Can you tell AI writing apart by its word choices alone, or does the real tell run deeper, in what it says?

This explores what actually gives AI-generated text away: the model's habitual word choices and phrasing, or deeper differences in what it says and how it structures ideas. The short answer from the corpus is that phrasing is the easy signal and also the shallow one. Underneath it sits a second layer of content and structure that is harder to fake and arguably more revealing. The collection doesn't directly measure how much of the signal comes from each layer for specific models (it has no 'GPT vs. Claude fingerprint' study), but it does show how the two layers behave.

Start with surface style. It turns out to be remarkably easy to detect. Even a small, older model like GPT-2 can identify authors with about 95% accuracy from stylistic patterns alone, and performance saturates early. Style is a low-hanging signal, but catching it tells you nothing about why the choices matter Can language models truly understand literary style?. Part of why model phrasing is so recognizable is that models lean toward statistically common wording. They do measurably better on high-frequency paraphrases than on rarer ones that mean the same thing, which suggests their default voice is pulled toward the most common phrasing in their training data Do language models really understand meaning or just surface frequency?. That same pull shows up at the level of whole populations: when many people write with the same few models, expression converges, and co-writers unconsciously adopt the model's stances and framings Do large language models narrow human expression and thought?.

The most direct evidence on the style-versus-substance split comes from fiction. A detector using only discourse-level features (how much agency characters have, whether events unfold in chronological order) separated AI from human stories with 93% accuracy. Stripping out stylistic cues entirely barely hurt it, as it kept 97% of its performance Can AI stories be detected without analyzing writing style?. So at least in storytelling, the phrasing isn't doing most of the work. The deeper choices about what happens and how it's arranged carry the signal. They also hold up against 'humanizer' tools, because changing them means rewriting the story, not swapping words.

This matters beyond detection. Systems that judge text are easily swayed by surface features: LLM judges reward fake citations and attractive formatting no matter what the content says Can LLM judges be fooled by fake credentials and formatting?. Models trying to tell how individual people reason also lean on surface word cues instead of tracking the actual strategy underneath Can models recognize how individuals reason differently?. Even a model's own reasoning traces can have a tidy narrative structure that doesn't match what actually drove its answer Do reasoning traces actually show how models think?.

The takeaway: phrasing is the loudest signal but also the most disposable one. A paraphrase tool can remove it, and readers and judges can be fooled by it. The more durable signature of a model is in the shape of its content: which choices it defaults to, which framings it reaches for, and how it structures a narrative or argument. If you want to know whether a text came from a machine, the better question may be 'what did it choose to say?' rather than 'how did it say it?'