What an AI detection score can—and cannot—tell you

Why results differ

A result from one tool cannot predict an exact result from another. Detectors may be evaluated on different generators, domains, and test conditions; the ACL RAID benchmark compares detector behavior across those varied conditions (ACL RAID benchmark). Short passages may also fall below a provider’s minimums. Sapling currently recommends at least 300 characters and marks inputs under 500 characters as low confidence (Sapling detector documentation).

Length requirements vary by provider. Turnitin’s AI Writing Report requires at least 300 words of qualifying prose; a shorter excerpt is not a complete Turnitin report (Turnitin AI Writing Report guidance).

Can Turnitin detect ChatGPT?

Turnitin says its AI writing detection is designed to help educators identify writing that may have been prepared with generative AI tools, including chatbots such as ChatGPT (Turnitin AI detector description). In its AI Writing Report, the percentage describes the share of qualifying prose its model classifies as likely AI-generated or likely AI-generated and modified by an AI paraphrasing or bypasser tool (Turnitin AI Writing Report guidance).

The report describes a text classification, not a record naming which AI model produced a passage. This is an inference: a Turnitin result by itself does not prove that ChatGPT specifically was used or who wrote the text. Turnitin also says its model can misidentify human-written, AI-generated, and AI-paraphrased text, and that the result should not be the sole basis for adverse action. Under its current guidance, results above 0% and below 20% are shown as *% without an exact score or highlights (Turnitin AI Writing Report guidance).

What to do next

Use a score to prompt review. Check whether a passage accurately expresses your reasoning, whether evidence supports its claims, and whether you can explain the argument. Do not replace sound reasoning just to move a number.

Keep your notes, source material, and drafts. If your work is questioned, those records can help you explain your writing process. Maricopa Community Colleges advises students to check course rules and preserve relevant writing records (Academic Integrity for Students).

For a practical way to keep those stages, follow our guide to saving essay drafts and version history in Google Docs and Word.

If a report raises a concern, ask how it will be reviewed and what evidence or context is relevant. Turnitin says its AI result should not be the sole basis for an adverse decision; the institution’s process and human review matter. Keep an unchanged copy of the submitted work and the drafts that show how you developed it (Turnitin AI Writing Report guidance).

How to compare two detector scores

For differences between an AI estimate and a text match, see AI detectors versus plagiarism checkers. For two specific report workflows, see GPTZero versus Turnitin.

If you compare two detectors, use the same non-sensitive passage, language, and text length where both tools allow it. Record the exact input, date, tool/version and settings when disclosed, and what each score represents. Otherwise, you may be comparing different samples or different measurements.

That check shows how two tools responded to one passage; it does not tell you which tool is more accurate. Accuracy claims need a shared test set with known labels and results that account for both human-written text incorrectly flagged and AI-written text missed. The ACL RAID benchmark uses multiple generators, domains, and test conditions to evaluate detectors (ACL RAID benchmark).