After about a year of daily use, something shifts. You start catching AI-written text not by running it through a detector, but by feel - a rhythm that's slightly too consistent, a paragraph that ends with a sentence summarizing what the paragraph just said, transitions that land a beat too smoothly.
This pattern recognition is showing up across heavy ChatGPT users, and it's worth understanding what they're actually noticing.
The Specific Tells
The observations cluster around a few recurring patterns:
The wrap-up summary sentence. ChatGPT frequently closes a paragraph with a sentence that essentially restates the paragraph's main point. It reads as thorough. It also reads as redundant once you know what to look for.
Engineered transitions. Words like "moreover," "furthermore," and "it's worth noting" appear at a frequency that feels statistically off. Any single instance is fine. The density is what gives it away.
Three-beat paragraph rhythm. Claim. Supporting detail. Transitional close. Repeat. LLMs - the large language models behind ChatGPT and similar tools - predict likely text continuations based on patterns in their training data. That process tends to reproduce the most statistically common structures, and the three-beat paragraph is one of them.
Confident hedging. A strong declarative claim immediately softened by a qualifier in the next sentence. "X is clearly the case. Of course, results may vary." It's the AI equivalent of covering your bases.
The Editing Problem
What makes this more than just an observation is what people report about editing. Even after substantially rewriting AI output - changing vocabulary, reorganizing sentences - the structural fingerprint often survives. The words change but the skeleton stays.
This matters practically. If you're producing content with ChatGPT assistance and care whether it reads as human-written, swapping synonyms isn't enough. The structural patterns - paragraph rhythm, transition placement, summary sentences - need to go too. Break the three-beat structure. Cut the wrap-up sentence. Put a transition somewhere unexpected, or drop it entirely.
The same intuition also works in reverse. Heavy users report becoming more self-conscious about AI-adjacent patterns in their own writing - catching themselves writing the summary sentence, or reaching for "moreover" - and editing it out.
AI detection tools try to quantify this algorithmically, but human pattern recognition built through sustained daily use seems to be running ahead of them, particularly on content that's been edited. The detectors struggle with structural fingerprints. People who've spent a year inside ChatGPT's output apparently don't.