Post #1506833

2026-04-17 10:20 UTC

I don&#39;t like the term &quot;hallucination&quot; in the context of LLMs for multiple reasons, but the top reason is it&#39;s unnecessarily anthropomorphising for something that sometimes needs to be said during technical discussion My suggestion is &quot;fault&quot;, inspired by &quot;page fault&quot;. What they have in common is that they are both unavoidable for deep technical reasons, and the best you can do is use various tricks to minimise them under normal conditions

Replies (3)

@BartoszMilewski@mathstodon.xyz 2026-04-17 13:47

@julesh But how is this a fault? The LLM provides the most plausible sounding answer to your question. It's painting a picture or writing a novel for you. You might not like it because you have a particular esthetic in mind. A novel is not wrong because it cites a non-existing paper to move forward the plot.

Open ##2103455
@bradheintz@mathstodon.xyz 2026-04-17 14:20

@julesh I dislike the term because it suggests derangement or an unexpected behavior, when non-factual output is not just expected, but provably unavoidable with current architectures. The word implies that a state of correctness is normal, when it is merely statistically more common than "hallucination”.

Open ##2103460
@cceckman@hachyderm.io 2026-04-18 13:25

@julesh Is there a counter that ticks over, that mechanically lets you know that it has happened, with 100% accuracy?

Open ##2103461