Claude AI Detector

We mark the sentences that read like AI and show how often we catch each Claude model.

0 / 2,000 words

How to check text for Claude

  1. Step 1

    Paste or upload

    Paste the text, or sign in to upload a .docx, .pdf or .txt file. You need at least 150 words (200 in Spanish, Portuguese and French) for a verdict.
  2. Step 2

    Read the score and colors

    You get an overall AI score. Red sentences read like AI, yellow ones are unclear, and grey text wasn't scored.
  3. Step 3

    Edit and check again

    Click the text to change it, then check again to see whether the score moved.

Claude's answers in a chat are often short. Paste the whole piece rather than a single paragraph: under 150 words (200 in Spanish, Portuguese and French) the detector won't give an overall verdict. If you only want to know about part of a document, the sentence colors show where the AI-like passages are.

Can it tell Claude apart from ChatGPT?

No, and we don't try. The detector answers one question: does this text read like it was written by AI? It gives the same kind of answer whether the text came from Claude, ChatGPT, Gemini or another model. It's the same check our main AI content detector runs.

Identifying the source model is a separate task, and we haven't tested whether we could do it reliably. A wrong guess about the source would be worse than no guess, so the result doesn't name a model.

If you need to know whether a particular tool was used, the writing itself won't settle it. Version history, drafts and a conversation about how the piece was written show things a detector can't, and they work the same whichever chatbot was involved.

How well we detect each Claude model

We generated 100 texts with each of four current Claude models, Fable 5.1 and the 5.5 versions of Opus, Sonnet and Haiku: a third each academic writing, blog posts and emails, none of them used when we tuned the detector. Anthropic has other models we haven't tested, such as Fable 5.2, so they aren't in the table. Half were 150 to 449 words long and half longer. The older Sonnet 4.5, which doesn't carry Anthropic's watermark, is shown separately.

Claude texts marked as likely AI
ModelTextsLikely AI95% intervalMixedLikely human
Claude Fable 5.110091 (91.0%)83.8% to 95.2%90
Claude Opus 5.510092 (92.0%)85% to 95.9%44
Claude Sonnet 5.510090 (90.0%)82.6% to 94.5%91
Claude Haiku 5.510097 (97.0%)91.5% to 99%12
Claude Sonnet 4.5 (older, no watermark)210209 (99.5%)97.4% to 99.9%10
Detection engine version w0p5-38c6d591, tested November 11, 2026. The four current models together: 370 of 400 (95% interval 89.5% to 94.7%).
By length
Model150 to 449 words450 words and up
Claude Fable 5.147 of 50 (94.0%)44 of 50 (88.0%)
Claude Opus 5.545 of 50 (90.0%)47 of 50 (94.0%)
Claude Sonnet 5.544 of 50 (88.0%)46 of 50 (92.0%)
Claude Haiku 5.547 of 50 (94.0%)50 of 50 (100%)
Claude Sonnet 4.5 (older, no watermark)109 of 110 (99.1%)100 of 100 (100%)

Most of the texts we missed came back mixed rather than human: across the four current models, 7 of 400 were called likely human. We also checked 10 shorter texts per model, under 150 words; none got a verdict, which is how the detector is meant to behave with too little text.

The models don't all behave the same way. Haiku 5.5 was the easiest to catch, especially in longer texts, where all 50 were marked likely AI. Fable 5.1 and Sonnet 5.5 had the most mixed results, 9 each out of 100, and Opus 5.5 had the most texts called likely human, 4 of 100. Fable 5.1 was caught less often in longer texts: 44 of 50 at 450 words or more, against 47 of 50 below that.

A mixed result means there's no confident verdict on the text as a whole. It can mean some sentences read like AI and others don't; the sentence colors show which. If you're checking someone else's work, treat mixed as a reason to look at the red sentences, not as a finding either way.

The interval is the range the true detection rate probably falls in. With 100 texts per model it's fairly wide, so small differences between models in this table aren't meaningful.

Claude, ChatGPT and Gemini side by side

On our main English test set, used to decide whether the detector was ready to launch, these are the results for unedited text from each model family. The same thresholds apply to every model.

AI texts marked as likely AI, by model
ModelTextsLikely AI
OpenAI GPT-6.14947 (95.9%)
Anthropic Claude Sonnet 5.55048 (96.0%)
Google Gemini 3.8 Flash5050 (100%)
DeepSeek V4 Pro4643 (93.5%)
Qwen 3.8 Max4847 (97.9%)
Mistral Small 3.23535 (100%)
English, November 2026.

Claude Sonnet 5.5 shows up in both tables. The two numbers come from different sets of texts, 48 (96.0%) on the main set and 90.0% here, which is the kind of variation you'd expect from samples this size.

For how other detectors did in published studies, see our roundup of the best AI detectors.

What Claude's writing tends to look like

We counted phrases in 150 texts written by Claude Sonnet 5.5 and compared them with human writing of the same kinds: academic writing, blog posts and emails. These turned up far more often in the Claude texts. The last row is the opposite case: “furthermore”, often called an AI giveaway, was far more common in the human writing.

