Education check
How to check an essay or assignment for AI
An AI detector cannot establish who wrote a student assignment. Use a score only to focus further review. An allegation or penalty requires the assignment policy, checkable process evidence, content verification and a fair opportunity for the student to explain.
Direct answer: check process and content
- Read the assignment policy first. Establish which AI use was allowed, prohibited or required to be declared. Without clear advance rules, a detector score cannot create a policy after submission.
- Check the input. Use the original PDF or DOCX and confirm headings, footnotes and paragraphs were extracted correctly. Do not treat a title, bibliography or short passage as representative of a full essay.
- Review process evidence. Ask for earlier drafts, notes, sources and revision history. Actively look for information that disconfirms AI use too.
- Discuss one content choice. Ask the student to explain an argument, source or revision. Apply the same procedure to comparable cases, then follow the existing academic-integrity process.
TEQSA advises educators to seek evidence that disconfirms AI use as well as evidence that may confirm it. Its guidance also identifies discussion of the content, earlier versions and writing-process evidence as relevant checks.
What a score does not prove
A high score is not the probability that AI wrote the assignment. Human writing can receive a strong signal, while edited or mixed AI text may receive little signal. OpenAI says its own detector research was not reliable enough for judgements with potentially lasting consequences for students.
Practical review order
- Preserve the original submission and record the detector and version.
- Check that citations and sources exist and support what the essay says.
- Compare relevant earlier versions without treating writing style as an identity test.
- Record the findings and provide a route for explanation or appeal.
Use the document detector for a complete PDF or DOCX, and the text detector for one coherent passage. Before acting, also review the detector's measured limits and false-warning rate.
Frequently asked questions
Can an AI detector prove a student used AI?
No. The output is a model-bound warning signal and is insufficient as the sole basis for an allegation or penalty.
What is stronger than one score?
The combination of advance rules, earlier drafts, sources, revision history and a conversation in which the student can explain their own choices.
Can you upload a student assignment?
Check the institution's privacy and approved-tools policy first. An assignment can contain personal data and unpublished student work.