A false negative is when a detector misses AI-written text and scores it as human. Lightly edited or paraphrased model output is the usual cause — small changes can be enough to drop a statistical detector's confidence below the threshold it uses to flag a passage.
Why this matters more than the false-positive case
Most discussion of detector errors focuses on false positives, because they hurt honest writers and feel unjust. False negatives are quieter but more interesting in academic settings, because they create the illusion that a workflow is fine when it isn't. A student whose laundered essay scores low on the detector concludes that the laundering worked; the examiner reading the essay forms a separate, less forgiving impression.
What humans notice that detectors miss
- Argument shape. An examiner can tell when paragraphs don't actually build on each other — when the conclusion doesn't follow from the body. Detectors can't see this.
- Source engagement. An examiner can tell when a citation has been name-dropped without the source being understood. Detectors can't.
- Voice mismatch with the rest of your work. If your essays usually sound one way and this one sounds different, the marker will notice — even if no detector flags it.
- Suspiciously even mastery. Genuine writing has the seams of the writer's actual learning curve. AI-edited writing is too smooth.
The harder question
If your goal is to submit work you can defend, the detector score is a tool, not a verdict. A low score doesn't certify a passage is honest, and a high score doesn't prove it isn't. The honest test is whether you could walk a sceptical marker through your essay sentence by sentence and stand behind every claim.