Newer, larger versions of AI chatbots are more likely to give false answers rather than admit ignorance. This is the finding of a study by José Hernández-Orallo and colleagues published in the journal Nature. The researchers looked at three major language models: GPT, LLaMA and BLOOM. They found that while the accuracy of responses increases with model size, reliability decreases.
The study also showed that humans often have difficulty detecting incorrect answers from chatbots. Between 10% and 40% of inaccurate answers were misclassified as correct. The researchers recommend improving the AI’s performance on simple questions and encouraging it to reject answers to difficult questions. This could help users better assess the reliability of AI systems.