New Study Reveals 13,000 AI Writing Tells as Models Struggle to Mimic Human Text

Thousands of New Indicators of AI-Generated Text Discovered

A comprehensive study by marketing firm Graphite has uncovered approximately 13,000 phrases common in AI content that serve as telltale signs of machine-generated prose1

1

2

2

. These AI tells represent words and phrases appearing at least twice as frequently in AI writing compared to human-produced text. The research examined frontier AI models including Anthropic's Claude Opus 5.5 and OpenAI's Astra, revealing distinct AI writing patterns unique to each system.

Graphite's chief AI officer Greg Druck explained the study's methodology involved analyzing 10,000 articles published before ChatGPT's release as a human-generated control group. Researchers then had different AI models rewrite these articles from summaries to eliminate source bias, creating matching samples that revealed how often certain constructions appeared across human-like writing versus AI-generated text1

1

.

Claude Opus 5.5 Emphasizes Significance Through Repetitive Phrasing

Claude Opus 5.5's most prominent tell is the phrase 'this matters,' appearing 116 times more frequently than in human samples1

1

2

2

. The model also uses 'why X matters' 92 times more often than human writers. Additionally, the word 'dependable' appears 23 times more frequently in Opus 5.5 outputs, while 'looking ahead the' occurs 40 times more often2

2

.

While Opus 5.5 has largely abandoned the 'it's not X, it's Y' construction, it still favors saying something 'is more than an X, it's a Y.' The model qualifies claims less often than competitors and shows a stronger tendency toward using superlatives2

2

. Despite Anthropic's claims that Opus 5.5 'communicates more naturally than prior models' with writing that early users 'found clearer and easier to follow,' these new tells in AI writing persist1

1

.

OpenAI's Astra Demonstrates Different AI Writing Patterns

OpenAI's Astra exhibits its own distinctive set of indicators of AI-generated text. The model's biggest tell is the phrase 'does not establish,' appearing 275 times more frequently in Astra outputs than in Claude-generated content2

2

. Astra favors describing 'another dimension' of topics and tends to hedge claims by saying an action 'may provide' or 'can provide' particular benefits1

1

.

The model also relies heavily on what Graphite calls 'corrective framing,' defining topics as 'not simply X' or offering alternatives 'rather than relying on X.' These constructions appear more than 100 times more commonly in Astra-generated prose compared to human writing1

1

. OpenAI made similar promises to Anthropic when releasing GPT-6 versions of Sol and Luna, claiming users could 'expect to see more clarity, less jargon, and fewer odd turns of phrase'1

1

.

Em-Dash Usage Plummets Across Frontier AI Models

All frontier labs have responded to criticism about excessive em-dash usage in AI-generated text. Opus 5.5 now uses the punctuation mark 99% less often than Opus 5, while Astra employs it 88% less than human samples1

1

2

2

. Gemini 3.1 Pro has almost completely eliminated em-dashes from its writing1

1

.

This dramatic reduction demonstrates that labs can address specific AI tells when they become widely recognized. However, the overall number of tells remains steady, creating a whack-a-mole dynamic where eliminating one set of phrases causes others to emerge2

2

.

Diverging Trends Between Claude and Astra Models

Graphite's research reveals contrasting trajectories for different frontier AI models. 'It turns out that Claude models are actually getting closer to the human word distribution over time,' Druck told TechCrunch. 'And for the GPT models, it's getting further away'1

1

2

2

.

This finding suggests Claude is moving toward more human-like writing while Astra trends toward less natural prose. The divergence raises questions about different training approaches and whether labs can fully control their models' linguistic quirks. '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

.

Why Detection Matters for Content Authenticity

Druck remains skeptical about whether labs can completely eliminate telltale constructions. 'A general hypothesis I have is that the labs are less able to control some of these things than you might expect,' he says. 'These are giant models with billions of parameters. They have some finite number of tests they can run, and things slip through'1

1

.

The persistence of AI tells has fueled a micro-economy of AI detectors now worth millions of dollars2

2

. As AI writing becomes ubiquitous, identifying these patterns helps readers distinguish between human and machine-generated content. Watch for continued evolution in how models construct sentences and whether the gap between AI and human prose narrows or widens in coming releases.