What counts as AI-generated or AI-assisted text?
“Was AI involved?” can mean generation, proofreading, translation, restructuring, or a small suggestion in an otherwise human draft. Those are different histories. A binary label can obscure the actual policy question you need to answer.
If your goal is quality, check the work directly: facts, reasoning, sources, and relevance. If your goal is disclosure under a stated policy, establish which kinds of assistance must be disclosed. Avoid changing the question halfway through a review because a tool produced a striking-looking number.
AI classifiers vs watermark verification
A general writing classifier makes a model-based judgment about text patterns. A provider watermark check evaluates a particular signal using a suitable method. A Unicode scanner lists covered character patterns. These outputs should not be described as if they carry the same evidential weight.
Research on text detectors documents limitations across settings, and OpenAI discontinued its earlier public text classifier because of low accuracy. That does not mean every method is identical or useless; it means that claims require validation for the task, input type, and consequences involved.
Check citations, facts, and document history
Open citations and verify that they support the statements attached to them. Confirm names, dates, amounts, quotations, and calculations. Look for missing context or contradictions. These checks identify concrete problems regardless of whether the draft was produced by a person, a model, or both.
When appropriate, review drafts, revision history, source notes, and the author’s explanation. Use the normal process for obtaining those records and avoid assuming that a missing record proves wrongdoing. The goal is a fair account of the work, not a search for any detail that fits an initial suspicion.
Why punctuation and hidden characters are not proof
Fluent prose, formulaic transitions, em dashes, and repeated structure can appear in many kinds of writing. Hidden characters can arrive through ordinary software. None of these is a reliable standalone author identifier.
If you inspect text here, keep the result narrow: report the code points found and any technical problem they explain. Do not convert a count into a probability of AI use. A clean scan is equally limited and should not be called a human-authorship certificate.
How to review AI detection results fairly
For a consequential review, record the policy, the specific concern, the evidence examined, and the remaining uncertainty. Give the person an opportunity to respond to the actual issue. Separate factual errors, attribution problems, and undisclosed assistance rather than using one label for everything.
For everyday editing, concentrate on making the final text accurate and useful. Character cleanup can improve portability, while editorial review improves the content. Neither should be marketed as a way to transform an unknown history into verified authorship.
Sources & further reading
Primary references for the technical points in this guide.