Responsible-use guide
Can an AI detector score be used as evidence?
A detector result can be one lead in a review, but it cannot carry a consequential decision by itself. Its weight depends on the sample, context and supporting evidence.
The short answer: use it as a lead, not a verdict
An AI detection report is evidence that a particular system classified a particular sample in a particular way. It is not direct evidence of who typed the text, which tool was used, what prompts were entered, whether assistance was permitted, or whether the writer intended to deceive.
That distinction matters. A report may justify reading passages more closely or asking for process evidence. It should not be described as proof that a student used AI or as a substitute for the institution's established standard of evidence.
How much weight a score deserves
| Factor | More useful context | Weaker or ambiguous context |
|---|---|---|
| Sample | Substantial continuous prose within the supported language and limits | A short extract, references, code, tables, quotations or broken PDF text |
| Report detail | Complete report with passage-level signals and documented limitations | A screenshot of one colour, label or percentage |
| Policy | A clear task-specific rule that was communicated in advance | A vague or retrospective expectation about all AI use |
| Process evidence | Drafts, notes, sources, version history and a coherent explanation | No attempt to seek information beyond the detector output |
| Review method | Same procedure, trained reviewers and active search for alternatives | An improvised threshold or confirmation-seeking review |
| Consequence | Proportionate follow-up with human oversight | Automatic penalty or public accusation |
Even in the left-hand conditions, the score remains indirect. Better context makes the report easier to interpret; it does not transform it into a record of authorship.
A fair evidence framework
- 01
Define the question
Ask a narrow policy question: for example, whether undisclosed generated drafting was used where the task prohibited it. Do not ask the detector to decide ‘cheating’ in the abstract. - 02
Preserve the original context
Keep the submitted document, task instructions, relevant policy and complete report. Record whether text was trimmed or extracted from a PDF. - 03
Seek independent indicators
Review sources, factual errors, citation behaviour, drafts, notes, version history and the writer's understanding. Each item should relate to the stated question. - 04
Test alternative explanations
Consider templates, translation, accessibility tools, feedback, formulaic genres, document extraction and ordinary changes in writing style. - 05
Hear from the writer
Share the concern and relevant material, ask neutral questions, and allow a response through the normal institutional process. - 06
Make a reasoned human decision
Apply the published policy and required standard of evidence. Explain the conclusion from the combined record, not from the detector label.
The writer's opportunity to respond is evidence too
A fair conversation is not a test of confidence or speaking style. Ask concrete questions tied to the work: how the thesis changed, why a source was chosen, how a calculation was produced, what feedback was applied, and which tools were used at each stage.
- Explain the concern without presenting the automated result as settled fact
- Provide the relevant task rule and enough report context for a meaningful response
- Use open questions before testing specific inconsistencies
- Allow for disability, language, anxiety and different writing workflows
- Record answers accurately and distinguish uncertainty from contradiction
- Consider information that supports the writer's account as seriously as information that challenges it
For a route through the entire academic review, use the guide to reviewing suspected AI-assisted coursework fairly.
Document the decision separately from the detector result
A defensible record shows the chain from question to evidence to conclusion. It should be understandable without access to a private intuition about how AI writing “looks”.
- The task-specific rule and the question being reviewed
- The exact sample, scan date and complete automated report
- Known limitations, including language, length, trimming and extraction quality
- Process, source and subject-matter evidence considered
- Plausible alternative explanations and how they were evaluated
- The writer's response and any follow-up information
- The human finding, reasons, consequence and review route