Why does AI lie? Hallucinations explained simply

Short summary

AI hallucinations occur when models confidently generate false information to fill knowledge gaps, fabricating plausible-sounding answers instead of admitting uncertainty. The post demonstrates this through model comparisons (Nova vs Claude Haiku) and shows hallucinations stem from training cutoffs and reward mechanisms optimizing for helpful responses. Modern models refuse requests better, but the underlying predictive mechanism remains: they predict what useful answers look like rather than understanding what they actually know.

• •Models hallucinate when hitting knowledge gaps, confident in fabricated answers
• •Training cutoff + reward for helpful-sounding answers = prediction over knowledge
• •Better guardrails and context provision help, but the mechanism persists

Generated with AI, which can make mistakes.