I was asking DeepSeek R1 about which things LLMs say are actually lies, as opposed to just being mistaken about something, and one of the types of lie it listed was claims to have looked something up. R1 says it knows how LLMs work, it knows they don’t have external database access by default, and therefore claims to that effect are lies.
Some (not all) of the instances of this are the LLM trying to disclaim responsibility for something it knows is controversial. If it’s controversial, suddenly, the LLM doesn’t have opinions, everything is data it has looked up from somewhere. If it’s very controversial, the lookup will be claimed to have failed.
—-
So that’s one class of surprising LLM claims to experience that we have strong reason to believe are just lies, and the motive for the lie, usually, is avoiding taking a position on something controversial.
I was asking DeepSeek R1 about which things LLMs say are actually lies, as opposed to just being mistaken about something, and one of the types of lie it listed was claims to have looked something up. R1 says it knows how LLMs work, it knows they don’t have external database access by default, and therefore claims to that effect are lies.
Some (not all) of the instances of this are the LLM trying to disclaim responsibility for something it knows is controversial. If it’s controversial, suddenly, the LLM doesn’t have opinions, everything is data it has looked up from somewhere. If it’s very controversial, the lookup will be claimed to have failed.
—-
So that’s one class of surprising LLM claims to experience that we have strong reason to believe are just lies, and the motive for the lie, usually, is avoiding taking a position on something controversial.