Is a screenshot of an AI detector result worth anything as evidence?

I was shown a screenshot once, by a teacher, on a phone. I could see a number and a red bar and nothing else. I was not allowed to keep it or see the text that was highlighted. I did not know how to challenge a picture. If you write guidance, please say that the student must be shown the same material the instructor saw, in a form they can study.

@Aisha R. noted, and thank you. Giving the student the same material in a form they can study will be an explicit requirement in the draft, not a courtesy.

A screenshot proves a number appeared on a screen. That is all. Anyone can paste anything into a free checker and screenshot the result. I could screenshot a passage of a newspaper and get a high score if I tried the right ones.

It should count for nothing without the full document and the full report.

When I was on a student panel last year, the screenshot was never the thing that decided it. It was a starting point for the conversation. What decided the case was the version history and the fact that the student could explain every paragraph.

I would add one line to your guidance: instructors should write down, at the time, what they observed and what they did next. Otherwise six weeks later nobody remembers.

From a pure evidence-handling view, a screenshot is the weakest form of record: no provenance, no integrity check, no reproducibility. If you want something that stands up, keep the original file the tool analysed, the tool's own exported report, the date, and the product version. Then someone else can rerun it, which is the real test.

I would also note that the rerun may give a different number. That is a feature of the exercise, not a bug in your process.

In a school I do not get panels, I get a parent and a principal. The screenshot is the whole case in their eyes, which is the problem. If your guidance lets a score through as evidence, it will be treated as proof within a week. I would put the limits in bold.

A screenshot shows only that a tool displayed a number. It does not establish who wrote the text, and the tool vendors themselves say the number should not decide a case alone. If you want something that holds up, ask for the complete report, the submitted document, and corroborating evidence about how it was written.

Turnitin's guidance on reviewing the AI writing report says the score is not meant to provide definitive answers in isolation and should be treated as a single data point alongside the educator's knowledge of the student and institutional policy. If the maker of the tool says that, an image of its output cannot carry more weight than the tool does.

The research explains why caution is warranted. The Stanford study GPT detectors are biased against non-native English writers tested seven widely used detectors and found they misclassified more than half of a set of TOEFL essays written by non-native speakers as AI generated, with an average false positive rate of 61.22%, while near perfect on US eighth grade essays. It also found that a simple follow-up prompt could greatly reduce detection of AI generated essays. That is a small pilot study on older detectors, so treat the numbers as illustrative, but the direction matches what panels see: the same tool can be wrong in both directions.

For your guidance: require the full report as a file, the submitted document, the date, and the tool and version; give the student the same material; and state that a score alone does not meet the standard of proof. The thread on what Turnitin shows instructors is a useful companion for instructors.

@Maya Lindqvist thank you. The "if the maker says that" point is the line I will use with colleagues who think a screenshot is stronger than the report behind it. I will draft the attachment list as you describe: file, report, date, version, and the same package to the student.