AI essay checkers: what the score on your draft does and does not mean

1What a checker measures

AI writing tools produce text by predicting likely words, as explained in how AI writing tools work. Checkers look for the statistical traces of that process: text that is unusually predictable, or patterns typical of model output. They return a score or highlighted passages, not a finding of fact.

2Why scores differ

Different checkers use different models, are trained on different examples and set different thresholds, so the same text can score very differently. Scores also change when tools update. Turnitin, one of the most widely used vendors, has published its own error rates: a document-level false positive rate of less than 1 per cent for documents with 20 per cent or more AI writing, and a sentence-level false positive rate of around 4 per cent, more common where human and AI text are mixed.

The same post advises instructors that there is no right or target score, to read the score alongside the institution's policy and the assignment rubric, and to use highlighted sentences to initiate a conversation, not to draw a conclusion. Some universities have chosen not to use detection at all: UCL's guidance says it is not investing in AI detection software and tells staff not to submit student work to such tools.

3If your own work is flagged

1. Stay calm and do not delete anything. Your files are evidence.
2. Gather your notes, outlines, drafts and version history, and any sources you used.
3. Be ready to explain the work: your argument, why you chose your sources, how the draft developed. That is usually what an investigation looks for.
4. If you used any AI in an allowed way, show your record and acknowledgement.
5. Read your university's misconduct procedure and ask for support from your students' union or advice service.

What else markers look at is covered in how teachers check work for AI. Prevention is simpler: the pre-submission checklist covers drafts and disclosure.

4What a score cannot tell anyone

Even a high score leaves the key questions open. It cannot say who typed the text, whether AI use was allowed in that assessment, whether a passage was a quotation, or whether the student edited a tool's suggestions heavily. Mixed documents are hardest: Turnitin notes that its sentence-level false positives occur more often where human and AI writing meet.

That is why policies treat a score as a reason to look closer, not as a finding. The evidence that settles a case is usually the student's process and their ability to discuss the work. For how essays are built honestly from the start, see drafting an essay with ChatGPT and thesis statements.

5Where to read accuracy studies

Independent tests of detectors vary widely by tool, text type and how the text was edited, and results go out of date as tools change. We do not restate benchmark scores here. For current comparisons, see the independent testing at humanrank.ai and aitruthbench.com. One related finding from research on AI text is worth knowing: Walters and Wilder note that no public detector they knew of checked for fabricated citations, which are a distinctive sign of generated text and something a marker can check by hand.

More on this in AI and writing.

6FAQ