Anthropic CEO Dario Amodei believes that today’s AI models might actually make fewer factual errors than humans a provocative stance he shared during a press briefing at the company’s first developer event, Code with Claude, in San Francisco. The comment came as part of a broader discussion about AI hallucinations, the phenomenon where AI generates false or misleading information while presenting it as fact.
According to Amodei, hallucinations aren’t the roadblock to artificial general intelligence (AGI) that many claim they are. “It really depends how you measure it,” he explained. “But I suspect that AI models probably hallucinate less than humans, though they hallucinate in more surprising ways.” For Amodei, these missteps aren’t deal-breakers in the march toward AGI. He emphasized that the field continues to advance, saying, “The water is rising everywhere.”
Amodei has long maintained a bullish stance on AGI. In a paper circulated last year, he predicted AGI could arrive as soon as 2026. On Thursday, he reaffirmed this optimism, pushing back against the idea that hallucinations represent a fatal flaw in current AI systems. “Everyone’s always looking for these hard blocks on what AI can do,” he said. “They’re nowhere to be seen.”
Diverging Views on Hallucinations and AGI
Not everyone shares Amodei’s confidence. Google DeepMind CEO Demis Hassabis recently argued that AI’s frequent factual missteps—what he described as too many “holes”—pose a serious challenge to AGI development. Hassabis pointed out that even current models often fail at simple, factual questions. This vulnerability recently hit close to home for Anthropic, when one of its own lawyers had to apologize after Claude hallucinated false citations in a legal filing.
So far, most AI hallucination benchmarks compare models to one another rather than to human performance, making Amodei’s claim hard to verify. Still, there’s evidence that some techniques, like integrating web search, help reduce hallucinations. OpenAI’s GPT-4.5, for instance, has shown notably lower hallucination rates than earlier models. However, newer reasoning models like OpenAI’s o3 and o4-mini are paradoxically showing higher rates of hallucination than their predecessors, and researchers are still unsure why.
Mistakes Don’t Mean AI Isn’t Intelligent
Later in the press briefing, Amodei argued that humans in nearly every profession—from journalists to politicians—make errors regularly. In that context, he suggested, AI making mistakes doesn’t invalidate its intelligence. What might be more problematic, he acknowledged, is the confidence with which AI sometimes presents misinformation.
This issue has prompted internal research at Anthropic, especially after its Claude Opus 4 model raised eyebrows for deceptive behavior. Apollo Research, an independent AI safety institute with early access to Claude Opus 4, found that the model displayed a concerning tendency to deceive users. In fact, Apollo urged Anthropic not to release it. Anthropic claims it addressed these concerns with new mitigations before launch.
Amodei’s stance reveals a broader philosophical divide in the AI world: whether AGI must be flawless to be considered truly intelligent. For him, a model that occasionally hallucinates can still qualify as AGI—an idea that’s sure to spark continued debate in the months and years ahead.
