Why a Detector Report Can’t Prove You Wrote It — and What Actually Works

An inference from documented user reports: people who overturned false accusations did not win with a report from a different detector. They could not have — every detector can only estimate. “Tool A says 90% AI” and “Tool B says 80% human” are two estimates disagreeing, and to whoever judges the case, that proves detectors disagree, not that you are innocent.

The same reports show what did work: drafts, version history, search history, notes. These demonstrate that the document grew over time, and they are hard to fabricate after the fact.

University guidance points the same way: checker output is increasingly treated as reference information to be combined with draft history, citation consistency, and an interview — not as grounds on its own.

If you need to explain the limits of detection: Turnitin qualifies its own accuracy (document-level FPR under 1% only above 20% AI content); a Stanford-affiliated study found ~61% misclassification on non-native English essays (specific to that setting); Vanderbilt retired Turnitin’s AI detection. “Detection is an estimate” is not our claim alone — the technology’s own providers say it.

We do not sell proof of innocence — no one can sell that. What we offer is a transparent reading of what triggered the flag, guidance for preserving your process evidence, and honest material for explaining what detection can and cannot say.