Claude's text tends to use certain connective phrases much more than human writers, and to avoid ellipses, exclamation marks and parentheses. No phrase is proof, because people write this way too, but a cluster of them is a signal. The figures below come from Fileverity's training corpus, comparing fully AI-written English documents from the Anthropic family with human writing.
Phrases Claude uses more than humans
Multipliers show how much more often the phrase appears than in human writing, after matching the genre and source mix.
- Contrast and emphasis: "far more" (11×), "yet the" (10×), "though the" (3.8×), "rather than" (5.2×)
- Hedging and tendency: "tends to" (4.2×), "depends on" (3.3×), "often" (2.8×), "plenty of" (2.9×)
- Practical framing: "practical" (6.5×), "right away" (6.4×), "matters" (4.2×)
- Ownership and agents: "its own" (3.9×), "their own" (2.5×), "someone who" (4.1×)
- Direction and scope: "toward the" (3.5×), "across the" (3.1×), "among the" (3×), "along with" (2.3×)
Phrases that set Claude apart from other AI
Many phrases are common to all AI text, so these compare Claude only with other AI families.
- "depends on" (3.6×), "depend on" (3×), "cannot be" (2.2×)
- "the whole" (2.7×), "a person" (2.7×), "a little" (2.3×), "a few" (1.7×)
- "since the" (2.6×), "its own" (2.1×)
- "i think" (2.6×), a more personal register than other families use
Structural markers
Against human writers, Claude uses:
- More three-item lists ("X, Y and Z"): 4.1 per 1,000 words against 2.6 for humans.
- More colon-introduced lists: 16% of texts against 11%.
- Far fewer ellipses (13× less often), exclamation marks (3.1× less), parentheses (2.5× less) and semicolons (2.5× less).
- Fewer question marks and fewer bullet or numbered lines in plain prose.
Against other AI families, Claude uses em dashes 3.7× less often and "not X but Y" contrasts 3× less often. If a text is full of em dashes and "not X but Y" lines, that points away from Claude, not towards it.
Read these tells with care
- Humans use all of these phrases. A single "rather than" means nothing.
- Style is a signal, not proof of authorship. Prompts, editing and genre change it.
- Short texts carry little signal, and a light human edit can hide the pattern.
- The AI texts in the corpus were generated as plain prose, so list and heading markup says little about real chat output.
- The same method gives different tells in French, German, Spanish, Chinese, Vietnamese and Arabic, so English tells do not transfer.
Why a phrase list is not a detector
A list of phrases can be matched by anyone and defeated by a quick edit. Veread, the detection model behind this site, reads far more than phrases: how sentences are built, how the text moves from one idea to the next, rhythm and structure across the whole text. It works across seven languages and shows which passages carry the signal. In published tests it caught 98.4% of fully AI-written Claude text and wrongly flagged 0.5% of human-written text; see the accuracy figures.
Check a text yourself
Paste at least 100 words into the Claude AI detector to see the AI verdict, the marked passages and the style similarity to Claude. To compare all seven model families, see the guide on how each AI model writes.