Source: David Pogue
Large language models are architecturally incapable of outputting uncertainty—they're trained to generate the next token with the highest probability, not to flag confidence gaps or abstain from answering. This creates a failure mode in professional contexts where stakes are real: an executive assistant getting wrong job titles for a conference presentation, or a lawyer citing fabricated case law, suffer not from occasional errors but from systems that confidently hallucinate rather than defaulting to honest ignorance. The fix isn't just better training; it requires redesign of how these models interface with users, potentially including explicit refusal mechanisms or confidence scoring that actually shapes output rather than appearing in afterthought disclaimers.