As AI-generated text becomes increasingly widespread, distinguishing between human writing and machine-generated content has become more difficult. Although some of the earlier indicators once relied upon by readers have faded, a recent study shows that AI models continue to leave behind recurring linguistic patterns that can be identified and analyzed.
اضافة اعلان
Graphite conducted the study by comparing human-written texts with content 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 a human-written article and an article generated by each model. This allowed them to directly compare the frequency of words and phrases and patterns of sentence construction.
Thousands of Signals Still Reveal AI Writing
The study found that Claude Opus 5.5 still has 2,548 linguistic signals that can indicate AI-generated writing, compared with 2,666 signals in Opus 5, a decline of just 4%. The number stood at 3,746 in Opus 4 before falling to 3,059 in Opus 4.6.
These signals 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 Is Important” Among Claude’s Key Signatures
The findings showed that Claude Opus 5.5 has a notable tendency to use phrases emphasizing the importance of a subject. The phrase “this is important” appeared 116 times more frequently than in human writing, while “why this is important” 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” appeared more than 32 times as often.
The study also identified specific words used by Claude Opus 5.5 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 often. Words such as “practical” and “consistent” appeared at rates of around 11 times the human level.
A word meaning “considered” appeared more than 9.2 times as frequently, while a word meaning “authentic” appeared at roughly the same rate. The word “broader” appeared about 8.9 times as often as in human writing.
Claude’s Language Moves Closer to Human Writing
The findings indicate that word distribution in Claude Opus 5.5 has become closer to human writing compared with previous versions. The divergence measure fell from 0.064 in Opus 5 to 0.052 in Opus 5.5, a decline of approximately 19%.
By contrast, the measure for GPT models rose from 0.101 in GPT-5.6 Sol to 0.109 in GPT-6 Astra, an increase of about 8%, indicating that Astra’s word distribution became relatively further from the human pattern according to this measure.
Clear Decline in Ornate Writing
The study found that Claude Opus 5.5 uses what researchers described as “ornate prose” less frequently than its previous version.
Opus 5.5 scored 10.57 out of 100, compared with 16.75 for Opus 5, representing a 37% decline. Its score, however, remains about 1.6 times the human average, while GPT-6 Astra stands at around 1.2 times the human level.
Em Dash Use Nearly Disappears
Another notable change was the sharp decline in the use of em dashes.
Claude Opus 5.5 used approximately 0.015 em dashes per 1,000 words, compared with 2.92 in Opus 5, a 99% decline.
GPT-6 Astra also uses em dashes at a rate far below that found in human writing, while their 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 writing patterns.
GPT-6 Astra tends to be more cautious when formulating claims. For example, a phrase meaning “may provide” appeared 18 times more frequently than in Opus 5.5, while “not necessarily” appeared more than 17 times as often.
Claude Opus 5.5, meanwhile, uses superlative expressions more frequently. A phrase meaning “most popular” appeared more than 45 times as often as in Astra, “perhaps the most” appeared 29 times more frequently, and “most powerful” appeared 24 times more often.
Help-Oriented Phrases Become More Common
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 often, “can help you” eight times as often, “makes things easier” six times as often, and “helps you avoid” five times as often as in the previous version.
Older Signals Decline but Do Not Disappear
The study tracked 11 linguistic features previously considered among the most prominent indicators of AI-generated writing, including specific words, formulaic conclusions and cautious expressions.
The average strength of these features declined by 6% in Opus 5.5 compared with Opus 5, and by 53% compared with Opus 4.
For GPT models, the strength of these indicators declined by 46% in Astra compared with GPT-4.1, but changed little after GPT-5.
Writing Signatures Change With Each Version
The study’s main finding is that AI writing signals do not disappear completely; instead, they change with each new model version.
As a model drops some well-known phrases or reduces their use, 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 linguistic structures at rates that are higher than those typically found in human writing.
