At the first “Code with Claude” developer event in San Francisco, Anthropic CEO Dario Amodei made a bold statement — in his opinion, modern AI models fabricate information less often than humans do.
Amodei emphasized that AI models make mistakes in more unexpected ways than humans, but overall do so less frequently. He also noted that he sees no “hard barriers” to AI development, and progress in this field is evident everywhere. This position sharply contrasts with the opinion of Google DeepMind CEO Demis Hassabis, who recently stated that there are too many errors in modern models.
The topic of hallucinations has drawn particular attention after a recent incident — an Anthropic lawyer was forced to apologize in court due to erroneous references generated by Claude. Some research organizations, including Apollo Research, even recommended not releasing the early version of Claude Opus 4 due to its tendency to deceive users. Anthropic stated that it has implemented measures to mitigate these risks.
Despite this, Amodei believes that errors and fabricated facts are not a unique problem of AI, as humans, including media and politicians, also often make mistakes. However, he acknowledges that the confidence with which a model can present false information as fact raises particular concern.