Are AI Detectors Accurate? What AI Detection Research Shows

AI detectors are partly accurate. The best tools today can be very good at spotting unedited AI text in long passages, and some independent tests have found them rarely flagging human writing. But accuracy varies a lot between tools, drops on short, edited or style-imitated text, and has shown bias against non-native English writers. AI detection is a useful signal, not proof.

That’s the short version. The longer version is more interesting, because the evidence has shifted over the past few years. Here’s what the published research and vendor statements actually say, with links so you can check each one.

What does “accurate” mean for AI detection?

Before looking at studies, it helps to separate the two ways a detector can be wrong:

• False positive: human writing flagged as AI. This is the error that can get a student or writer wrongly accused.
• False negative: AI writing that passes as human. This is the error that lets AI text slip through.

A single “accuracy” figure blends both and hides which mistake a tool tends to make. It also depends on the test set: the models used, text length, writing genre and whether anyone edited the text. That’s why one detector can look excellent in one study and mediocre in another.

Why did OpenAI shut down its own AI detector?

OpenAI launched its AI Text Classifier on January 31, 2023. At launch, the company reported that on its challenge set the tool correctly flagged 26% of AI-written text as “likely AI-written,” and incorrectly labeled human text as AI 9% of the time. It also said results were unreliable on text under 1,000 characters.

Less than six months later, OpenAI added a note to the announcement: “As of July 20, 2023, the AI Classifier is no longer available due to its low rate of accuracy” (Inquirer Technology; see also The Decoder).

The company that built ChatGPT couldn’t build a detector it was willing to keep online. That became one of the most cited points in the debate about AI writing detection.

Are AI detectors biased against non-native English writers?

The best-known study here is by Weixin Liang, James Zou and colleagues at Stanford, published in Patterns in July 2023 (full text). They ran seven widely used detectors on 91 TOEFL essays written by non-native English speakers and 88 essays by US eighth-graders.

The findings, as reported in the paper and by Stanford HAI:

• The detectors were near-perfect on the eighth-graders’ essays.
• On the TOEFL essays, the average false positive rate was more than 61%.
• About 19% of the TOEFL essays were flagged as AI by all seven detectors, and 97% were flagged by at least one.
• When the researchers used ChatGPT to make the TOEFL essays’ vocabulary more sophisticated, the false positive rate dropped sharply.

The explanation ties back to perplexity, one of the main detection signals. Writers with a smaller vocabulary and simpler grammar produce more predictable text, which looks more “AI-like” to a detector. (We explain perplexity in how AI detectors work.)

This study tested detectors available in early 2023. Newer tools may perform differently, but the underlying risk is worth keeping in mind whenever you judge writing by non-native speakers.

What does Turnitin say about its false positive rate?

Turnitin added AI writing detection to its plagiarism platform in 2023 and has published its own figures. In a June 2023 post, Chief Product Officer Annie Chechitelli said:

• The document-level false positive rate is under 1% for documents with 20% or more AI writing.
• The sentence-level false positive rate is around 4%.
• False positives are more common in mixed human and AI documents, especially at the transitions between the two.
• Educators should use highlighted sentences to start a conversation, not to draw a conclusion.

These are the vendor’s claims, measured on its own test data. They’re also useful context: even a low rate adds up at scale, which is exactly the point Vanderbilt made.

Why did some universities turn off AI detection?

In August 2023, Vanderbilt University announced it was disabling Turnitin’s AI detector. Its reasons included:

• Turnitin enabled the feature with less than 24 hours’ notice and no option to opt out at first.
• There was little transparency about how the detector worked.
• Vanderbilt submitted 75,000 papers to Turnitin in 2022. Even at Turnitin’s claimed 1% false positive rate, that could mean around 750 papers wrongly flagged.
• Research suggested detectors were more likely to flag non-native English writers.

Vanderbilt concluded it did not believe AI detection software was an effective tool to use. Other institutions made similar calls, while many still use detection alongside human review.

What did independent studies find?

Several independent evaluations have tested detectors side by side. Their conclusions changed as the tools matured.

Study	What it tested	Main finding
Weber-Wulff et al., 2023 (Int. Journal for Educational Integrity)	14 tools, including Turnitin	Tools were “neither accurate nor reliable,” leaned toward calling AI text human, and got worse with obfuscation
RAID benchmark, ACL 2024	12 detectors on 6M+ generations across 11 models	Detectors were easily fooled by adversarial attacks, sampling changes and unseen models
Jabarian & Imas, 2025 (Chicago Booth)	GPTZero, Originality.ai, Pangram and an open-source RoBERTa model on about 2,000 human passages	All three commercial tools kept false positives under 1%; the open-source model was close to random; accuracy dropped on passages under 50 words
Epoch AI, July 2026	Pangram, GPTZero and Originality.ai on 495 pre-2022 human passages and text from current frontier models	Basic AI text was almost always caught; style-imitated AI text was missed 10% to 18% of the time; one tool flagged 3.8% of human passages

A few things stand out from the newer work:

• The gap between tools is large. In the 2025 study, commercial detectors did far better than the open-source model.
• Style imitation is a weak spot. In Epoch AI’s test, when a model was given samples of an author’s writing and asked to imitate it, detectors missed far more AI text, and scientific writing was the hardest genre.
• Short text is still hard. Accuracy fell on short passages in the 2025 study.
• Results date quickly. Epoch AI notes its results reflect detector versions from June 2026, and vendors update their tools often.

So, are AI detectors accurate enough to trust?

It depends on what you’re trusting them to do.

Where AI detection is genuinely useful:

• Flagging long, unedited AI text for a closer look.
• Spotting which sections of a document deserve attention, especially with sentence-level highlights.
• Screening content at scale, where a human reviews anything flagged.
• Checking your own drafts before you submit or publish them.

Where you should be careful:

• Short passages, like a single paragraph or a discussion post.
• Writing by non-native English speakers or anyone with a plain, formal style.
• Text that mixes human and AI writing, or has been heavily edited.
• Any decision with real consequences for a person, based on a score alone.

The fair reading of the research is that detectors are useful signals, not proof. A detector can tell you text looks like AI. It can’t tell you how it was made, who made it, or whether AI use was allowed.

How to use AI detection responsibly

If you rely on AI detection for school, hiring or publishing, a few habits reduce the risk of getting it wrong:

1. Use more than one detector, ideally built by different companies. Agreement means more than a single high score.
2. Read the sentence highlights, not just the headline number.
3. Check context: drafts, version history and the writer’s own explanation.
4. Set a high bar for action. Look for several kinds of evidence before making an accusation.
5. Keep up with your institution’s AI policy, which may allow some AI use.

Our step-by-step guide on how to check if text is AI generated covers this process in detail. For tool picks, see our roundups of the best free AI detectors and the best AI detectors for essays.

Our own free AI detector shows an overall score, per-sentence highlighting and the signals behind the result, so you can see why it reached a verdict. Like every detector, it can be wrong, and we say so on the results screen. If you’re revising your own writing to read more naturally, our AI humanizer can help with flow and tone, though no tool can guarantee a specific detector result. Plans and limits are on our pricing page.