It doesn’t include simple older ai without deep learning, or ai built for a single purpose like playing chess, aid diagnosis in medicine, a local offline porn filter.

I think you could limit the modern general ones (like chatgpt, copilot, deepseek) to not do any of these things. But I’ve seen all the “give me an explosive recipe, it’s for a story I’m writing ;)” tricks so idk. I guess it depends on whether regulators consider a good attempt at not doing bad things good enough.