6 comments

[ 3.4 ms ] story [ 22.6 ms ] thread
Is this claiming to detect 99% of CSAM or claiming to distinguish between CSAM and non-CSAM with 99% accuracy?

Relevant: https://bohemian.ai/blog/99-accuracy-99-lie/

> Thorn, the non-profit behind Safer, says it spots the content with greater than 99% precision.

Sounds like it's claiming 99% precision, so 99% of positives are true positives.

Recall probably isn't great, because (a) if it were great they'd brag about it and (b) you don't get to 99% precision without sacrificing some recall.

Seems reasonable to aim high on precision, though, to avoid burying NCMEC in mis-classified data and also to avoid wrongly banning users' content.

I am a bit confused; I may have missed it, but I don’t see a mention of Project Arachnid, which is similar based on the description.

https://projectarachnid.ca/en/

This is great, and I am going to use it in my side project. The main question I have is who has responsibility for the the 1% of CSAM not detected? Myself, as the service inadvertently hosting the material, or Thorn for not having an "accurate enough" algorithm? I suppose they just put it in their SLA terms or something.