Plausible Is Not the Same as True | Visual AI Learning Guide

So the context keeps narrowing the options. But narrowed toward what? Toward whatever continues the text most plausibly, not toward whatever is true.

0:00 / 0:00

Auto-play

The model's objective at every step is to score how likely a token is to continue the text, not to check whether the resulting statement is true. A sentence can therefore be assembled from tokens that each fit the context perfectly while the claim they form is false. Fluency and factual accuracy are produced by different things: fluency comes from the next-token mechanism, accuracy would require information the mechanism does not verify.