How to Check if Text Was Written by ChatGPT
Teachers, editors and hiring managers all ask the same thing: was this written by ChatGPT or a person? There is no certain answer, but a careful reading combined with a statistical check gives a reasonable picture. Here is how it works and how to avoid unfair conclusions.
What AI text detectors measure
- Burstiness: people mix long and short sentences. Language models tend to produce even, similar-length sentences.
- Predictability: model output uses the most likely next word more often than a human would.
- Vocabulary diversity: AI text can repeat the same safe words and transitions.
- Stock phrasing: phrases such as "in today's fast-paced world", "it is important to note" and "delve into" appear much more often in model output.
Signs you can spot yourself
- Every paragraph has the same shape: claim, three points, tidy conclusion.
- It is fluent but vague, with no specific example, name, date or personal detail.
- It hedges on everything and never commits to an opinion.
- Facts or quotations are confident but cannot be verified. Invented citations are a classic sign.
- The voice differs sharply from the author's earlier writing.
Human tells that lower the odds
Typos, sentence fragments, local slang, strong opinions, specific memories and small inconsistencies all point towards a human author. Modern tools can imitate some of these, so none is absolute.
Run a quick check
Paste the passage into the AI text detector. It shows the individual signals behind its estimate so you can judge them yourself. Longer samples (at least a few hundred words) are far more reliable than a single paragraph.
Why you should never treat a score as proof
Detectors produce false positives. Formal writing, text from non-native English speakers and heavily edited drafts are often flagged wrongly. Our article Can AI text detectors be trusted? explains the limits. A fair process is:
- Use the score only to start a conversation.
- Ask the author to explain their sources and reasoning, or to write a short sample.
- Compare with earlier work, drafts and version history.
- Make decisions on several kinds of evidence, never on one number.
If you are accused unfairly
Keep drafts, notes and document version history. Those show your process far better than any detector can. Ask which evidence was used, and point out that these tools are not reliable on their own.