Post #1802968

2026-04-09 11:18 UTC

@UlrikeHahn @pbloem Asking "do you?" upsets me. I'm not arguing that I'm smarter than an LLM, I'm questioning your point that "AI is operating in substantially different modes when it's hallucinating to when it's retrieving factual information." If I want a tool for retrieving and operating on facts, I would insist it have a trusted knowledge graph linked to sources! I'm biased, 'cuz I worked for years on Google's KG. But, seriously: I want a tool with verifiable results, and a trusted authority who's accountable for accuracy. I buy what you're saying about an "internal knowledge graph" and source tracing after the fact. I, too, have a fuzzy memory of where my ideas came from, and have to fact check myself sometimes! But these models are trained on fact, fiction, and falsehood, all blended together without judgment in the same "sorta KG." Unless the LLM is explicitly told to fact check itself, and which sources to trust, then it will occasionally make stuff up.

Replies (2)

@ngaylinn@tech.lgbt 2026-04-09 11:19

@UlrikeHahn @pbloem I suppose I'm saying recent models can do much better here, but that they still don't have a clear conception of fact, or the separation between factual and non-factual modes of communication. Maybe that could be added on, using RAG and the sort of research Ulrike is referencing for analyzing model uncertainty, but this is not currently common practice. Also, even if it becomes the norm, we would need to be very careful about who the arbiters of truth are here. These models are decidedly not neutral. Look at Grok and Grokipedia, for instance. There LLMs are being used to launder misinformation and establish an alternate history. :(

Open ##1802969
@pbloem@sigmoid.social 2026-04-09 11:29

@ngaylinn @UlrikeHahn Sorry if saying "do you?" came across as aggressive or combative. It wasn't meant that way. Perhaps I should have said "do we?" I simply wanted to make the comparison between LLMs and humans. In most of these discussions people paint the LLMs as missing something fundamentally necessary for producing facts. This is usually by contrast with humans. My point is that there are no _fundamental_ differences.

Open ##1802970