What makes ChatGPT text recognizable
Every large language model has a house style, and ChatGPT's is one of the most documented on the internet — mostly because it's the model the largest number of people have used to write things they didn't want to write themselves. That popularity is what makes detection tractable: teachers, editors, and hiring managers have now read enough ChatGPT output that certain patterns jump out before any tool gets involved. This detector scores text against those patterns and shows you which sentences are driving the score, so you can judge for yourself instead of taking a single number on faith.
The specific tells this tool looks for
- Rule-of-three structure
- ChatGPT defaults to grouping ideas in threes — three examples, three adjectives, three reasons — even when two would do or four would be more accurate. A paragraph that keeps landing on triads is a mild signal on its own and a strong one when it repeats across an entire document.
- Hedge-then-assert phrasing
- Constructions like "It's important to note that" or "While there are many factors to consider" show up disproportionately in ChatGPT output, usually right before a fairly ordinary claim.
- Overused transition words
- "Moreover," "furthermore," "in conclusion," and "overall" appear far more often in ChatGPT text than in typical human writing at the same reading level, especially stacked at the start of consecutive paragraphs.
- Leftover assistant framing
- Phrases like "as an AI language model" or "I don't have personal opinions, but" are the clearest tell of all when they survive into a final draft — a sign the text was pasted with little to no editing.
- Symmetrical sentence length
- Human writing tends to vary — a short sentence, then a longer one, then a fragment for emphasis. ChatGPT's default output is comparatively even, with fewer of the length swings that come naturally when a person writes quickly.
How the score is built
Paste text into the box above and the detector analyzes it sentence by sentence, weighing vocabulary choices, structural repetition, and phrasing against patterns associated with ChatGPT output specifically — not AI writing in general. You get an overall probability score and highlighting on the individual sentences contributing most to that score, so a single suspicious paragraph in an otherwise human document doesn't get lost in an aggregate number.
We don't publish an accuracy percentage, and you should be skeptical of any detector that does. Detection is inherently probabilistic: short passages, heavily edited AI text, and writers whose natural style happens to overlap with ChatGPT's patterns can all produce scores that don't match reality. Use the score as one input into a judgment call, not a verdict.
Who uses this, and what a good workflow looks like
- Teachers reviewing a submission that reads unusually polished for the student, using the sentence-level highlighting to start a conversation rather than issue an accusation.
- Editors checking freelance or guest submissions for unedited AI drafts before they enter a publishing pipeline.
- Hiring managers spot-checking cover letters or writing samples that feel generic.
In every case, treat a high score as a prompt to look closer, not as proof. Ask the writer about their process, compare against earlier work of theirs if you have it, and remember that a false positive has real consequences for the person on the other end.
If your own writing is scoring as AI-generated
It happens, particularly to non-native English speakers and to writers who lean on formal, structured prose by habit. If a detector flags your own work, the fix isn't to fight the tool — it's to add the texture that's missing: vary sentence length on purpose, cut a hedge phrase, replace one stacked list with a plain sentence. The ChatGPT Humanizer is built for exactly that kind of revision if you want a starting point.