Claude detector

A quieter voice

Anthropic tunes Claude to sound thoughtful rather than enthusiastic. Compared with other assistants it uses fewer exclamation marks, fewer emoji and less marketing language, and newer models deliberately vary their sentence structure. The result reads more like a careful human writer — which is exactly what makes it harder for pattern-based detection.

It still has habits. Longer Claude answers often open by restating or reframing the question, then lay out considerations in balanced pairs (“On one hand… on the other…”), qualify claims with “generally”, “often” or “in many cases”, and use em dashes for asides. It likes phrases such as “it’s worth noting”, “that said” and “the key insight is”, and it ends with a short synthesis rather than a formal “In conclusion”. Wikipedia editors have documented many of these as general signs of AI writing.

What our test data shows

Our evaluation set includes encyclopedia-style passages from Claude Opus 5 and Claude Haiku 4.5 written in 2026, alongside human Wikipedia text from before 2022, academic writing by non-native English speakers and public-domain literature. The standard check — signs and statistics only — misses most Claude Opus 5 passages. The deep scan, which adds a neural model, catches all of the Claude Haiku passages and a minority of the Opus ones, without raising false alarms on the human sets above about 2%.

In other words: a high score on suspected Claude text is meaningful, but a low score is weak evidence that a human wrote it. Turn on Deep scan for Claude-suspect text, and look at the accuracy table below for the per-model breakdown.

When the stakes are high

For coursework, hiring tests or publishing decisions, combine any detector with evidence of process: drafts and version history, notes and sources, and a conversation with the writer about the ideas in the text. Writers who used Claude as a tutor or editor rather than a ghostwriter will usually show a clear path from first draft to final version.

Signs worth reading closely in Claude text

Because Claude’s vocabulary is less formulaic, the structural signs carry more weight: a reframing first sentence, paired considerations, frequent em-dash asides, and hedged generalisations (“this often depends on…”). Look at the highlighted copy for these patterns and at the paragraph-uniformity statistic, which stays elevated even when word choice looks natural. A long text that shows several of them, plus a high deep-scan score, is a much stronger signal than any one alone.

One practical tip: check long documents in sections. Claude is often used to draft one part of a piece — an introduction, a summary, a cover letter paragraph — and a whole-document score averages that part with the human-written rest.