Study Finds Opus 5.5 Still Writes Like AI
Research agency Graphite identified about 2,500 writing tells in articles generated by Anthropic's Opus 5.5 model.
Updated on Nov. 12, 2026 in Artificial Intelligence

Live Poll
Do you trust that you can usually tell AI-written text from human writing?
Graphite compared 10,000 articles generated by Opus 5.5 with 10,000 human-written articles on the same topics and found roughly 2,500 writing tells. The model uses em dashes at a 99% lower rate than its predecessor, Opus 5.
Why it matters
The findings show that reducing one stylistic habit does not eliminate the broader patterns that make AI-generated text recognizable to readers. For publishers and writers evaluating AI-assisted content, the study offers a measurable benchmark of how far current models still diverge from human prose.
The analysis compared two matched sets of 10,000 articles and defined a tell as a word or pattern occurring at least twice as often in AI writing as in human writing. Graphite's measure of mannered prose is its own scale, and the study does not yet show how these metrics translate to reader perception.
The players
Anthropic
Anthropic is an artificial intelligence company that develops the Claude family of large language models, including Opus 5.5.
Graphite
Graphite is a research-based growth agency that conducted the comparative analysis of AI-generated and human-written articles.
The details
Graphite, a research-based growth agency, defined tells as words, phrases and patterns that occur at least twice as often in AI writing as in human writing. The top one-word tell, 'dependable,' appeared 23 times more often in Opus 5.5 output than in human writing, while the top two-word tell, 'matters because,' appeared 146 times more often.
Timeline
• Article on the Opus 5.5 writing-tells study published November 12, 2026.
Big Picture
Opus 5.5 cut em dash usage by 99% and mannered prose by 37% compared with Opus 5, yet still produced roughly 2,500 detectable tells in Graphite's comparison. The result suggests targeted fixes to individual habits do not fully close the gap with human writing.
Writers and editors who rely on AI drafting should expect that stylistic cleanup alone may not make output indistinguishable from human prose. Readers encountering AI-assisted articles may still notice recurring phrasing even when obvious punctuation habits are removed.
The takeaway
If you edit AI-generated text, check for repeated stock phrases and inflated connectors, not just punctuation. Measuring output against human samples on the same topics gives a more reliable sense of how distinctive the prose remains.
Further reading
For more on the models behind these findings, see Artificial Intelligence coverage on Wisevoter.
Source note: This article includes information reported by Culpeper Star-Exponent.
Live Poll
Do you trust that you can usually tell AI-written text from human writing?



