Claude Opus 5.5 Trades Old AI Tells for New Ones

The competitive question for frontier AI is no longer only which model can produce the most fluent prose. It is whether that fluency can remain flexible when the same system is asked to write an email, explain a technical issue, summarize research or make a persuasive case. If each model develops a recognizable vocabulary across those tasks, then better imitation of human writing may not be the same as genuinely humanlike versatility.

“It turns out that Claude models are actually getting closer to the human word distribution over time,”

That is the implication of a new analysis from Graphite, which compared human and AI writing across thousands of matched topics. Graphite’s research examined 9,974 topic pairs and identified 2,548 writing patterns associated with Claude Opus 5.5. The findings suggest that the model has reduced some familiar habits, but has not escaped the broader problem of machine generated prose leaving statistical traces.

The most prominent examples are ordinary words and phrases. Graphite reported that Opus 5.5 used “dependable” 23 times more often than the human comparison material. It used “this matters” 116 times more often. The phrases are not incorrect, unusual or inherently artificial. Their significance comes from repetition at a scale that makes them useful as clues.

Opus 5.5 also appears to have changed one of the most visible stylistic signals associated with earlier AI writing. Graphite reported a 99 percent reduction in em dash use compared with Opus 5. That change matters because punctuation habits became an informal way for readers and editors to suspect that a passage had been generated by a model. The newer system may be less likely to trigger that particular suspicion, even as other patterns become more noticeable.

Anthropic’s own positioning helps explain why the distinction is important. On its official Claude Opus model page, Anthropic says Opus 5.5 “communicates more clearly,” leads with what matters, avoids jargon and produces writing that is easier to follow. The phrases identified by Graphite appear to sit close to those stated goals. A model that regularly explains why something matters may be following a deliberate communication strategy, rather than accidentally repeating a quirk.

That creates a difficult measurement problem. A behavior can be useful to readers and still function as an AI tell. Leading with the important point is generally good advice. Avoiding jargon can improve accessibility. But if a model applies the same framing across too many subjects, the technique starts to look less like judgment and more like a template.

The tell has moved, not disappeared

TechCrunch reported that the Graphite comparison also found distinctive patterns in OpenAI’s Astra and Google’s Gemini 3.1 Pro. That comparison places Opus 5.5 in a wider industry pattern. Each frontier model can sound polished while still producing recurring choices that distinguish it from both people and competing systems.

The result resembles an arms race in style. Model developers improve systems partly by removing behaviors that users associate with earlier generations. Those changes can make output feel fresher, but they can also expose new regularities. A phrase that appears helpful in one response becomes suspicious when it turns up across many unrelated subjects.

Greg Druck, Graphite’s chief AI officer, told TechCrunch that AI tells are not decreasing. In his account, models are removing well known signals while new ones appear with each version. That interpretation casts stylistic evolution as a moving target for detection. The model does not need to preserve its old fingerprint. It only needs to develop another one.

For businesses, the consequences are practical. Companies increasingly use generative AI to prepare customer messages, reports, marketing copy and internal documents. If a model’s vocabulary is too consistent, organizations may produce material that sounds uniform even when it is technically accurate. That can weaken trust, particularly in areas where readers expect expertise, personal judgment or a distinctive institutional voice.

The issue also affects efforts to identify AI generated text. A detector trained to look for old signals, such as excessive em dashes, may become less useful after a model update. New phrases could offer fresh clues, but those clues may not remain stable. A system that flags “this matters” today could become outdated if a future version learns to avoid the phrase.

A useful finding with a commercial caveat

The strongest criticism concerns not the existence of recurring model patterns, but how much confidence readers should place in Graphite’s interpretation. Graphite sells AI related products, so its research also draws attention to a problem that is relevant to its commercial position.

Gizmodo writer AJ Dellinger viewed the findings cautiously, pointing to Graphite’s status as a marketing company and describing the evolution of AI tells as a game of whack a mole. That skepticism is important because the analysis identifies correlations in word use, not a final test for whether a specific passage was written by a machine.

A phrase can be disproportionately common in model output without proving that every passage containing it was generated by AI. Human writers can use “dependable” or explain why an idea “matters” for entirely ordinary reasons. The same is true of punctuation. A sharp reduction in em dash use shows a change in the model’s output distribution, but it does not establish that the model now writes with greater stylistic independence.

The research is therefore more convincing as a description of model behavior than as a universal detector. It shows that Opus 5.5 has a measurable style. It does not, on the information provided, show that the model’s prose can be reliably separated from human writing in every context.

That distinction points to the larger question facing Anthropic and its rivals. Natural communication is not simply the removal of conspicuous habits. It requires choosing different structures, levels of explanation and tones because the situation calls for them. A model that replaces one repeated signal with another may sound less mechanical at first, while remaining just as patterned underneath.

Opus 5.5 may be better at hiding the tells that readers learned to recognize. The next test is harder: whether it can stop sounding like the same writer when the subject, audience and purpose change.