How to tell if ChatGPT wrote an essay
The clearest way to tell if ChatGPT wrote an essay is to check how it was entered rather than how it reads: a single paste event covering 60-90% of the final word count, inter-keystroke gaps consistently below 150ms (a sign of transcribing rather than composing), and a 1,000-word essay finished in under 8 minutes are all objective, hard-to-fake indicators. Reading the prose for 'AI style' is far less reliable, since ChatGPT's phrasing changes with every model update and often flags non-native writers by mistake.
Why 'it reads like ChatGPT' isn't good evidence
Text-based AI checkers try to spot ChatGPT by scoring how statistically predictable the wording is. That signal erodes every time OpenAI ships a new model, and it regularly misfires on formal, simple, or non-native English, flagging human students as machine-generated. Several universities have restricted these tools in disciplinary proceedings specifically because of high false-positive rates.
The process signals ChatGPT essays actually leave behind
A student who generates an essay in ChatGPT and submits it typically pastes it in one or two actions, so 60-90% of the final word count arrives via paste rather than keystrokes. If any typing happens around the paste, it's often unusually uniform - inter-keystroke gaps under 150ms across a sustained run suggest transcription, not original thought. And the total session time is frequently implausible for the length: an 1,000-word essay wrapped up in under 8 minutes is a meaningful outlier for genuine drafting.
What to do when you spot these signals
No single signal proves ChatGPT was used - a student might paste from their own earlier notes, or draft in another app first. Treat two or more corroborating signals as a prompt for a conversation, not a verdict.
FAQ
- Can a text-based AI detector reliably catch ChatGPT essays?
- Not reliably. Text-based detectors estimate probability from prose style, and that accuracy degrades every time ChatGPT's underlying model changes. They also disproportionately flag non-native and simple writing as AI-generated, which is why several universities have limited their use in disciplinary decisions.
- What's the single strongest sign an essay came from ChatGPT?
- A large share of the final text - typically 60-90% - arriving in one or two paste events, combined with a total writing session far shorter than the word count would suggest for genuine drafting.
- Can a student disguise a ChatGPT essay by retyping it?
- Manually retyping suppresses the paste signal, but it often leaves unusually uniform keystroke timing behind - consistent gaps below 150ms are more typical of transcribing existing text than composing original thought.
Want to see this on real writing?