Do AI detectors actually work?

AI detectors work sometimes, on some kinds of text, and they fail in ways that can hurt real people. Independent studies have found high error rates, easy workarounds and bias against non-native English writers, and OpenAI withdrew its own detector in 2023 for low accuracy. The best commercial tools now do much better on unedited AI text.

Even so, no AI detector score is proof. Use one as a reason to ask questions, never as a verdict. Here is what the research shows, and what students, teachers and publishers should do instead.

01How AI detectors work

Many detectors lean on a measure called perplexity: how surprised a language model is by each next word. AI text tends to be predictable. The trouble is that some human writing is predictable too, especially plain, formulaic or carefully learned English. That overlap is where most false accusations come from.

02What the research says about AI detector accuracy

The evidence has shifted over three years, so dates matter.

1. January 2023OpenAI launches a classifier that catches 26% of AI text and wrongly flags 9% of human text
2. July 2023OpenAI withdraws it, citing its low rate of accuracy
3. August 2023Vanderbilt University switches off Turnitin’s AI detector
4. December 2023A test of 14 tools finds none above 80% accuracy
5. September 2024Simple edits cut detectors’ average accuracy from 39.5% to 22%
6. September 2025Three commercial detectors wrongly flag under 1% of human passages
7. February 2026Turnitin and Originality score 61% and 69% accuracy on a mixed test set

Two lessons stand out. First, the tools improved: do not judge a 2026 product by a 2023 test. Second, the hardest case has not gone away. Mixed writing, where a person drafts, an AI polishes and the person edits again, is increasingly common, and that is exactly where the 2026 study found both tools weakest.

03The false positive problem

A false positive is human writing flagged as AI. In a school or workplace, it means a false accusation.

• Small rates add up. Turnitin says its detector has a false positive rate under 1%. That is the vendor’s own claim, from 2023. Vanderbilt did the math when it turned the tool off: it had sent 75,000 papers through Turnitin in 2022, so around 750 could have been wrongly flagged.

• Non-native writers are hit hardest. In the Liang study, all seven detectors agreed that 19.8% of the non-native essays were AI-written, and at least one flagged 97.8% of them. When the researchers used ChatGPT to enrich those essays’ vocabulary, the false positive rate fell from 61.3% to 11.6%. Polishing human writing with AI made it look more human.

• A probability cannot be disproved. There is no file to inspect, only a score. The 2026 study reached the same conclusion: detectors are unsuitable as the sole basis for misconduct decisions.

04Why AI detectors are easy to fool

Detectors are in an arms race with tools that rewrite text. In the 2023 study, human editing cut detection of AI text from 74% to 42%, and machine paraphrasing dropped it to 26%. In the 2024 study, adding spelling errors and making sentence lengths more varied worked especially well.

The newest tools push back. The 2025 University of Chicago working paper found that Pangram stayed accurate on passages of 50 words or fewer and on text run through “humanizer” tools, which rewrite AI text to dodge detection. The authors still warn that results will shift as detectors, models and humanizers keep competing. It is a working paper and has not been peer reviewed.

05People can be good detectors too

A 2025 study presented at ACL, a leading language-technology conference, asked people to judge 300 nonfiction articles. Annotators who often use ChatGPT for writing were remarkably good at it. A majority vote of five of them misclassified only 1 of the 300, beating most commercial and open-source detectors, even on paraphrased and “humanized” text.

They noticed typical “AI vocabulary”, but also subtler things such as formality, originality and clarity. Our takeaway: a teacher or editor who uses these tools, and knows the writer, is a strong check in their own right.

06Watermarking versus detection

Detection guesses after the fact. Watermarking plans ahead: the model hides a statistical signal in its word choices as it writes, and a matching checker looks for it later.

Google’s SynthID-Text is the best-documented example. Google published it in Nature in October 2024, after a live test on nearly 20 million Gemini responses showed no change in quality. The paper notes that edits such as AI paraphrasing weaken the watermark.

OpenAI built a text watermark too, and in August 2024 said it was still considering it while researching alternatives. It held up against local edits such as paraphrasing, but was less robust against translation or rewording by another model. OpenAI also worried it could stigmatize non-native English speakers who use AI as a writing tool.

Images and audio are further along. Signed content credentials plus watermarks such as SynthID, which OpenAI began adding to its images in 2026, can show which tool made a file. Our guide to content credentials explains how to check them.

07What to do instead: advice for students, teachers and publishers

Students. Keep your drafts, notes and version history, because they are the strongest evidence of your own work. Learn the AI rules for each course. If you are flagged, stay calm, ask what evidence exists beyond the score, and offer to talk through your work. Our guide to AI for students covers using AI without crossing the line.

Teachers. Say plainly what AI use each assignment allows. Build in process: outlines, drafts and short conversations about the work. Use a detector, if at all, to start a conversation, and remember the evidence of bias against non-native writers.

Publishers and editors. Judge accuracy and usefulness rather than origin. Google takes a similar line on AI-assisted content, as our guide to SEO in the age of AI explains. Ask contributors to disclose AI use, and check every fact, quote and reference, because invented citations are the real risk.

A short conversation is often the fairest check. This prompt turns a piece of writing into questions that test understanding, not authorship.

Here is a student's essay. Write five short questions I can ask the student in a two-minute conversation to check they understand what they wrote.
Ask about their reasoning, the sources they used and one choice they made, not about memorized facts.
Do not judge whether the essay was written by AI.

Essay:
[paste the essay with the student's name removed]

08FAQ

Next, see how students can use AI without cheating, or how content credentials show where an image came from.