Effective CSAM filters are impossible because what CSAM is depends on context
rys.io
rys.io
The hesitation to effectively communicate after CSAM may actually cost lives.
This article is essentially referring only to edge cases which require context to determine if images are CSAM. There are other cases where context is not needed and determining if something is CSAM is sadly easy.
For example, there is no appropriate context for images or videos of adults raping toddlers or infants.
unlike "CSAM," "radioactive" implies only content - content that has to be deliberately mined and enriched in a handful of factories, not something that can be accidentally generated by billions of families.
also, material can become CSAM without starting off as CSAM. imagine a baby monitor captures video of a toddler changing clothes. that's not CSAM yet - it's merely indecent. but now imagine a pedophile hacks the feed and shares it on a CP website. now it absolutely is, because intent (sexual gratification) and abuse (stealing and sharing the photos) have occurred.
we need nuanced labels to implement a nuanced policy.
The solution to have human-in-the-loop appeals processes. And because companies don't want to pay for this, it should be enforced as regulation.
It's the invasion of privacy not accuracy rates of the model
One is that it has false positive and negatives, others, like you mentioned involve privacy. Some people even feel like it's an invasion of privacy and then justify their argument by saying that it has false positives and negatives.
Is it csam? I doubt it, those photos are for me and my wife who had to put up with it.
He finally wears underwear now.
Citation please. Once it is explained to them how it could work, what the failure cases are, what false positives cause (with real life examples, since it has already happened), etc, I doubt this would be even remotely true.