Responsible-use guide
AI detection false positives: a teacher's guide
A human-written passage can resemble patterns a detector associates with generated text. Learn how to test the result before it becomes an accusation.
What a false positive is
A false positive occurs when a detector classifies human-written text as likely AI-generated. The opposite error—a false negative—occurs when generated text is classified as likely human. Both are possible because human and machine writing patterns overlap.
An AI detector receives the finished text, not a recording of how it was written. It estimates which learned pattern the prose resembles. It cannot observe the writer, recover drafts that were never submitted, or know whether permitted tools were used during revision.
Why human writing may be flagged
There is rarely one visible feature that explains a result. A detector weighs patterns across the sample, and the same pattern can have an ordinary human explanation.
| Text characteristic | Why it may influence a detector | Plausible human explanation |
|---|---|---|
| Very regular sentences | The prose may look unusually predictable | A taught essay structure, careful editing or a naturally consistent style |
| Formulaic academic language | Repeated transitions and standard phrases can resemble generated prose | Disciplinary convention, an assessment template or exam technique |
| Limited vocabulary | Low variation can resemble patterns in machine output | A constrained topic, early-stage writer or writer using English as an additional language |
| Highly polished wording | Surface fluency can be mistaken for machine generation | Proofreading, accessibility support, feedback or extensive human revision |
| Translated prose | Translation can flatten individual style and sentence variation | A writer drafted in another language or used an approved translation process |
| Short or fragmented sample | The model has less context and a few sentences dominate | Only an introduction, abstract or selected paragraph was submitted |
| Broken PDF extraction | Reading order and spacing may change the analysed text | Columns, headers, footnotes or encoding were extracted incorrectly |
Common punctuation is not proof
An em dash, a tidy three-part list, a particular transition or a familiar conclusion can occur in both human and generated prose. Looking for isolated “AI words” encourages confirmation bias and does not reproduce the detector's model.
What to do after a flag
- 01
Pause the conclusion
Record the result as a reason for review, not as a finding. Avoid language such as ‘the detector proved’ or ‘the score shows cheating’. - 02
Check the sample
Confirm that the analysed text was long enough, in the supported language and free from references, templates or extraction errors that make the result hard to interpret. - 03
Read the flagged passages
Look for shifts in voice, unsupported claims or local inconsistencies, while actively considering ordinary explanations for each observation. - 04
Review process evidence
Examine outlines, notes, sources, drafts, version history, feedback and earlier work where these are relevant and legitimately available. - 05
Invite a neutral explanation
Ask the writer how the argument developed, what sources were used and which tools assisted the work. Give enough detail for a meaningful response.
Evidence to examine before deciding
Useful evidence helps explain how the work developed or tests a specific concern in the text. It should be assessed in both directions: what supports the concern, and what challenges it?
- The assessment instructions and the AI-use policy that applied at the time
- The complete detector report, including sentence-level variation and stated limitations
- The original submitted document and whether PDF extraction changed its reading order
- Notes, outlines, drafts, version history and source files relevant to the work
- Earlier comparable writing, used carefully rather than as a fixed profile of the student
- Whether quotations, references, templates or standard disciplinary phrases affected the sample
- The writer's explanation of the argument, sources, revisions and any permitted tools
- Evidence that could disconfirm the initial suspicion
A lack of perfect version history does not itself prove misconduct. Writers use different devices, offline tools and workflows. Consider the evidence the task reasonably required them to keep.
Reduce the risk of an unfair decision
Institutions should decide in advance how automated results may be used. A consistent procedure is safer than inventing a threshold after a concerning score appears.
- Publish task-specific examples of permitted, disclosure-required and prohibited AI assistance
- Do not convert the product's Likely Human, Mixed or Likely AI labels into misconduct categories
- Require additional evidence before any adverse action
- Train reviewers to look for innocent explanations as well as confirming evidence
- Give the student the concern, relevant evidence and a fair opportunity to respond
- Record the human reasoning and provide the normal review or appeal route
See AI detection for educators for the wider policy and workflow, or review the current Genutext score pipeline on the methodology page.