Do AI Detectors Work? AI and Academic Integrity

Do AI detectors work? Not well enough to be treated as proof. They estimate how likely a text is to be machine-written from statistical patterns, and two published tests show they can miss most AI text while flagging real human writing. This guide, part of our academic integrity hub, covers how the tools work, what the numbers say, whether using AI counts as cheating in your course, and how to protect yourself.

How do AI detectors work?

Most AI detection tools are classifier models trained on large samples of human and AI text. They look at how predictable each word choice is (often called perplexity) and how much sentence length and structure vary (burstiness). Text that is very predictable and very even gets a higher AI score. The output is a probability or a percentage, not a finding of fact.

That design explains the main weakness. Predictable, even prose is not unique to machines. Plenty of careful human writing is predictable too, especially short texts, formulaic assignments and writing in a second language.

How reliable are AI detectors?

Two sources give hard numbers on how accurate AI detectors are, and both point the same way.

Source	What it tested	What it found
Liang et al. 2023	91 TOEFL essays by non-native English writers, run through seven GPT detectors	An average false-positive rate of 61.22%, and 97.8% of the essays were flagged as AI by at least one of the seven detectors. The authors warn that detectors may unintentionally penalize writers with constrained linguistic expressions (arXiv:2304.02819).
OpenAI’s AI classifier	A challenge set of English texts	It correctly identified 26% of AI-written text and wrongly labeled 9% of human-written text as AI. OpenAI withdrew the tool on July 20, 2023 because of its low accuracy (OpenAI announcement).

Read together, the two results describe both failure modes. The OpenAI classifier missed roughly three out of four AI-written samples. The detectors in the Liang study flagged human essays at a very high rate. Neither number tells you how accurate a specific current tool is, so treat any vendor’s accuracy claim as something students and instructors should review against independent testing.

Which writing styles do AI detectors flag as false positives?

An AI detector false positive happens when a detector scores human text as machine text. The same pattern check that catches smooth AI prose also catches human writing that is simple, regular or heavily edited. Three situations come up most often.

Situation	Why it can trigger a flag
Writing in a second language	A smaller working vocabulary and plainer sentence patterns look predictable to the model, which is the pattern the Liang study documented.
Short or formulaic assignments	Lab reports, definitions and five-paragraph essays follow a fixed shape, so there is little variation for the detector to read as human.
Heavy grammar or paraphrasing tools	Tools that smooth your sentences make the text more uniform, even when every idea is yours.

OpenAI’s own limits list adds three more. Its classifier was very unreliable on texts below 1,000 characters, performed significantly worse outside English, and could not reliably identify text that is very predictable. A short answer, a list or a formulaic paragraph is exactly the kind of text where an AI detector review is least trustworthy.

Can paraphrasing tools and rewriting beat AI detectors?

Yes, and that is the second half of the reliability problem. The Liang authors showed that simple prompting strategies can bypass GPT detectors, and OpenAI noted that AI-written text can be edited to evade its classifier. So the same tools that wrongly flag honest students can pass text that a student did not write. Paraphrasing or rewriting by hand also changes the pattern a detector reads, which is why a clean result does not show that no AI was used.

What do institutional policies say about AI detection scores?

Students ask whether a detection score is accurate enough to count as evidence, and the sources give a clear answer. OpenAI said its classifier should not be used as a primary decision-making tool, only as a complement to other methods of working out where a text came from. The Liang authors caution against using detectors in evaluative or educational settings. Institutions still set their own rules, so check how your school treats a detector score in its academic misconduct process. A score can start a review, but your drafts and your ability to explain the work carry more weight than a percentage.

Can Turnitin detect ChatGPT?

Turnitin offers an AI writing indicator, and many schools switch it on. It is a separate reading from the similarity percentage and works differently. If you are unsure which number you are looking at, our guide to the Turnitin similarity score explains the report.

Feature	Similarity score	AI flag
What it measures	Text that matches other sources	A statistical estimate that the writing pattern looks machine-made
Can you check it?	Yes, each match points to a source you can open	No, there is no source to compare against
What it can show	That wording overlaps with a known source	Nothing on its own, it is a prompt for a conversation

So Turnitin can flag text it thinks resembles AI output, but a flag is an indicator and not a verdict. The answer to “can Turnitin detect ChatGPT” is that it can try, and it can be wrong in both directions.

Is using ChatGPT plagiarism or cheating?

It depends on the course policy. Classic plagiarism means presenting someone else’s words or ideas as your own, and AI-generated text has no human author to credit, which is why the rules are written course by course. Using AI is cheating when the policy bans it or when you hide it where disclosure is required. The same use can be fine when the instructor permits it and you say so. Our guide to academic dishonesty covers the wider category these cases fall under.

Courses usually set one of a few policy levels.

Policy level	What it usually means for you
No AI allowed	Any AI-generated text or ideas count as unauthorized help. Using it is cheating under the course rules.
Limited use	AI is allowed for named tasks such as brainstorming, grammar checks or outlining, but not for drafting graded text.
Use with disclosure	AI is allowed if you state what you used and how. Skipping the disclosure is the violation.
Open use	AI is part of the assignment. You are still responsible for accuracy and for citing it.

If the syllabus says nothing, ask the instructor in writing before you submit. “I did not know” is a weak defense, and a short email costs you nothing.

How to protect yourself and prove you did not use AI

You cannot prove a negative with a detector, so build evidence of your own process while you write.

• Keep your drafts. Save notes, outlines and each major version, not only the final file.
• Write in a tool with version history. Google Docs and Word both record timestamped edits, which show text growing over time.
• Keep your sources. Saved PDFs, highlighted passages and a reading log show where your ideas came from.
• Disclose any AI use. If you used AI for something the policy allows, say so in a note, and cite it the way your style guide requires. Our guide on how to cite ChatGPT shows the formats.
• Do not rely on a detector to clear you. A “0% AI” result from one tool does not prove anything either, because the same tools are unreliable.

What to do if a detector flags your work

1. Stay calm and ask to see the exact report and the school’s policy on how AI scores may be used.
2. Gather your drafts, version history, notes and sources before the meeting.
3. Explain how you wrote the piece and be ready to talk through your argument and sources out loud.
4. Point out the known limits of detectors, including the Liang and OpenAI figures above, and note that a score is not evidence of how the text was written.
5. Ask for the decision and any appeal route in writing.

FAQ