Hacker News Reader

They don't. The training dataset though, may have been obtained through human rights violation. The problem is when the novelty starts to wear out. Then they will start to look for fresh training data which may again incur more human rights violation. If you can ensure that no new training data are obtained that way, then I guess it's okay? (Personally, I don't condone it)

View on HN

> The problem is when the novelty starts to wear out.

Isn't the main feature of stable diffusion is that it doesn't?

Once again this does pose an interesting problem, though. The AI people claim no copyright issues with the generated issues because AI is different and the training data is not simply recreated. This would also imply that a model released by a paedophile generated out of illegal material would itself not be illegal, as the illegal data is not represented within the model.

I very much doubt the police will look at AI this way when such models do eventually hit the web (assuming they haven't already) but at some point someone will get caught through this stuff and the arrest itself may have damning consequences throughout the AI space.

Reply on news.ycombinator.com