Is Getting Better. The Learning Is Not..