Claude AI Detector: Can GPTZero, Turnitin and Other Tools Detect Claude?

Summary: Leading AI detectors name Claude as a model they cover, and the best tools caught almost all unedited Claude text in a 2025 University of Chicago benchmark. Accuracy drops on edited or "humanized" text, and no score proves who wrote something.

Can AI detectors detect Claude? Yes: most major AI detectors now explicitly list Anthropic's Claude among the models they detect, and on clean, unedited Claude output the strongest tools miss only a few percent of passages. The catch is that results vary a lot by tool, by Claude model and by how much the text was edited afterward.

This guide covers which detectors name Claude, what independent research found on Claude text, whether Turnitin and GPTZero catch it, free options, and what to do if human writing gets flagged. Vendor figures are labeled as vendor claims. For background on the company behind the model, see what Anthropic is.

💡 Not sure which AI tool is actually right for your business?

Get the free guide, Which AI Tool Should You Actually Use? It is a straightforward breakdown of the leading AI tools to help you pick the right one for your needs.

Subscribe to AI Business Weekly for the guide, plus daily coverage of AI trends, acquisitions, and product launches.

Get the Free Guide →

AI detectors can detect Claude-written text with high accuracy when the text is long and unedited. In the largest independent benchmark that tested Claude, Pangram and GPTZero each missed under 3% of AI passages from Claude Opus 4 and Claude Sonnet 4 at a standard threshold, while Originality.ai missed far more. Edited and humanized text is much harder.

Anthropic's current lineup, according to its models overview, is Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5. Independent research so far covers older Claude 4 models; the newest have been tested mainly by detector vendors. For a practical look at the assistant itself, see the guide on how to use Claude.

GPTZero, Pangram, Originality.ai, Copyleaks, Winston AI and Turnitin all name Claude on their own pages, and Codequiry names Claude for code. Only some publish Claude-specific accuracy numbers, and most of those numbers come from the vendors. The table below separates what each company states from what independent research measured.

Every vendor Claude text test above reports detection over 95%, but the latest ones included no human-written samples, so they say nothing about false positives.

AI detectors are highly accurate on unedited Claude text in independent testing, but the gap between tools is wide. A University of Chicago benchmark found Pangram and GPTZero missed under 3% of Claude passages at a 0.5 threshold, while Originality.ai missed 19% to 37%, depending on the Claude model.

The study, by Brian Jabarian and Alex Imas (NBER working paper No. 34223, September 2025), compared 1,992 human passages with AI versions from GPT-4.1, Claude Opus 4, Claude Sonnet 4 and Gemini 2.0 Flash. It is the largest independent test with Claude-specific results we could find.

At a 0.5 detection threshold, the false negative rates (AI text wrongly passed as human) on Claude text were:

An open-source RoBERTa detector missed even less Claude text, but it flagged about 95% of human passages as AI, so that result is meaningless. When the authors capped false positives at a strict 0.5%, Pangram was the only tool that still caught nearly all Claude text from both models.

The Chicago test used 2025 versions of each detector, and vendors update their models often. Originality.ai's own October 2026 study reported detecting 325 of 327 Claude Opus 5.5 responses, or 99.4%, at its "AI Allowance 15%" setting, though the page states it did not include human-written controls. That is a vendor claim, and it has not been independently confirmed.

For how the two biggest assistants compare as writers, see ChatGPT vs Claude.

Turnitin says it detects Claude. Its English AI writing detection FAQ lists seven Claude models: Claude 3 Haiku, Claude Sonnet 3.5, Claude Sonnet 3.7, Claude Haiku 4.5, Claude Sonnet 4.5, Claude Sonnet 4.6 and Claude Opus 4.6. Turnitin publishes no Claude-specific accuracy figure, and no Claude 5-series model appeared on the list when checked in November 2026.

A missing model does not mean Turnitin cannot flag its output, but Turnitin has not publicly confirmed coverage for Claude Opus 5.5, Sonnet 5.5 or Fable 5.1.

The same Turnitin FAQ sets out several limits that apply to Claude text like any other:

Students cannot check their own work with Turnitin's AI detector, because the indicator is visible only to instructors and administrators at licensed institutions.

GPTZero can detect Claude, and its homepage states that it detects AI content from Claude. Its own October 2025 test on 4,500 Claude Sonnet 4.5 texts reported over 97% recall at a 1% false positive rate, and the independent Chicago benchmark found it missed under 3% of unedited Claude Opus 4 and Sonnet 4 text.

GPTZero ran a second, smaller test in July 2026 on 30 samples from Claude Opus 4.8, Claude Sonnet 4.6 and Claude Haiku 4.5. It flagged all 30 as AI, including samples written to imitate a human author's style. The post describes the sample as small and included no human-written controls.

