“Negative outcomes” is very broad, so that terminology makes me suspicious. What if the LLM gives you accurate information, but you’re a retard and use that accurate information poorly? I do think that it’s obviously bad if the LLM is giving outright lies out, although given the way they hallucinate sometimes on basically any topic, I’m not convinced that that’s something the company can be blamed for so much as a ubiquitous, unavoidable flaw in the early versions of an emerging technology.

One of the scariest things about LLMs is how many people treat an obviously fallible, obviously open to bias algorithm as an arbiter of truth. I’ve seen people do things like respond to twitter posts about bias in a Wikipedia article with “@grok, is there bias in this article?” of course, Grok ingests loads of training data from Wikipedia itself and the various MSM sources they cite. There’s no guarantee Grok wouldn’t just have that same bias, and yet people assume it is an authoritative source. That’s an obvious issue that is worrying at a societal level, but how do you legislate for “there’s a really amazing technology with a lot of potential, but many people are dumb?”