AI Content Detectors: Do They Work?

AI content detectors: do they work? How they work, why they get it wrong and what to do cover graphic

AI content detectors do not work reliably enough to prove that a person used AI. They give a probability based on patterns in the text, and the vendors themselves say they can be wrong. OpenAI shut down its own AI text classifier in July 2023 because of its low rate of accuracy, and a Stanford study found that detectors often flagged writing by non-native English speakers as AI-generated.

This guide explains how detectors score text, why they make mistakes, what four well-known detector makers say about their own limits, and what to do if you are flagged or if you are the person reading a score. It is organised by question, not ranked, and it does not recommend any detector as the best. It also does not cover tools that claim to bypass detection, because the sensible response to a flag is evidence of your process, not disguise.

How we researched this: we did not run hands-on tests for this article. It is based on the vendor’s own documentation and public sources, checked on 17 November 2026. Features and prices change, so confirm details on the vendor’s site before you buy.

Key takeaways

  • A detector score is a probability, not proof. GPTZero’s own explainer says a high AI score indicates shared patterns, not definitive proof.
  • False positives are real. Turnitin says it cannot mitigate the risk of false positives completely, and it marks low scores with an asterisk because they are less reliable.
  • Non-native English writers may be flagged more often. A Stanford study reported by The Markup found seven detectors labelled such writing as AI-generated about 61 percent of the time.
  • Short texts, lightly edited AI text and mixed human and AI writing are all harder to judge, according to detector vendors.
  • If you are flagged, gather your drafts, notes and version history. If you are the reader, use the score to start a conversation, not to decide a case.

Quick answer: what the evidence says

Question Short answer Evidence
Can a detector prove AI use? No GPTZero, Turnitin and Originality.ai each describe scores as signals or probabilities, not proof
Do detectors make false accusations? Yes, sometimes Turnitin states a document false positive rate under 1 percent for documents with over 20 percent AI writing and a sentence level rate of about 4 percent, in a May 2023 post
Is any group flagged more often? Possibly non-native English writers The Stanford study; Turnitin says its own unpublished research finds no such bias
Did any major AI lab release a reliable detector? OpenAI withdrew one Withdrawn July 2023 for low accuracy
Is there a more reliable approach? Watermark checks work only for content made with the matching tool Google says SynthID Detector identifies content made with Google AI

How we chose what to include

We chose the detectors people ask about most in schools and publishing: Turnitin, GPTZero, Originality.ai and Copyleaks. We added OpenAI’s withdrawn classifier for context and Google’s SynthID Detector as a different kind of check. We read vendor blog posts and help pages, plus news coverage of a Stanford study.

We did not run any text through any detector, so we cannot give accuracy numbers of our own. All accuracy figures below are the vendors’ own claims or reported studies. Several vendor pages could not be opened by our tools and are noted as unconfirmed. Our related guide how to spot AI-generated text and images covers reading clues by eye. This article focuses on detector software and what to do with its scores.

How AI detectors work

GPTZero’s explainer says detectors examine wording, rhythm and structure and compare the text with large datasets of known AI and human writing. Results are usually labelled AI, human or mixed, with a probability score.

Perplexity

GPTZero describes perplexity as a surprise meter. Unexpected wording raises perplexity and is linked to human writing, while predictable text is linked to AI. Predictability comes from how large language models choose likely next words.

Burstiness

Burstiness measures how perplexity varies across a document. GPTZero says human writing mixes short and long sentences, while AI text tends to be more uniform.

Classifiers and embeddings

GPTZero says it began with perplexity and burstiness but now uses a multilayered deep learning system trained on human and AI text from recent models. The article also describes classifiers that sort text by learned patterns and embeddings that represent words as vectors. Other vendors use their own models, and we could not confirm the details of each.

OpenAI’s classifier: a cautionary tale

OpenAI launched an AI Text Classifier on 31 January 2023 and made it unavailable from 20 July 2023, saying it was discontinued due to its low rate of accuracy, per Search Engine Land. At launch OpenAI had warned that it was impossible to reliably detect all AI-written text, and that it had not thoroughly assessed content co-written with humans. OpenAI said it was researching more effective provenance techniques for text.

The lesson is not that detection is hopeless in every case. It is that even the company that makes the generator did not find a reliable text detector.

Turnitin

Turnitin is built into many school and university workflows. Its figures below come from a post by its chief product officer dated 23 May 2023, which may be out of date, so check Turnitin’s current documentation.

What it says

  • For documents with over 20 percent AI writing, a document false positive rate of less than 1 percent.
  • A sentence level false positive rate of about 4 percent, mostly next to real AI text, especially at transitions between human and AI writing.
  • When less than 20 percent AI writing is detected, false positives are more common, so the indicator shows an asterisk.
  • The minimum length for detection was raised from 150 to 300 words.
  • Turnitin says it cannot mitigate the risk of false positives completely and that these metrics may change.

