Opus 5.5 features a shift toward shorter, simpler expressions, making it feel more accessible while still retaining recognizable AI writing cues. While this makes it feel more accessible, the model still retains certain recognizable AI writing cues.
A recent study by Graphite examined the writing habits of advanced AI models, revealing unique quirks and signs that often give away AI-generated content. Researchers found an astonishing 13,000 phrases that appeared at least twice as often in AI-generated text compared to human writing. Greg Druck, Graphite’s chief AI officer, noted that these patterns persist even as models attempt to craft more human-like prose. “It turns out that Claude models are actually getting closer to the human word distribution over time,” Druck said. “And for the GPT models, it’s getting further away.”
One of the most striking observations about Opus 5.5 is its frequent use of the phrase “this matters,” which pops up an impressive 116 times more often than in human writing. Similarly, the construction “why X matters” appears 92 times more frequently. This consistent focus on relevance may indicate the model’s effort to engage readers by emphasizing importance.
Interestingly, while Opus 5.5 has moved away from certain phrases like “it’s not X, it’s Y,” it still favors constructions that frame ideas in contrasting terms. For instance, it often states something “is more than an X, it’s a Y.” These linguistic choices highlight a broader trend among AI models as they seek to express thoughts in ways that resonate with human readers.
OpenAI’s Astra, on the other hand, exhibits a different set of writing habits. This model often introduces concepts with the phrase “another dimension” and tends to hedge its claims using qualifiers like “may provide” or “can provide.” A notable feature of Astra-generated text is its use of “corrective framing,” where topics are frequently presented as “not simply X” or framed as alternatives with phrases like “rather than relying on X.” Graphite’s research indicates that these constructions are more than 100 times more common in Astra’s outputs than in human writing.
A notable trend across the board is the decreased use of em-dashes. Opus 5.5, for example, employs them 99% less often than its predecessor, while Astra has reduced usage by 88%. Meanwhile, Gemini 3.1 Pro has almost completely eliminated em-dashes from its writing. This decline reflects an awareness among AI developers of complaints about overused punctuation in AI-generated text.
Despite some individual markers changing or disappearing, the overall number of identifiable patterns remains relatively stable. Druck expressed doubt about the labs’ ability to fully eliminate these distinctive constructions. “A general hypothesis I have is that the labs are less able to control some of these things than you might expect,” he stated. “These are giant models with billions of parameters. They have some finite number of tests they can run, and things slip through.”
As AI writing continues to develop, it’s clear that while models like Opus 5.5 strive for a more natural communication style, they still retain specific markers that can reveal their origins. These are not just quirks; they serve as reminders of the ongoing interplay between human-like expression and algorithmic patterns. Recognizing these cues can help readers and critics alike discern the subtle differences in AI-generated content.
