Originally published September 4, 2023. Updated October 15, 2026. This documentation-only revision revisits OpenAI’s 2023 educator guidance and checks its current help documentation. We did not test detectors or assess student work.

Cover image: illustration retained from this article’s 2023 publication for historical context. It does not show a detector test or a real authorship finding.

Asking ChatGPT whether it wrote an essay cannot establish who wrote it. OpenAI’s educator FAQ, checked for this update, warns that ChatGPT can invent answers to questions about whether content was generated by AI. A confident reply is therefore unsuitable as evidence of authorship.

The 2023 context

On August 31, 2023, OpenAI published Teaching with AI, introducing classroom examples, suggested prompts and a companion educator FAQ. Its discussion covered model limitations, detector reliability and bias. This GPTOnline article originally appeared four days later.

The teaching guide also described an educator encouraging students to verify ChatGPT’s answers against primary sources. That approach supports an important distinction: checking a statement’s accuracy is a different task from determining who wrote it.

The retired classifier

OpenAI had introduced a separate AI text classifier on January 31, 2023. Its original announcement now records that the tool was withdrawn on July 20, 2023, because of poor accuracy. This was a dedicated classifier, separate from asking ChatGPT to judge a passage.

In the English-text evaluation reported at launch, it identified 26% of AI-written texts as likely AI-written and incorrectly flagged 9% of human-written texts. OpenAI also warned about short inputs and advised against using the classifier as the primary basis for decisions.

Those figures belong to that historical model and evaluation. They do not measure the performance of every detector available today, and they are not a prediction for a particular assignment.

What the FAQ says now

OpenAI’s current educator guidance describes false positives in its own detector research and says that research did not establish sufficient reliability for consequential judgments about students. It expressly declines to assess other developers’ tools.

The FAQ also discusses documenting AI use, citing sources and reviewing how students question or check an AI response. These are ways to discuss learning and the writing process; a conversation record alone does not prove how an entire assignment was produced.

A classroom process

The following is GPTOnline’s suggested planning checklist, not an OpenAI protocol or a tested method for identifying AI writing. Adapt it to the assignment and the institution’s approved procedures:

  1. Make expectations concrete. Before work begins, explain whether AI may be used for brainstorming, language support or drafting, and what disclosure is required. Give an example of an acceptable disclosure.
  2. Separate the questions. Review factual accuracy, source attribution and compliance with the assignment’s AI rules separately. Permitted, disclosed AI assistance should not automatically be treated as plagiarism.
  3. Discuss the reasoning. Invite the writer to explain a claim, the source behind it and a revision they made. Use this to understand the work, rather than treating fluency or hesitation as an authorship test.
  4. Keep the discussion private. Use the institution’s approved channels. Do not ask students to publish full chat histories or upload identifiable work to an outside service for this exercise.
  5. Keep conclusions narrow. A missing citation can be corrected as a citation problem. A detector label or chatbot answer should not become a finding of misconduct. Any formal concern belongs in the institution’s established review process.

A fictional example

Imagine an assignment comparing two arguments about a public park. The rules allow brainstorming with AI but require the submitted argument to be the writer’s own work, with sources cited. This example is hypothetical and has not been tested in a classroom.

A submitted paragraph contains a claim about tree cover without a source. A useful next step is to ask where the claim came from and what evidence supports it. The writer can locate a suitable source, revise an overstated claim or remove a claim they cannot support. If AI was used, they can explain that use under the assignment’s disclosure rules.

The intended outcome is a better-supported paragraph and a clearer account of the process. The exercise does not establish authorship or determine whether a student violated a policy.

Reading detector reports

For context on a current third-party service, our GPTZero guide explains its reports, limitations and privacy considerations. That separate guide is a documentation review, not an endorsement or a comparative accuracy test. Keep the 2023 OpenAI classifier’s results separate from any present-day product claim.