0XP
All concepts
Concept 08 of 090/3 mastered

Hallucination & Grounding

Why it invents things so convincingly

The model always produces the most plausible-sounding continuation. When it has the facts, plausible and true line up. When it does not, it still produces something plausible, in the same confident voice.

Why it invents things, so convincingly
Page 01 / 05

Why it invents things, so convincingly

A model does not know the difference between recalling something and producing something that sounds like the sort of thing it would recall. From the inside those feel identical, because mechanically they are identical.

Ask for a citation and the machinery does what it always does, which is produce the most plausible next fragments. A plausible citation has a plausible author, a plausible journal, a plausible year and a plausible title, so every piece of it is individually reasonable while the paper itself has never existed.

ONE SENTENCE, TWO KINDS OF TEXTThe policy was introduced in March 2019 and applies to all orders over £50.ACTUALLY LEARNEDCommon, widely repeated in trainingCONSTRUCTEDA plausible date, a plausible figure
Think of it like this

Like recounting a film you half-remember

You are fairly sure of the plot. You confidently name an actor. Some of it is right, some is another film entirely, and crucially you cannot feel the join. You are not lying. You have no internal marker separating the parts you actually remember from the parts your brain reconstructed to make the story hold together.

The word 'hallucination' is a bit misleading
It suggests a malfunction, the system going wrong in some special way. In fact nothing goes wrong at all: the model produces plausible text, exactly as it does when the answer is correct, and we only call it a hallucination on the occasions when that plausible text turns out to be false.