How to tell if Claude wrote something

Sometimes: people who use AI writing tools often spot it remarkably well, while readers who seldom or never write with AI, tested on GPT-4o articles, guessed at roughly chance level (56.7% of AI articles caught, 51.7% of human ones flagged).

The clearest numbers come from a 2025 paper by Russell, Karpinska and Iyyer. Each batch held 60 English nonfiction articles under 1,000 words: 30 from eight US publications and 30 generated by a model from the same title and subtitle. Five readers who often lean on AI for writing tasks judged every article.

• Claude batch: the five readers' majority vote labeled all 30 Claude 3.5 Sonnet articles as AI and none of the 30 human ones.
• Whole study: across 300 articles from several models, that vote got only one wrong.
• Non-writers with AI, GPT-4o batch only: 56.7% of AI articles caught, 51.7% of human articles flagged. They were not tested on Claude.

Read those results with their limits in mind. It is one study with one prompt template, short nonfiction only, a 2024 Claude model rather than the ones that carry today's watermark, and readers chosen because they already use AI often. The authors themselves do not claim the results carry over to other kinds of writing.

Sources: Russell, Karpinska and Iyyer, arXiv 2501.15654 (2025), Tables 1 and 2 (read 25 Oct 2026)

Readers lean on seven clues: four that do point toward Claude, and three that misled the readers who used them. None proves anything on its own, and the weights below are our own reading of the evidence.

1. The "this isn't X, it's Y" turn. Anthropic says models seem fond of it. People use it too, in speeches and ads. Weight: moderate, and only when it keeps coming back.
2. "Quietly", again and again. Anthropic notes models use the word more than you would expect. One "quietly" means nothing. Weight: low, only in volume.
3. Emily or Sarah in invented examples. In the 2025 study, 70% of the Claude 3.5 Sonnet articles named one of the two. The paper gives no figure for human articles, and the habit belongs to one prompt and one older model. Weight: low.
4. Words on an "AI vocab" list. The one expert who leaned on such words, without knowing Claude's own favourites, dropped to catching 80% of Claude articles while flagging 30% of human ones. Weight: low; lists go stale.
5. A neutral, even tone. Readers who did not write with AI marked neutral prose as machine-made and produced many false alarms. Weight: very low.
6. Fancy or rare words. The same group read any unusual word as a sign of AI. Weight: very low.
7. Grammar, judged either way. Those readers also believed people write more grammatical sentences than machines, a belief that misled them. Weight: none.

What carries weight is a cluster: several of the stronger clues in one passage, from a writer whose earlier work reads differently. Even then it is a reason to ask questions, not a finding.

Sources: Anthropic, Claude text watermark explainer (clues 1 and 2) · arXiv 2501.15654 (clues 3 to 7; Table 2 for annotator 2)

Because people write the same way for reasons of their own, and the habits that gave a model away last year fade as new versions arrive.

Second-language writers carry a known risk. A 2023 paper by Liang and colleagues found that style detectors "consistently misclassify non-native English writing samples as AI-generated."

Human readers fall into the same trap. In the 2025 study, the group with little or no AI writing habit treated a calm tone and any uncommon word as a sign of a machine. Even among the frequent users, the one who relied most on stock "AI words" missed a fifth of the Claude articles and flagged nearly a third of the human ones; the authors report this as an observation, not a tested cause.

Clues also age. The study used a Claude model from 2024; the models Anthropic marks today are newer, and a list built on old output tells you little about them.

Sources: Liang et al., arXiv 2304.02819 (2023) · Russell, Karpinska and Iyyer, arXiv 2501.15654

No. Anthropic says nothing is added to Claude's text and it contains no hidden characters, so a marked reply looks exactly like an unmarked one.

The mark sits in which words Claude picks, using a version of the SynthID-Text method Google DeepMind published in 2024. Per Claude's help center, it goes along when the text is copied and pasted, and may survive some editing. The mechanics are on our SynthID text watermark page. Anthropic's free online checker reads files only; its page states plainly that the tool does not check text.

Sources: Anthropic explainer, updated 1 Sep 2026 · Claude help center · claude.com/check-content (all read 25 Oct 2026)

Only organizations Anthropic admits to its detection API, and what they get back is a likelihood that Claude helped write the text, not proof of who wrote it.

The API is in private preview. Anthropic's help center lists regulators, law enforcement, media outlets, fact-checkers, independent researchers, educational organizations and EU civil society groups, plus companies with their own duty to verify content. Anthropic says it plans to open access further over time.

The answer has limits Anthropic spells out. It responds to one question, how likely it is that Claude wrote part of the passage, and it "doesn't confirm whether the text was human-written." Short samples hold too few word choices to judge well, while confidence rises as a passage grows. A light edit probably leaves the mark in place; replacing every word takes it out. How this compares with the free tools sold as checkers is on our Claude AI detector page.

Sources: Anthropic explainer · Claude help center, eligible organizations (read 25 Oct 2026)

Gather evidence of how the piece was made and talk to the writer before you decide anything.

1. Read the whole piece and count how many of the stronger clues cluster, rather than stopping at the first one.
2. Set it beside the person's earlier writing: vocabulary, rhythm, the mistakes they usually make.
3. Ask for drafts, notes and version history. A document that grew over days tells you more than any phrase.
4. Ask openly how they worked and whether an assistant helped.
5. If the stakes are high, ask an organization with access to Anthropic's key, and treat its answer as a likelihood.
6. Never decide on one clue or one score. Writers working in a second language are often misjudged by style detectors.

Source: Liang et al., arXiv 2304.02819 (step 6)

Put it in your own words, and let the rewrite tab of our Claude watermark remover do the first pass.

The rewrite replaces the wording and returns a receipt: words in, words out, hidden characters removed, and the share of your original three-word runs it kept, so you can see how far the text has moved. Claude adds no hidden characters, so the free cleaning tab on its own leaves Claude's wording untouched; new phrasing is what changes it. Then read it through and make it yours: your examples, your opinions, your rhythm.

Your first four rewrites of up to 250 words are free after an email sign-in, and the character cleaner is always free. Plans start at $12 a month on yearly billing (Basic is $18 monthly, Pro $27, Ultra $48); details are under plans and pricing. Whatever your school, employer or publisher asks you to disclose about AI help still applies to the final text.