GPTZero's weak spot is humanized text: in the Chicago humanizer test, it missed 44% to 77% of processed Claude passages, depending on genre.

Yes, several detectors that name Claude offer free checks. GPTZero scans up to 10,000 characters without an account, Copyleaks scans up to 25,000 characters free, Originality.ai gives 10,000 free words once, and Pangram's homepage cites 20 free checks a day after sign-up. Winston AI offers a 14-day trial with 2,000 credits.

Free tiers suit spot checks, not bulk scanning. Pangram's pricing page lists only a 7-day trial, so its free allowance may change. Teachers comparing paid options can use the AI Business Weekly guide to the best AI detector for teachers, which covers LMS integrations and student privacy.

AI detectors catch edited or humanized Claude text far less reliably than raw output. In the 2025 Chicago benchmark, Claude text run through a humanizer tool slipped past GPTZero 44% to 77% of the time, while Pangram missed 0% to 4% and Originality.ai missed about 3% to 10%.

Heavy human rewriting is harder still, because the result genuinely mixes human and machine work. Turnitin's FAQ notes that short documents of a few hundred words tend to be scored "all or nothing."

Some code-specific detectors claim to spot Claude-written code, but at lower rates than for text. Codequiry lists an 84% detection rate for code from Claude Opus 4.8, Sonnet 4.6 and Haiku, below its 90% rate for ChatGPT, and says 1.3% of human code is falsely flagged. Both figures are vendor claims.

Turnitin's FAQ says its text model does not reliably detect AI-generated code. The guide to the AI code detector category covers the code-specific tools in more depth.

Claude detector results can be wrong because every AI detector estimates probability from writing patterns rather than proving authorship. Plain, formal or formulaic human writing can look machine-like.

The clearest warning comes from Stanford. Researchers ran 91 TOEFL essays by non-native English speakers through seven detectors and found they misclassified 61.22% of them as AI-generated on average. Turnitin's FAQ adds that repetitive text with little structural variation can trigger false positives.

A high AI score means "likely AI," not "definitely Claude." The full research record on error rates is in the AI Business Weekly guide on whether AI detectors are accurate, and the wider debate is covered in AI content detection.

If your own writing is flagged as Claude-generated, respond with evidence of your process. Vendors themselves say a score should not decide a case alone, so drafts and version history usually carry more weight than any percentage.

Teachers and editors should compare flagged work with earlier samples and talk with the writer before drawing conclusions. For policy guidance, see AI for education.

Yes. GPTZero, Pangram, Originality.ai, Copyleaks, Winston AI and Turnitin all name Claude as a model they detect. In a 2025 University of Chicago benchmark, Pangram and GPTZero missed under 3% of unedited Claude Opus 4 and Sonnet 4 passages, while Originality.ai missed 19% to 37%. Edited or humanized text is harder to catch.

Turnitin says it does. Its English AI detection FAQ lists seven Claude models, from Claude 3 Haiku to Claude Opus 4.6 and Claude Sonnet 4.6, but no Claude 5-series models as of November 2026. Turnitin publishes no Claude-specific accuracy figure and needs at least 300 words of prose to score a document.

Yes. GPTZero states it detects Claude, and its own test of 4,500 Claude Sonnet 4.5 texts reported over 97% recall at a 1% false positive rate. Independently, a 2025 Chicago benchmark found it missed under 3% of unedited Claude text, but 44% to 77% of humanized Claude passages.

Yes. GPTZero checks up to 10,000 characters free without an account, Copyleaks scans up to 25,000 characters free, and Originality.ai gives 10,000 free words once. Pangram's homepage cites 20 free daily checks after sign-up, and Winston AI offers a 14-day trial. Free tiers suit spot checks rather than bulk scanning.

Yes. AI detectors estimate probability from writing patterns and can flag human work. A Stanford study found seven detectors misclassified 61.22% of essays by non-native English speakers as AI on average. Most detectors also cannot reliably name the model behind a passage, so a high score means likely AI, not proof of Claude.

Every major detector now names Claude, and the best performers in the 2025 University of Chicago benchmark missed under 3% of unedited Claude passages, though humanized text exposed big gaps between tools. Turnitin's list stops at Claude Opus 4.6 and Sonnet 4.6, and Claude 5.5 results come only from vendor tests. Treat any "Claude detected" result as a reason to look closer, not as proof of authorship.

Written by the AI Business Weekly Editorial Team. Reviewed, edited and published by Sameer Khan. This article was AI-assisted, then reviewed by Sameer Khan before publishing.