GenAI Is Fluent in Everything, but Faithful in Nothing
I, Cringely, Monday, June 8th, 2026
LLMs produce plausible text but optimize for likelihood, not truth, so they cannot reliably know facts.
The author argues that language models predict probable word sequences from training-data patterns rather than possessing genuine knowledge. They cannot distinguish real from fabricated information because they have only ever seen words, not the world itself.
Truth requires correspondence between claims and reality, which is impossible for systems with no access to that reality. The proposed solution is to separate language generation from knowledge verification, keeping 'the saying and the knowing in separate rooms' and anchoring claims to verifiable sources rather than expecting machines to arbitrate truth from contradictory internet data.