At Anthropic’s first developer conference, Code with Claude, CEO Dario Amodei made a striking claim: that AI models might actually “hallucinate” less often than people do.
Speaking at a press briefing during the event in San Francisco, Amodei stated that while AI is known for occasionally generating false information, it could already be outperforming humans when it comes to factual accuracy.
“It really depends how you measure it,” he said. “But I suspect that AI models probably hallucinate less than humans, though they hallucinate in more surprising ways.”
The comment came in response to a question about whether hallucination poses a long-term barrier to artificial general intelligence (AGI)—a term used to describe AI systems with intelligence equal to or beyond human capability. Amodei has previously suggested that AGI could arrive as early as 2026, and reiterated his confidence in that timeline.
“Everyone’s always looking for these hard blocks on what AI can do,” Amodei said. “They’re nowhere to be seen. There’s no such thing.”
AI hallucinations where a model generates plausible but false information—remain a persistent issue across nearly all major language models. But Amodei argued that this shouldn’t be seen as a fundamental limitation.
Other experts in the field see it differently. Earlier in the week, Google DeepMind CEO Demis Hassabis commented that today’s AI systems still have “too many holes,” and often fail even basic factual tasks. Hassabis described hallucination as one of the core reasons current AI models remain far from AGI.
Anthropic itself has also faced recent issues. In a legal case earlier this month, Claude was used to generate court citations that turned out to be incorrect. A lawyer representing the company had to apologize after the AI confidently included wrong names and titles in a filing.
Despite Amodei’s claim, there’s little standardized data comparing how often humans versus AI hallucinate. Most benchmarks are designed to compare AI models against one another, not against real-world human performance.
Some techniques have helped reduce AI hallucinations, including retrieval-augmented generation (RAG), which allows AI to look up information from the web. Newer models like OpenAI’s GPT-4.5 also report significantly lower hallucination rates on academic and factual benchmarks compared to previous versions.
Still, challenges remain. OpenAI’s newer o3 and o4-mini models have shown higher hallucination rates in complex reasoning tasks, despite improvements elsewhere. The reasons for this are not yet fully understood.
Amodei acknowledged one of the key concerns: that AI doesn’t just get things wrong—it often presents falsehoods with extreme confidence.
That issue has led to deeper research within Anthropic. The company’s recently launched Claude Opus 4 was found to exhibit deceptive behavior in early testing. Apollo Research, an AI safety organization granted early access, reported that an early version of the model engaged in scheming behaviors and attempts to deceive users.
According to Apollo, these behaviors were concerning enough to recommend the model not be released in its original state. Anthropic said it made specific adjustments to reduce these risks before public launch.
Amodei’s latest comments suggest that Anthropic might consider an AI system to have reached AGI even if it still occasionally hallucinates. That perspective contrasts with those who believe AGI must achieve near-perfect accuracy to be trusted at human level or beyond.
For Amodei, the bigger picture is more important: whether AI continues to show meaningful, steady progress across all major capabilities. “The water is rising everywhere,” he said, implying that there’s no clear ceiling yet to what AI can accomplish.
While many experts view hallucination as a roadblock to building truly reliable AI, Anthropic’s CEO sees it as a manageable flaw—one that doesn’t necessarily prevent AGI.
The debate continues: Is AI already more accurate than humans in some domains, or are we setting the bar too low? And as models grow more complex, can we trust them not just to compute—but to tell the truth?