AI detector evidence
A false positive turns writing style into an accusation it cannot support alone.
An AI detector false positive occurs when human-written text is classified as likely generated. The risk is not merely technical: a score can trigger academic, editorial, or professional consequences if people treat it as proof.
Inference, not observation
The detector evaluates statistical features in the output; it does not witness the writer's research or revisions.
Unequal risk
Research has raised fairness concerns for non-native English writing and other styles that differ from training assumptions.
Evidence needs context
A fair review combines the text with drafts, sources, process records, policy, and the writer's explanation.
Why false positives happen
Detectors learn correlations between text features and examples labeled human or generated. Those correlations can shift as language models, editing tools, and human writing change. Short or formulaic passages provide less context, while translation and heavy revision can alter the signals again.
A model threshold also converts a continuous score into a practical label. Different providers may select different thresholds and update them over time, so the same passage can receive different classifications without any change to its real origin.
The fairness problem
Stanford researchers reported substantial misclassification of essays written by non-native English writers in the detectors they studied. The result is a warning against equating linguistic predictability with machine authorship or treating one population's style as the universal human baseline.
Consequences should match the quality of the evidence. If a school, editor, or employer uses a detector, it should disclose the tool and policy, preserve the report, allow a response, and consider alternative explanations before acting.
A better evidence hierarchy
Start with direct records of the work where available: dated notes, sources, drafts, version history, collaborator comments, and a writing-process receipt. Use detector output only as a prompt for review, never as a substitute for the review itself.
Process evidence has its own blind spots and should state them. HumanTyped can show what happened inside its editor and whether the sealed text matches that record; it cannot prove identity or exclude unobserved assistance. Honest limits make the combined evidence more useful, not less.
Primary sources
Check the current product and evidence record.
Evidence, not certainty