LLMs are useful. You might not be able to train ChatGPT without stealing, but you can totally train a useful base model for task specific fine tuning without stealing.