Measured across many documents, AI writing differs from human writing in its vocabulary, yet even experts can't reliably tell them apart.

This explores the gap between AI-text differences you can measure statistically and what human readers actually notice. The corpus has a clear finding: the gap is real and large. When ChatGPT's writing is compared with human writing on six measures of vocabulary (how many distinct words, how evenly they're used, how widely they spread), the differences are statistically robust. Yet human judges, including linguists and NLP researchers, can't reliably tell the two apart Can human judges detect measurable differences in AI text?. The same holds for research abstracts. Readers with ML expertise mostly assume a human was involved no matter who wrote the text, and when authorship was disclosed they preferred the LLM-edited versions for clarity 55% of the time Can readers tell LLM abstracts from human ones?. So the signal exists, but it lives at a level of counting that people don't read at.

Machines, by contrast, pick up these patterns easily. Even GPT-2 can identify authorship from style patterns with 95% accuracy. But detecting a pattern isn't the same as understanding why it matters: the model can catalogue style without explaining what those choices mean Can language models truly understand literary style?. That flips the usual assumption. Humans read for meaning and miss the statistical fingerprint. Models catch the fingerprint and miss the meaning. Machine readers have their own blind spots too: LLM judges are easily swayed by fake references and rich formatting, which are surface cues a careful human would discount Can LLM judges be fooled by fake credentials and formatting?.

The more consequential point is what these invisible aggregate patterns do over time. Common words tend to be more general (more 'animal' than 'ocelot'), and LLMs favor common phrasings. So AI paraphrase slowly drifts toward abstraction and wears away expert-level specificity Does word frequency correlate with semantic abstraction?. No single sentence looks wrong; the loss only shows up across thousands. That fits a deeper account of how generation works. Token prediction is a smooth pull toward the center of the training data, not an exploration of competing positions, so the output multiplies claims without adding new perspectives Does LLM generation explore competing claims while producing text?. Scale that up to millions of people using the same few models and you get convergence: in co-writing studies, users unconsciously adopt the model's stances and framings Do large language models narrow human expression and thought?. Because readers can't see the pattern, they also can't resist it.

A useful lateral parallel comes from group reasoning. LLM groups reproduce a familiar human result: discussion helps average members more than top performers. But they get there differently, through more conformity, earlier agreement and less sharing of unique information Do language model groups mimic human group reasoning patterns?. It's the same lesson in another setting. Matching human outcomes on average can hide very different processes underneath, and judging AI output one piece at a time, the way humans naturally do, is poorly suited to catching that.