The proposed CSAM scanning used perceptual hashing to try to identify CSAM material known to law enforcement.

It was not a tool to identify private images as being underage. That’s an impossible task.

cortesoft 5 hours ago | parent | next [-]

It's impossible to do with perfect accuracy, but that doesn't mean it isn't done.

Just ask the dad who was investigated for taking pictures of his toddler for the doctor: https://www.koffellaw.com/blog/google-ai-technology-flags-da...

▲

trollbridge 5 hours ago | parent | prev [-]

I'd say modern AI tools could probably do this pretty effectively. They're very effective at describing anything else about an image. I have a workflow that churns through large amounts of images, describes them, and then looks for things I specifically want (in my case, auction listings that are not described accurately on the auction website).

	▲	philipkglass 5 hours ago \| parent [-]
		I need to process a modest amount of imagery (about 25 million images, and growing) for NSFW content and general captioning/description. About 5% of it contains nudity or partial nudity, and about 10% of that 5% contains sexual activity. In theory, modern vision language models could classify human nudity and sexual activity very thoroughly. But every model I have tried is reluctant to clearly describe what is notable about sexualized/nude images. The models are deliberately under-exposed to nude and sexualized content during training and further RLHF'd away from generating straightforward descriptions of such images. Models also occasionally hallucinate WTF captions for ordinary adult sexual activity. I recently ran a baseline test with frames extracted from adult videos and about 1/3000 frames was mis-captioned as involving a child according to Gemma 4 12b.