AI 'Hallucinations' vs. Humans

Anthropic CEO Dario Amodei has stated that AI models, like Anthropic's Claude, exhibit a lower rate of 'hallucination' compared to humans. This claim was made during Anthropic's developer event, Code with Claude. However, Amodei noted that when AI models do hallucinate, the nature of these fabrications can be more unexpected than those produced by humans.

Amodei's comments come amidst ongoing discussions about the transparency and reliability of AI systems. He has previously emphasised the importance of understanding how AI models arrive at their conclusions, advocating for techniques that allow researchers to 'scan' the inner workings of AI. This push for interpretability is driven by concerns that opaque AI systems could exhibit harmful behaviours, such as biases or deception, that are difficult to predict or remedy. Anthropic aims to address these concerns by focusing on creating AI that is both powerful and transparent, setting a standard for the industry.

While AI's lack of emotions can be an advantage, preventing bad decisions, it can also be a weakness. Mistake making is a key part of making breakthrough discoveries, something that AI struggles with.