How it says the score should be used

Turnitin says educators should use the AI score to start a dialogue with students and not as a standalone finding, and should view highlighted areas as a whole. It points educators to guides on handling false positives.

Bias concern

The Markup reported that Turnitin was not among the detectors in the Stanford study, and that Turnitin says its tool was trained on multilingual writing and its unpublished internal research finds no bias against non-native speakers. The Markup also described a Johns Hopkins instructor who saw Turnitin label over 90 percent of a student’s paper as AI-generated before the student showed drafts and notes. We could not verify these individual cases.

Pricing

Turnitin is sold to institutions, and we could not confirm any price.

GPTZero

GPTZero is a detector aimed at teachers, students and writers. Its own explainer lists limits that are useful for any detector.

Limits it states

  • Detectors can produce false positives and false negatives.
  • No tool can claim 100 percent accuracy.
  • Mixed human and AI documents and unconventional or highly creative writing are harder to assess.
  • Short passages offer fewer patterns, which reduces reliability.
  • Most detectors are trained mainly on English, which can reduce effectiveness in other languages.
  • Many detectors struggle to tell light grammar help from fully AI-generated text.
  • Using a detector alone can lead to unfair outcomes, so combine it with other checks and human judgment.

Pricing and accuracy

We did not read an official GPTZero pricing page, so we do not quote prices. Its accuracy claims are self-reported and we did not verify them.

Originality.ai

Originality.ai is popular with publishers and content teams.

What it says

Its blog page on false positives claims 99 percent plus accuracy for its latest models, says false positives are exceptionally low but still possible, and gives no specific false positive figure on that page. It says AI scores should be treated as a signal, not proof of misconduct, and recommends clear AI policies at the start of a project.

Advice for accused writers

It suggests gathering research notes, version history and a writer replay, reviewing the writing agreement and AI policy, checking whether a detection score or tool was agreed in advance, and showing earlier work. For students it also suggests reviewing the school’s academic integrity policy and escalating to administration if unresolved. This advice comes from a vendor, but it matches the sensible approach for anyone.

Pricing

We could not confirm pricing from an official page.

Copyleaks

Copyleaks offers an AI detector alongside plagiarism checking. Search results showed it publishes a limitations page and a blog on dos and don’ts of using AI detectors, but the help page we opened did not load, so we cannot report what it says. We also could not confirm its accuracy claims or pricing. If you use it, read its limitations page first.

Watermark checkers such as SynthID Detector

A different approach is to check for a hidden watermark rather than guess from style. Google announced SynthID Detector on 20 May 2025 as a portal that checks media for a SynthID watermark and can highlight the parts most likely to be watermarked. It covers images, audio, video and text created with Google’s AI tools, and Google describes it as identifying content made with Google AI. Early access was via a waitlist at announcement, and we did not check its current availability.

The limit is built in. A watermark check can only find marks added by the matching tool, so it does not detect text from other systems and it says nothing about unmarked content. We cover Google’s music watermark in our music generators guide.

Why detectors get it wrong

Plain, formal or simple writing looks predictable

If a detector links predictable wording with AI, writers who use simple vocabulary or formal patterns can be flagged. The Stanford study, as reported by The Markup, found that seven detectors labelled non-native English writing as AI-generated about 61 percent of the time and that all seven gave the same wrong label on about 20 percent of papers. The study’s authors wrote that the design of many detectors inherently discriminates against non-native authors. Originality.ai published a response calling the study flawed, which we did not evaluate.

Mixed writing

Many real documents mix human and AI text, or use AI only for grammar. GPTZero says detectors struggle with this. Our guide to Grammarly covers one tool that edits rather than writes.

Short text

Short passages have fewer patterns. Turnitin raised its minimum from 150 to 300 words for this reason.

Other languages

GPTZero says most detectors are trained mainly on English, which can reduce effectiveness in other languages.

Newer models

GPTZero says traditional clues become less obvious as AI writing becomes more human-like.

Comparison tables

What each source says about its own limits

Source Stated position
OpenAI (classifier) Withdrawn July 2023 for low accuracy
Turnitin Cannot mitigate false positives completely; use score to start a dialogue
GPTZero No tool can claim 100 percent accuracy; combine with human judgment
Originality.ai Treat scores as a signal, not proof of misconduct
Copyleaks Not confirmed from pages we could open
Google SynthID Detector Identifies content made with Google AI

Documented thresholds and claims

Vendor Claim or threshold Date or note
Turnitin Document false positive rate under 1 percent above 20 percent AI writing; sentence rate about 4 percent Post dated 23 May 2023
Turnitin Asterisk below 20 percent; 300 word minimum Same post
Originality.ai 99 percent plus accuracy claimed for latest models Self-reported, no false positive figure on page
GPTZero Multilayered deep learning system Self-reported
Stanford study About 61 percent of non-native English samples flagged by seven detectors Reported by The Markup, August 2023; Turnitin not tested

