What a detector score actually means
Every detector in this directory publishes an accuracy figure. None of those figures answers the question people actually have, which is: if this tool flags my document, how likely is it to be wrong?
Accuracy is the wrong number
"99% accurate" usually means: across a test set the vendor assembled, 99% of documents were classified correctly. Two problems. The vendor chose the test set, and accuracy blends together two errors that matter very differently - missing AI text, and falsely accusing a human.
The number you want is the false positive rate, and then something the detector cannot know: how much human writing is in the pile it is being pointed at. Run a detector with a 1% false positive rate over 2,000 genuinely human essays and it will flag roughly 20 innocent people. That is not the detector malfunctioning. That is the detector working as advertised.
Who gets falsely flagged
Not randomly. Detectors measure how predictable text is, and predictable writing is produced by a lot of people who did nothing wrong:
- Non-native English writers, who use a smaller, more common vocabulary and simpler sentence structures. This is the best-documented bias in the field.
- Technical and formulaic writing, where the conventions exist precisely to make text predictable.
- Heavily edited text, including anything run through a grammar checker.
- Short documents. Most detectors need several hundred words before their output means much at all, and many will still return a confident-looking percentage for a paragraph.
What the percentage is not
A detector returning "87% AI" is not saying there is an 87% chance a model wrote it, and it is not saying 87% of the document is machine-written. It is a classifier confidence score on a scale the vendor chose. When two detectors both return 87%, they are not agreeing about anything in particular.
What a fair process does with a flag
- Treats it as a prompt to look, never as a finding.
- Looks for corroboration a human can evaluate: draft history, version history, the student's or writer's ability to discuss their own work.
- Tells the person the result and lets them respond before any conclusion.
- Does not run one detector, get the answer it expected, and stop.
- Records that detectors have a known false positive rate, in writing, in the policy.
If you are wrongly accused
Ask which tool was used and what score it returned. Ask what the institution's policy says about detector evidence. Offer what a detector cannot fake: version history, drafts, notes, the ability to talk through your argument. Point to the vendor's own documentation, which in almost every case says the output should not be used as sole evidence.
Why we still list detectors
Because they are used, and a directory that only listed the tools for evading them would be taking a side in a fight it is meant to be documenting. What we can do is record each accuracy claim as a claim, attach this page to every detector listing, and never write the sentence "accurate detector" as though it were a property of the software.