AI Detector Robustness: Data and Methods from Our RAID Reanalysis
Full aggregate tables from Hi Violet's RAID reanalysis: detector accuracy under 11 attacks, per-document flip rates, per-domain false positives, methods.
Loading
AI research 5 posts.
Hi Violet research and field notes filed under “AI research.”
Can a Paraphraser Beat AI Detection? (Data, Not Hype)
Measured on 480,000 generations: paraphrasing makes one detector 5.4pp more accurate and another 17.0pp worse. Evasion is a detector lottery, not a strategy.
Where Detectors Break: The Domain Map
One pooled 5% setting, eight writing domains of truth: measured per-domain detector false-positive rates range 0.96% to 20.4%.
The Flip Rate: Why Detector Accuracy Hides the Collapse
Measured on 480,000 generations: one symbol substitution flips 93.7% of a detector's per-document verdicts. Aggregate accuracy hides the collapse.
You Cannot De-Watermark Prose: The Science Behind AI Text Provenance
Why synonym shuffling cannot guarantee clean provenance—plus RAID measurements of 11 attacks collapsing AI detectors, and Hi Violet's integrity protocol.
Bring the market problem into focus.
Want this kind of research run on your product? Tell us what you want to move and where it's stuck, and we'll reply by email.