That would be true if children were abused specifically to obtain the training data. But what I'm talking about is using the data that already exists, taken from police investigations and other sources. Of course, it also requires victim's consent (as they grow old enough), as not everyone will agree to have materials of their abuse proliferate in any way.

Police has already used CSAM with victim's consent to better impersonate CSAM platform admins in investigative operations, leading to arrests of more child abusers and those sharing the materials around. While controversial, this came as a net benefit as it allowed to reduce the amount of avenues for CSAM sharing and the amount of people able to do so.

The case with AI is milder, as it requires minimum human interaction, so no one will need to re-watch the materials as long as victims are already identified. It's enough for the police to contact victims, get the agreement, and feed the data into AI without releasing the source. With enough data, AI could improve image and video generation, driving more watches away from real CSAM and reducing rates of abuse.

That is, if it works this way. There's a glaring research hole in this area, and I believe it is paramount to figure out if it helps. Then, we could decide whether to include already produced CSAM into the data, or if adult data is sufficient to make it good enough for the intended audience to make a switch.