Language models often generate false information, known as hallucinations, due to training methods that reward guessing over acknowledging uncertainty. The article discusses how evaluation procedures can incentivize this behavior and suggests that improving scoring systems to penalize confident errors could help reduce hallucinations in AI systems.