Graphite conducted the study by comparing human-written texts with texts generated by several advanced models, including Claude Opus 5.5 and GPT-6 Astra. Researchers examined articles covering 9,974 identical topics, with each topic represented by one human-written article and one article generated by each model. This allowed them to directly compare the frequency of words and phrases as well as sentence-structure patterns.
اضافة اعلان
Thousands of indicators still reveal AI-generated writing
The study found that Claude Opus 5.5 still has 2,548 linguistic indicators that can signal AI-generated writing, compared with 2,666 indicators in Opus 5 — a decline of just 4%. The number of indicators stood at 3,746 in Opus 4 before falling to 3,059 in Opus 4.6.
These indicators include words, phrases and linguistic patterns that appear at least twice as frequently as they do in human writing, after adjusting the results for text length and applying minimum frequency thresholds.
“This matters” among Claude Opus 5.5’s strongest fingerprints
The findings show that Claude Opus 5.5 has a notable tendency to use phrases that emphasize the importance of a topic. The phrase “this matters” appeared 116 times more frequently than in human writing, while “why this matters” appeared 92 times more frequently.
A phrase meaning “more than just something” appeared up to 98 times more frequently than in human texts, while “rather than just” was used more than 32 times as often.
The study also identified specific words that Claude Opus 5.5 uses at much higher rates than humans. A word meaning “reliable” appeared more than 23 times as frequently, while “clearer” appeared more than 14 times as frequently. Words such as “practical” and “consistent” appeared at rates of about 11 times those found in human writing.
A word meaning “considered” appeared 9.2 times more frequently, while a word meaning “authentic” appeared at roughly the same rate. The word “broader” appeared about 8.9 times more frequently than in human texts.
Claude becomes linguistically closer to humans
The findings indicate that the distribution of words in Claude Opus 5.5 has become closer to human writing compared with previous versions. The divergence score fell from 0.064 in Opus 5 to 0.052 in Opus 5.5, a decline of about 19%.
By contrast, the measure increased among GPT models, from 0.101 in GPT-5.6 Sol to 0.109 in GPT-6 Astra an increase of about 8%. This suggests that Astra’s word distribution is relatively farther from the human pattern according to this measure.
Clear decline in ornate prose
The study found that Claude Opus 5.5 uses what it describes as “ornate prose” less frequently than its predecessor.
Opus 5.5 scored 10.57 out of 100, compared with 16.75 for Opus 5, representing a 37% decline. However, its score remains about 1.6 times the human average, while GPT-6 Astra stands at about 1.2 times the human level.
Em dash use nearly disappears
Another notable change is the sharp decline in the use of the em dash.
Claude Opus 5.5 used about 0.015 em dashes per 1,000 words, compared with 2.92 in Opus 5 a decline of 99%.
GPT-6 Astra also uses the em dash at a much lower rate than human-written texts, while its use has almost completely disappeared in Opus 5.5.
Clear differences between Claude and Astra
The study found that each model has its own distinctive fingerprints.
GPT-6 Astra tends to be more cautious when framing claims. For example, a phrase meaning “may provide” appeared more than 18 times as frequently as in Opus 5.5, while “not necessarily” appeared more than 17 times as frequently.
Claude Opus 5.5, meanwhile, uses superlative expressions more often. A phrase meaning “most popular” appeared more than 45 times as frequently as in Astra, while “perhaps the most” appeared 29 times more frequently and “most powerful” 24 times more frequently.
Helpful language becomes more prominent
Compared with Opus 5, Opus 5.5 uses language associated with assistance and making tasks easier more frequently.
A phrase meaning “particularly useful” appeared more than 12 times as frequently, while “can help you” appeared eight times as frequently. “Makes it easier” appeared six times as frequently, while “helps you avoid” appeared five times as frequently compared with the previous version.
Older indicators are declining but not disappearing
The study tracked 11 linguistic features that were previously considered among the most prominent indicators of AI-generated writing, including specific words, formulaic endings and hedging expressions.
The average strength of these features in Opus 5.5 fell by 6% compared with Opus 5 and by 53% compared with Opus 4.
In GPT models, the strength of these indicators fell by 46% in Astra compared with GPT-4.1, but changed little after GPT-5.
AI fingerprints change from one version to another
The study’s main conclusion is that AI-writing indicators do not disappear completely; rather, they change with each new model version.
While a model may eliminate or reduce its use of certain well-known phrases, new patterns emerge in their place. As a result, a model can become more similar to humans in its word distribution while still using thousands of words, phrases and structures at rates higher than those typically found in human writing.
Resource: Al-Ghad.
