New Graphite Study Finds AI Models Still Have Distinctive Writing Tells, Even After Scrubbing Em-Dashes

The em-dash is basically extinct in AI writing now. That was supposed to be the tell. But according to new research from marketing firm Graphite, killing off one obvious quirk just made room for a dozen subtler ones to take its place.

Graphite's study, led by chief AI officer Greg Druck, set out to measure something that's been mostly vibes-based until now: what words and phrases make AI-generated prose feel like AI-generated prose. The team built a corpus of 10,000 articles published before ChatGPT existed, using them as a human-written control group. Then they had various frontier models rewrite those same articles from summaries, a clever trick meant to cancel out source bias so the comparison is actually about word choice and sentence construction, not just different source material.

The result: 13,000 phrases that showed up at least twice as often in AI output as in human writing. Graphite calls anything that clears that bar a "tell." That's a surprisingly large number, and it suggests the stylistic fingerprint of these models runs a lot deeper than "delve" and semicolons.

Claude Opus 5.5's signature habit is oddly specific. The model uses the word "dependable" 23 times more often than human writers do. It's also apparently kicked the "it's not X, it's Y" construction that got called out in an earlier round of AI-writing-tell coverage back in April. Old habits die hard, though: Opus 5.5 has simply pivoted to a close cousin, telling readers something "is more than an X, it's a Y."

But the real standout is Opus's compulsion to narrate its own relevance. The phrase "this matters" shows up 116 times more often than in human samples, and "why X matters" clocks in at 92 times more common. If you've ever felt like a chatbot was narrating its own importance to you unprompted, this is probably why.

How the Labs Compare

OpenAI's Astra model has its own set of fingerprints, and they're different enough from Opus's that you could probably build a rough classifier just off vocabulary alone.

• Astra loves invoking "another dimension" of whatever topic it's discussing.
• It hedges constantly, leaning on phrases like an action "may provide" or "can provide" some benefit.
• Its biggest tell by far is what Graphite dubs "corrective framing": constructions like "not simply X" or "rather than relying on X." Those patterns showed up more than 100 times more often in Astra's writing than in human samples.

That's a notably different flavor of AI-speak than Opus's sincerity-forward "this matters" energy. Astra sounds more like it's perpetually correcting a strawman; Opus sounds like it's trying to convince you of its own importance.

Where the models agree is on the em-dash purge. Opus 5.5 uses the punctuation mark 99% less than its predecessor, Opus 5. Astra has cut em-dash usage by 88% compared to human baseline rates. Gemini 3.1 Pro has essentially eliminated it. Whatever the labs did here, between RLHF tuning, system prompt tweaks, or just awareness of the meme, it clearly worked, at least for this one specific tell.

Here's the more interesting finding, though: killing the em-dash didn't shrink the overall tell count. Druck says the total volume of detectable quirks has stayed roughly constant even as individual signatures evolve. "It's not like the tells are decreasing," he told TechCrunch. "They are managing to remove the most well-known tells, but other ones pop up. And every model version has its own."

There's also a divergence in direction worth flagging. Graphite found that Claude models are trending closer to the human word distribution with each release, while GPT models are trending further away from it. That's a strange split given both labs are presumably optimizing for similarly human-sounding output, and it suggests different training approaches are producing genuinely different stylistic trajectories, not just different flavors of the same drift.

This all sits awkwardly next to what the labs themselves have claimed. Anthropic's Opus 5.5 release notes said the model "communicates more naturally than prior models," citing early users who found the writing "clearer and easier to follow." OpenAI made nearly identical claims when it shipped the GPT-6 versions of Sol and Luna, promising "more clarity, less jargon, fewer odd turns of phrase." Both of those claims are presumably true in some narrow sense. Neither stopped a 13,000-phrase tell list from existing six months later.

Druck's explanation for why these patterns persist despite deliberate effort to stamp them out is pretty simple: scale makes precision hard. "A general hypothesis I have is that the labs are less able to control some of these things than you might expect," he said. "These are giant models with billions of parameters. They have some finite number of tests they can run, and things slip through."

Key Takeaways

For anyone building detection tools, content moderation pipelines, or just trying to figure out if a document was AI-assisted, this is a useful reminder that tells are a moving target, not a fixed signature. The specific words change release to release. The underlying tendency, some statistical residue of RLHF and instruction tuning leaking into surface-level word choice, doesn't seem to be going away.

Practically speaking, if you're building a classifier around known AI tells, expect to retrain it every time a model version ships. Em-dash detection is basically dead as a signal now. But "this matters," "dependable," and "another dimension" are, for the moment, live wires. Check back in six months and the list will probably look completely different, while staying exactly as long.