New Study Uncovers 13,000 AI Tells: Opus 5.5 Says 'This Matters' 116x More Than Humans

AI Writing Patterns Exposed Across Frontier Models

A 1

1

has uncovered approximately 13,000 AI tells—words and phrases appearing at least twice as often in AI-generated text compared to human writing. The research examined frontier AI models including Anthropic's Claude Opus 5.5, OpenAI's Astra, and Gemini 3.1 Pro, revealing distinct writing patterns that distinguish AI-generated text from human prose1

1

. Greg Druck, Graphite's chief AI officer, noted that while labs successfully eliminate well-known tells, new ones consistently emerge with each model version1

1

.

The marketing firm designed a careful methodology to identify these AI writing patterns. Researchers compiled 10,000 articles published before ChatGPT's release as a human-generated control group, then had different AI models rewrite the articles from summaries to eliminate source bias1

1

. This approach allowed precise comparison of word frequency and sentence construction patterns between human and AI writing.

Claude Opus 5.5's Distinctive Linguistic Fingerprints

Anthropic's Claude Opus 5.5 exhibits specific phrases that reveal AI writing, with "this matters" appearing 116 times more frequently than in human samples2

2

3

3

. The phrase "why X matters" occurs 92 times more often, while "matters because" appears 146 times the human rate3

3

. Single-word tells include "dependable" (used 23 times the human rate), "clearer" (14 times), and "quietly" (11 times)3

3

.

Despite Anthropic claiming Opus 5.5 "communicates more naturally than prior models," the research identified upwards of 2,548 tells in this version alone4

4

. While this represents improvement from Opus 4's 3,746 tells, the model still relies heavily on contrast-heavy constructions4

4

. Opus 5.5 frequently uses the pattern "is more than an X, it's a Y" and phrases like "especially valuable" (136 times the human rate) and "builds trust" (83 times the human rate)1

1

3

3

.

OpenAI's Astra Moves Away From Human-Like Writing

OpenAI's Astra demonstrates a different set of AI writing patterns and, surprisingly, shows increasing divergence from human prose. The model's word-distribution divergence is 8 percent higher than GPT-5.6 Sol, meaning Astra is actually trending toward less human-like writing4

4

. This contradicts OpenAI's claims that users could "expect to see more clarity, less jargon, [and] fewer odd turns of phrase"1

1

.

Astra's biggest tell involves corrective framing, where topics are defined as "not simply X" or offered as alternatives "rather than relying on X"—constructions appearing more than 100 times more often than in human writing1

1

. The model also loves describing "another dimension" of topics and hedges claims by saying actions "may provide" or "can provide" particular benefits1

1

. The phrase "does not establish" appears 275 times more frequently in Astra outputs compared to Claude's text2

2

. Astra qualifies claims far more often than Opus 5.5 and uses superlatives less frequently2

2

.

The Whack-a-Mole Dynamic of AI Tells

Frontier labs have clearly responded to criticism about em-dash usage. Opus 5.5 uses em dashes 99 percent less often than Opus 5, while Astra employs them 88 percent less than human samples1

1

4

4

. Gemini 3.1 Pro has almost completely eliminated the punctuation mark from its writing1

1

. Both Opus 5.5 and Astra now use em-dash less than human writers4

4

.

However, eliminating one set of tells creates a whack-a-mole dynamic where new patterns emerge. "It's not like the tells are decreasing," Druck explained. "They are managing to remove the most well-known tells, but other ones pop up. And every model version has its own"1

1

. The research suggests labs may have limited control over these linguistic patterns despite their billions of parameters. "These are giant models with billions of parameters. They have some finite number of tests they can run, and things slip through," Druck noted1

1

.

Diverging Trajectories in Human-Like Writing

The Graphite study reveals contrasting trends between major AI models. Claude models are moving closer to human word-distribution over time, with Opus 5.5 showing 19 percent lower word-distribution divergence compared to Opus 51

1

4

4

. Meanwhile, GPT models are getting further away from human writing patterns1

1

.

This matters because as AI-generated content becomes ubiquitous, the ability to distinguish AI-generated text from human writing remains crucial for transparency and authenticity. The persistence of thousands of AI tells suggests that despite rapid advances, frontier AI models still have considerable distance to travel before achieving truly indistinguishable human-like writing. Watch how labs address these newly identified patterns in upcoming releases, and whether the trend toward more or fewer tells continues across different model families.