How often phrases appear per 100,000 words
PhraseExample from our Claude samplesClaude Sonnet 5.5, per 100,000 wordsHuman writing, per 100,000 words
“This essay argues that”“This essay argues that the economics of adaptation in low-income nations is best understood as a problem of financing…”18.90
“critics”“Critics, however, note that predictive validity is modest and uneven.”24.60.3
“honestly”“And honestly, there’s a gentle lesson in being made to wait.”23.30.8
“Warm regards”The sign-off in 18 of our 48 Claude emails19.20
“defensible”“Perhaps this is the most defensible assessment: refugee law has succeeded as a language of legitimacy…”11.50
“a modest”“Even a modest sunny patch of about four by eight feet is enough for a productive first garden.”11.30.2
“rests on”“The case for adaptation rests on favorable benefit-cost ratios.”9.80.1
“I’ve been thinking about”“I’ve been thinking about this since June.”9.60.5
“I hope this message finds you well”The first line after the greeting in 8 of our 48 Claude emails8.50
“furthermore” (for comparison)Rare in the Claude texts1.615.3
Our English test collection, November 2026. Rates are averaged across academic writing, blog posts and emails.

Claude likes to announce its thesis (“This essay argues that”), to bring in critics and answer them, and to soften claims with “a modest” or call a position “defensible”. In emails it reaches for the same openings and sign-offs. It also didn't use a single em dash in our samples, against about one per 1,000 words in the human writing, so the idea that dashes give Claude away doesn't hold for these texts.

The detector combines sentence scores with writing measurements, including stock phrases. These phrase counts alone don't determine the result, and a text with none of them can still read like AI. For the wider picture, see how AI detectors work.

Does Claude watermark its text?

Yes, for current models. Anthropic's help center says Claude Fable 5.1 and the 5.5 versions of Opus, Sonnet and Haiku add a statistical watermark to the text they write, while older models such as Opus 4.6, Sonnet 4.5 and Haiku 4.5 don't. The watermark is a pattern in which words get chosen, not hidden characters, so there's nothing visible to delete. Anthropic notes that heavy editing can make it undetectable.

Checking for the watermark needs Anthropic's own tool, which is in private preview for eligible organizations such as regulators, researchers, media and educational organizations. Our detector doesn't read it; it judges the writing. Anthropic also notes that a detected mark shows Claude processed the text but isn't conclusive, and that a missing mark doesn't mean a person wrote it. (Help center checked November 11, 2026.)

Claude-edited text and false positives

Two cases come up a lot, and we haven't measured either separately for Claude. Human writing that Claude has polished keeps the writer's structure but picks up Claude's phrasing, so it can come back mixed. A Claude draft that a person has rewritten heavily can move toward mixed or human as more of the wording becomes the writer's own.

False positives happen too. These are the detector's results on human writing in our English test set:

Human writing in our English test set
TextsHow manyResult95% bound
Human-written texts1,0961 flagged as likely AI (0.09%)At most 0.52%
Essays by non-native English writers6000 flagged as likely AIAt most 0.64%
English, November 2026.

If your own writing is flagged, our guide to AI detector false positives walks through what to do.

From a Claude draft to your own wording

If you used Claude for a first draft and want it to sound like you, start with the red sentences. Rewrite them yourself, or send just those to our AI humanizer, which rewrites only what you select and shows every change next to the original. The ChatGPT humanizer works the same way on Claude text.

A few things help a Claude draft sound like you, whichever way you do it:

  • Cut the announcement. Start with your point instead of “This essay argues that”.
  • Replace general claims with something you know: a number from your own research, an example from class, a place you've been.
  • Keep your own sign-off and greeting in emails rather than the stock ones.
  • Read it aloud. Anything you wouldn't say, change.

Before you do, check the rules where you're submitting. Many schools and publishers allow some AI help but expect you to say how you used it.

Frequently asked questions

Is this Claude detector free?

Yes. Without an account you can check up to 2,000 words at a time, 10 times a day. A free account raises that to 10,000 words and 30 checks.

Can it tell if text came from Claude or ChatGPT?

No. It tells you whether text reads like AI, not which company's model wrote it. Telling models apart reliably on short or edited text is a harder problem we don't claim to solve.

Does it detect Claude Fable 5.1 and Opus 5.5?

Yes. On November 11, 2026, it marked 91 of 100 Fable 5.1 texts and 92 of 100 Opus 5.5 texts as likely AI. Most of the rest came back mixed.

Does Claude add a watermark to its text?

Anthropic says recent models, including Fable 5.1 and the 5.5 versions of Opus, Sonnet and Haiku, do. Detection is limited to approved organizations, and our detector doesn't read watermarks.

Can Turnitin detect Claude?

Turnitin lists Claude models among those it covers, if your school has its AI detection switched on. Our Turnitin article explains how it works.

Can it detect Claude text that I edited?

Sometimes. It depends on how much of the wording is now yours; we haven't published separate results for edited Claude text.

Why was my own writing flagged as AI?

Detectors make mistakes, especially on short, formulaic or very polished writing. Our guide to AI detector false positives covers what to do.

Do you store my text?

Without an account, text you only check isn't saved in our database. With an account it's saved as a draft in your history, which only you can see, until it expires or you delete it. We don't use it to train models.

Claude, ChatGPT, Gemini and Turnitin are trademarks of their owners. AI Checker Humanizer isn't affiliated with any of them.

Try it on your own text

Check for AI