ChatGPT output tends to share a handful of recognizable writing patterns. Here is what detectors look at, and what to do if your text gets flagged.
The five specific tells
AI detectors do not read for meaning. They measure writing signals, and ChatGPT output tends to show the same five signals over and over.
1. Hedging density
ChatGPT over-uses transitional and hedging phrases: 'while,' 'however,' 'moreover,' 'it is worth noting that,' 'in essence,' 'importantly.' Human writers use these too, but usually less often. Detectors compare the rate of these phrases against typical human and AI baselines.
2. Em-dash frequency
ChatGPT, especially since GPT-4, uses em dashes far more often than most human writers do. Not every human writer leans on em dashes like this, which makes density a useful signal. If your output has em dashes in every paragraph, that pattern alone can raise an AI score.
3. Tier-1 AI vocabulary
ChatGPT reaches for a specific vocabulary band more often than humans: 'multifaceted,' 'leverage,' 'paramount,' 'ensure,' 'facilitate,' 'navigate,' 'robust,' 'comprehensive,' 'strategic.' Detectors keep lists of stock AI vocabulary and count occurrences. High density flags heavily.
4. Uniform sentence length
ChatGPT writes medium-length sentences at high consistency. Human writing alternates more aggressively: short punchy sentences next to long complex ones. That variation, often called burstiness, tends to be much lower in ChatGPT output than in human writing at the same register. Combined with predictability measures, it is the backbone of most detectors.
5. The "not X but Y" construction
ChatGPT loves structural oppositions: 'this is not about speed but about depth,' 'not only efficient but also ethical.' A few of these per page is normal human writing. One per paragraph, which is typical of ChatGPT output, is a fingerprint.
What changes across GPT versions
GPT-3.5 has the most obvious tells: heavy hedging, stock transitions, rigid structure. GPT-4 reduced the obvious ones but kept the statistical signature. GPT-4o softened rhythm but still uses stock AI vocabulary at AI-typical rates. GPT-5 continues the trend toward more human-like output while still being statistically distinguishable from human writing.The practical takeaway: newer ChatGPT versions are harder to spot by eye, but detectors updated after each release still pick up the underlying patterns. A human editor who finds your text 'reads great' is not running the same test a detector runs.
Why paraphrasing ChatGPT output usually fails
A standard paraphrase, swapping synonyms and rearranging clauses, leaves the underlying distribution intact. Predictability stays low, sentence length stays uniform, and stock vocabulary gets replaced with more stock vocabulary. The text reads differently but scores nearly the same.A real cleanup has to change rhythm, break parallel structure, and replace stock phrases. Leap's free humanizer runs in your browser and does a cleanup pass: it replaces stock AI phrases, removes em dashes and invisible characters, adds contractions, and marks sentences of uniform length so you can vary them. It does not guarantee passing any detector, but it targets exactly the signals described above.
What to do if your writing got flagged
If you wrote it yourself and it got flagged, check whether your natural writing has unusually uniform sentence length, low vocabulary variety, or heavy hedging. These can make original human writing score as AI. False positives also cluster around topic terminology that overlaps with common AI phrasing.Research from Stanford found that detectors flagged non-native English essays as AI-written at far higher rates than native-English writing, so if English is not your first language, context matters more than any single score. A detector score is a signal, not proof, and it should never be the only basis for an accusation.If you edited ChatGPT output lightly, detectors can still catch the underlying signature. Either rewrite at the structural level or run the text through a cleanup tool and check the score again before submitting. Focus on rewrites that change rhythm and structure, not cosmetic word swaps.
Why this matters beyond academic settings
Flagging matters in contexts students often miss. Journals like Nature have updated their AI authorship policies to require disclosure of LLM assistance, and COPE, the publishing ethics body, has guidance for authors and reviewers that mirrors it. The reputational cost of a silent flag after publication is higher than the cost of disclosure at submission.If you write in an academic or publishing context where policies apply, think about disclosure and editing standards first, not only about detection scores.