Strengths and weak spots

Approach Strength Weak spot
Statistical text detectors Work on any pasted text Probabilities, prone to false positives and negatives
Watermark checkers Can give a clearer signal for marked content Only for matching tools, and not for unmarked text
Process evidence such as drafts Shows how work was made Needs to be kept before a dispute

What to do: flagged writers and readers

If you are flagged

  • Stay calm and read your school’s or client’s AI policy.
  • Gather drafts, notes, outlines, research and version history. Google Docs and Word keep history.
  • Ask which tool and which score was used, and whether the score was agreed in advance.
  • Offer to explain your work or write a short passage on the same topic.
  • Cite vendor guidance that scores are signals, such as Turnitin’s advice to use the score to start a dialogue.
  • Escalate through your institution’s process if needed. This is general information, not legal advice.

If you are the reader

  • Treat a score as a prompt for a conversation, not a verdict.
  • Be careful with low scores and short texts, which vendors say are less reliable.
  • Take extra care with non-native English writers.
  • Compare with the person’s earlier work and ask them to explain it.
  • Write a clear AI policy in advance, as Originality.ai also recommends. See AI in education: this year’s shifts.

If you use AI and want to be open about it

Keep a record of how you used it. Our guide on documenting AI-assisted decisions shows one way, and auditing AI-written content for E-E-A-T signals covers publishing quality. Students can start with our students shortlist guide.

If you run a team, school or publication

Decide your rules before a dispute starts. Write down whether AI help is allowed, for which tasks, and how it must be disclosed. Say whether detector scores will be used at all, and if so, what happens next: a conversation, a request for drafts, or a review by a second person. Originality.ai recommends clearly defined AI policies at the start of a project or assignment, and Turnitin recommends using its score to open a dialogue. Both fit with a simple three step process: notice, ask, review.

  1. Notice: a score or a hunch prompts a closer look, nothing more.
  2. Ask: invite the writer to share drafts, sources and an explanation of how they worked, in a respectful tone.
  3. Review: a person who is not the original reader weighs the evidence, and the writer has a way to appeal.

Put the policy in writing and apply it the same way to everyone. Treat non-native English writers with extra care, since the research we cite suggests they are at higher risk of being wrongly flagged. Keep records of decisions so that you can show how you reached them, and revisit the policy at least once a year, because detector claims and AI tools both change quickly.

Mistakes and limits to watch for

Treating a percentage as a fact

A score of 80 percent AI does not mean 80 percent of the text was AI-written, and it does not prove anything. Read the vendor’s own explanation.

Using one detector alone

GPTZero warns that relying on a detector alone can lead to unfair outcomes. Different tools can disagree.

Ignoring language and length

Short or non-English text is harder, according to vendors.

Ignoring your own process

Writers who keep drafts have the best evidence. Writers who paste a final text from nowhere have none.

Publishing unchecked AI text

Whether or not a detector flags it, AI text can contain errors. Check output using our guide to checking AI output for mistakes.

Believing marketing claims

Accuracy claims such as 99 percent plus are self-reported. Look for independent, recent studies and for false positive rates, not just accuracy.

Frequently asked questions

Do AI detectors work?

They find patterns and give probabilities, but vendors acknowledge errors. OpenAI withdrew its own classifier for low accuracy. Treat results as a signal.

Can an AI detector be wrong about human writing?

Yes. Turnitin states a sentence level false positive rate of about 4 percent and says it cannot mitigate false positives completely.

Are AI detectors biased against non-native English speakers?

A Stanford study reported that seven detectors labelled non-native English writing as AI-generated about 61 percent of the time. Turnitin was not tested, and says its own unpublished research finds no such bias.

What is the most accurate AI detector?

We cannot say. We did not test any, and vendors report their own accuracy figures. Look for independent evidence and false positive rates.

What should I do if my work is flagged as AI?

Gather drafts, notes and version history, read the policy, ask which score and tool were used, and request a conversation. This is general information, not legal advice.

How many words does a detector need?

Turnitin raised its minimum to 300 words in 2023. Other vendors differ, and we did not confirm their limits.

Can detectors spot text from any AI?

Statistical detectors try to, with mixed reliability. Watermark checkers such as SynthID Detector only find content made with the matching tool.

Should schools ban AI detectors?

That is a policy decision we do not make here. Vendors themselves advise not using scores alone, so any use should include human review and an appeal route.

Conclusion

AI content detectors can be a useful nudge, but they are not evidence on their own. Their makers say as much, the best known attempt from OpenAI was withdrawn, and research suggests non-native English writers may be flagged unfairly.

If you write, keep your process. If you read, ask questions before you decide. For help choosing any AI tool, read how to choose an AI tool.

Sources


Keep reading