I appreciate such tests & their visibility, but this is a predictable failure of LLMs, which have no intent, so framing it as "lying" misleads readers.

I believe journalists should follow the advice of @emilymbender.bsky.social & @nannainie.bsky.social: www.techpolicy.press/we-need-to-talk-about-how-we-talk-about-ai/

February 18, 2026 - 19:14 UTC

2
0
19
The "hyperparameters" that define the data associations in an LLM are calibrated (at some level, often poorly) by humans with intent.
0
0
0
Um is the information the chatbot returns the truth, or some kind of other thing?
4
0
2
Some kind of other thing. It is output based on its inputted training data, so pretty much like a fancy version of the autocomplete on a mobile phone (before that became LLM-led).

If a topic is widely factual in the training data, the output is more likely to be factual.
0
0
0
they're nondeterministic and just generate the most likely content for previous context, so it could give you the truth one time or hallucinate another because there's no real difference to it
1
0
6
It returns an answer-shaped object.
1
0
25
It is not a truth or even a falsehood, because to be a falsehood something must be possible to be truth, and truth requires intent.

Throwing scramble tiles on the ground cannot result in truth, even if the statements appear to be factually correct.
1
0
4