AI Detectors: How They Work and Why They Are Unreliable as Proof

When AI writing tools took off, so did a market for tools that claim to detect them. Some institutions subscribe to such detectors, some ban their use in misconduct decisions, and many students live with a low level of worry that an honest essay will be flagged. The truth about these tools is less comforting than the marketing, and understanding it can help you protect yourself.

This guide explains how detectors work, why they are error-prone, what institutions do with the results and what to do if your own work is flagged. It describes general principles, since individual tools and policies change quickly.

What an AI detector does

An AI detector analyzes text and estimates how likely it is to have been produced by an AI system. It does not know how the text was written. It looks for statistical patterns that tend to differ between human and machine writing, and returns a score, a percentage or a label.

It is a classifier, trained on examples of human and AI text. Like any classifier, it will sometimes be wrong, and the more different the new text is from its training examples, the more likely it is to err.

How they generally work

Details vary and are often proprietary, but common ideas include:

• Predictability. Text generated by AI tends to be statistically more predictable word by word. A detector may measure how surprising each word is to a language model.
• Variation. Human writing often varies more in sentence length and structure, while AI text can be more uniform.
• Learned patterns. Some detectors are trained directly on large sets of labeled human and AI texts, and learn features that distinguish them.
• Watermarks. Some research explores hidden statistical patterns added deliberately by AI developers, which a detector could look for. This only works if the text came from a system that adds them and has not been altered.

None of these gives certainty. Predictable text is not necessarily machine text, and varied text is not necessarily human.

Why they make errors

Because of these errors, detection scores are better seen as weak signals than as findings.

Bias against some writers

Research has found that some AI detectors are more likely to misclassify writing by non-native English speakers as AI-generated. One reason is that writers with a smaller range of vocabulary or more cautious sentence structures can produce text that looks more predictable to a detector. Students who write in a second language, who are neurodivergent or who have been taught a rigid academic style may be at higher risk of a false flag. This is one of the main reasons why many universities and educators warn against relying on detectors.

What they cannot tell you

A detector cannot tell you:

• Who wrote the text.
• How it was produced, such as whether AI was used for brainstorming or for drafting.
• Whether use of AI was allowed in the course.
• Whether the student understands the work.

A score is not evidence of misconduct. At most, it can prompt a conversation.

How institutions use them

Practice varies widely.

• Some institutions license a detector and make results available to instructors, who treat them as one piece of information.
• Some have disabled or banned detection tools because of reliability and fairness concerns.
• Some provide guidance that a detector flag must never be the only evidence in a misconduct case.
• Instructors may use their own judgment, comparing the work with your previous writing, asking questions about the content or discussing your process.

Check your institution's policy, which should say how detection results are used and what process follows.

If your work is flagged

Being flagged is stressful, and it is not proof. A calm, organized response helps.

1. Read the notice carefully, and note deadlines and what is alleged.
2. Do not panic, and do not ignore it.
3. Gather evidence of your process: drafts, outlines, notes, research, browser history, version history in your word processor and any messages about the work.
4. Be ready to explain your work: the argument, the sources, the choices you made. Instructors may ask questions about the content.
5. Ask what evidence exists, including the report itself, and how it was interpreted.
6. Point to the known limits of detectors, politely, and ask how the institution treats a flag on its own.
7. Seek support from a students' union advocacy service, an adviser or an ombudsperson.
8. Be honest. If you did use AI in a way you think may be a problem, say so and explain.

Our guide to academic integrity and misconduct describes how cases usually work and your rights.

Protect yourself in advance

You cannot control a detector, but you can control your evidence.

• Write in a tool with version history, such as a word processor that saves revisions.
• Keep your notes, outlines and drafts.
• Save your sources and annotated readings.
• Write in your own voice, and be able to explain every part of the work.
• Follow the course AI policy, and disclose any permitted use.
• Avoid pasting in text from tools, even if you plan to edit it, unless allowed.

This documentation also shows your learning process, which is useful for your own revision.

Should you check your own work?

Running your own essay through a detector can be tempting, but the results are unreliable. A low score does not prove that the work will pass, and a high score on honest work can cause needless anxiety. Rewriting your work to lower a score can make it worse, and trying to avoid detection by paraphrasing tools is a bad idea. A better approach is to write honestly and keep records. Privacy is also a consideration, since pasting your work into an online detector may store it. See our guide to privacy when using AI tools.

What instructors can do better

Many educators prefer approaches that do not depend on detectors: assignments with personal or local components, drafts and process checkpoints, oral discussion of submitted work, in-class writing and clear AI policies. If your instructor uses such methods, take part in them seriously, since they reduce the chance of misunderstandings.

A note on tone with your instructor

If you are asked about your work, stay courteous even if the question feels like an accusation. Most instructors want to understand, not to punish. Offer to walk through your outline, show your sources and describe how the argument developed. A student who can discuss the work fluently and show drafts usually resolves a flag quickly, and a respectful conversation keeps the process fair for everyone involved.

Common mistakes

• Treating a detector score as proof.
• Panicking and deleting drafts.
• Not keeping version history.
• Rewriting honest work to chase a lower score.
• Using a detector on someone else's work and accusing them.
• Assuming a flag means you cannot defend yourself.
• Not checking the institutional policy.

A quick checklist

• Do I know how my institution uses detectors?
• Do I keep drafts, notes and version history?
• Can I explain every part of my work?
• Do I follow the course AI policy and disclose permitted use?
• Do I know where to get support if flagged?

Final thoughts

AI detectors are imperfect statistical tools, and they should not decide anyone's academic future on their own. The best protection is to do your own thinking, follow the rules, keep a trail of your process and know your rights. If a flag does come, a calm, well-documented explanation is the strongest response.