Your definition of theft doesn't necessarily match the legal definition. It still remains to be seen whether training foundation models is fair use or not, and it's disingenuous to suggest otherwise.
Inference is where things get more muddled. If I download a pirated copy of La La Land, that's illegal. However, if I watch it and then explain the plot of the movie to you in detail, that's fair use. It remains an open question whether or not inference via an LLM represents fair use or not.