> If you can't tell when a big expensive llm is subliminally grooming you to like/dislike something or is selective with information

this is already there and in prod but called AI "safety" (really corporate brand safety). The largest LLMs have already been shown to favor certain political parties based on the preferences of the group doing the training as well. Even technical people who should know better naively trust the response of an LLM well enough to allow to make API calls on their behalf. What would prevent an LLM provider to train their model to learn and manipulate an API to favor them or a "trusted partner" in some way? It's just like in the early days, "it's on the Internet, it has to be true".