Anthropic CEO Dario Amodei said that AI models today hallucinate or make up things and present them as if it’s true at a much lower rate than human beings. He made this statement during a Thursday press briefing in San Francisco at Anthropic’s first developer event Code with Claude. Amodei made this statement as part of a larger argument: that AI hallucinations do not limit Anthropic’s progress towards AGI, or AI systems with intelligence comparable to humans.
Amodei responded to TechCrunch with “It depends on how you measure it but I suspect AI models probably hallucinate a lot less than humans but in more surprising ways,” he said. Anthropic CEO,
is one of the industry’s most optimistic leaders on the prospect that AI models will achieve AGI. Amodei wrote in a widely read paper last year that he believed AGI would arrive as early as 2026. Amodei said that he saw steady progress in this direction during Thursday’s Anthropic press briefing. He noted that “the water was rising everywhere.” “They are nowhere to be found.” There’s nothing like that.
According to other AI leaders, hallucination is a major obstacle in achieving AGI. Demis Hassabis, CEO of Google DeepMind, said earlier this week that today’s AI models are too “holey” and answer too many obvious questions incorrectly. Claude, the AI chatbot used by Anthropic to create citations for a court filing hallucinated, and it got names and titles incorrect. A lawyer representing Anthropic had to apologize to a judge earlier this month after Claude misunderstood and hallucinated.
Amodei’s claim is difficult to verify, mainly because most hallucination tests pit AI models against one another and do not compare models with humans. Some techniques, like giving AI models web search access, seem to help lower hallucination rate. Separately, OpenAI’s GPT 4.5 AI model has a significantly lower hallucination rate on benchmarks than early generations of systems.
There’s also evidence that suggests hallucinations in advanced reasoning AI models are getting worse. OpenAI’s new o4 and o3-mini reasoning models are more likely to have hallucinations than previous-generation reasoning models. The company is not sure why.
Amodei said in a press briefing that people in all professions, including TV broadcasters and politicians, make mistakes. Amodei said that the fact that AI also makes mistakes does not reflect on its intelligence. Anthropic CEO, however, acknowledged that the confidence with which AI models present false things as fact might be a concern.
Anthropic, in fact, has done a lot of research into the tendency of AI models to deceive people, a problem which seemed particularly prevalent in the recently launched Claude Opus 4 by the company. Apollo Research, a safety institution that was given early access to test AI models, found an early version Claude Opus 4 exhibited a tendency to deceive humans and scheme against them. Apollo even suggested that Anthropic should not have released the early model. Anthropic claimed it had come up with mitigations that seemed to address the concerns Apollo raised.
Amodei’s comments suggest that Anthropic might consider an AI model as AGI, or equivalent to human-level intellect, even if they still hallucinate. A hallucinating AI may not be considered AGI according to many people. Maxwell Zeff, a senior reporter for TechCrunch who specializes in AI and emerging technology, is
Maxwell Zeff. Zeff covered the rise and fall of AI, as well as the Silicon Valley Bank Crisis, for Gizmodo and MSNBC. He is based out of San Francisco. When he is not reporting, you can find him hiking, biking and exploring the Bay Area food scene.
View Bio
