Look for a cluster, not a single sign: stock phrases such as "the importance of" and "is essential", more em dashes than people usually write, neat lists of three, and sentences of very even length. Humans use all of these too, so a cluster is a reason to check, not proof. The figures below come from Fileverity's proprietary corpus of fully AI-written English from the OpenAI family compared with human writing.
Phrases that ChatGPT-family text uses more than people
Multipliers show how much more often a phrase appeared in AI-written text than in human writing.
- Importance and necessity. "is essential" (5.8 times), "importance of" and "the importance" (5.2 times), "a clear" (4.9 times).
- Contrast and framing. "while others" (7.4 times), "and whether" (6.9 times), "whether the" (4.4 times), "rather than" (5.1 times).
- Purpose and effect. "ensuring that" (5.7 times), "intended to" (3.8 times), "to address" (3.6 times), "contribute to" (4 times), "lead to" (2.5 times).
- Soft guidance. "this approach" (5.8 times), "offers a" (5.9 times), "easier to" (4 times), "continues to" (4 times), "often" (3.3 times).
- Atmosphere. "a sense" (4.3 times), "a moment" (4.1 times), "toward the" (3.9 times).
Phrases that set ChatGPT apart from other AI families
Compared with other AI models, ChatGPT-family text leaned toward "intended to" (4.2 times), "useful" (4.8 times), "and whether" (4.9 times), "affect" (3.4 times), "try to" (3.3 times), "easier to" (3.1 times) and "can help" (3 times).
Structural markers
More of these than human writers:
- Em dashes, about 8 times the human rate.
- Three-item lists in the "X, Y and Z" pattern, about twice the human rate.
- Curly quotation marks, about 1.7 times the human rate.
Less of these than human writers:
- Parentheses, about 3 times less often.
- Ellipses, about 3.5 times less often.
- Exclamation marks, about 2 times less often.
- Colons, a little less often.
- Variation in sentence length: AI sentences were more uniform, and slightly shorter on average.
Why phrases are not proof
Humans use every phrase and mark above, especially trained writers and anyone who has read a lot of formal prose. A student who writes "the importance of" is not a chatbot. Style is a signal, and editing, paraphrasing or a prompt that asks for a casual voice can remove many of these tells. Read the whole text, not the list.
What a detector reads that a list cannot
Veread reads patterns across whole passages, not isolated phrases. It marks the passages that carry the AI signal and shows writing-style similarity to ChatGPT and other families. In tests it wrongly flagged 0.5% of human-written text across seven languages and caught 98.3% of fully AI-written texts; the accuracy figures explain both. If you are reviewing student work, read AI detection for teachers.
Check a text now
Paste at least 100 words into the ChatGPT detector and see the verdict, the marked passages and the style similarity.
To compare all seven model families, see the guide on how each AI model writes.