The CEO of Anthrope, Dario A Amodei, believes that today’s models hallucinates, or invent things and present them as if they were true, at a lower pace than humans, he said during a press conference at the first Anthrope developer event, Code With Claude, in San Francisco on Thursday.
Amodei said all this in the midst of a bigger point he was doing: that the hallucinations of AI are not a limitation on the path of Anthrope towards agi-aii systems with intelligence at a human or better level.
“It really depends on how I measure it, but I suspect that the models of AI probably hallucinate less than humans, but hallucinate more surprisingly,” said Amodei, answering TechCrunch’s question.
The CEO of Anthrope is one of the most bullish leaders in the industry about the perspective that the AGI AGI models. In a widely circulated paper he wrote last yearAmodei said he believed that AGI could come as soon as 2026. During the informative session of the press on Thursday, the anthropic CEO said he was seeing constant progress for that purpose, pointing out that “the water is climbing everywhere.”
“Everyone always looks for these hard blocks in what [AI] You can do, “said Amodei.” They cannot be seen anywhere. There is no such thing. “
Other leaders of AI believe that hallucination presents a great obstacle to achieve AGI. Earl this week, Google Deepmind CEO, Demis Hassabis, said Today’s models have too many “holes” And get too many questions obvious badly. For example, earlier this month, a lawyer who represents Anthrope was forced to apologize in court after they used Claude To create quotes in a judicial presentation, and the hallucinated chatbot and was wrong with names and titles.
It is difficult to verify Amodei’s statement, largely because most of the hallucination reference points face the AI models; They do not compare models with humans. Certain techniques seem to be helping the lowest hallucination rates, such as giving access to AI models to the web search. Separately, some AI models, such as OpenAi’s GPT-4.5They have remarkably lower hallucination rates at reference points compared to the first generations of systems.
However, there is also evidence that suggests that hallucinations really get worse in advanced reasoning AI models. OPENAI O3 and O4-MEMORYS have Higher hallucination rates than OpenAI’s previous generation reasoning modelsAnd the company really doesn’t understand why.
Later in the newspaper of the press, Amodei said that television broadcasors, politicians and humans in all kinds of professions make mistakes all the time. The fact that AI also makes mistakes is not a blow to its intelligence, according to Amodei. However, Anthrope’s CEO recognized the confidence with which AI models present false things, since the facts could be a problem.
In fact, Anthrope has carried out a good amount of research on the trend of AI models to deceive humans, a problem that seemed especially frequent in Claude Opus 4’s company recently launched by the company. High tendency to plan against humans and deceive them. Apollo came to suggest that Anthrope should not have launched that early model. Anthrope said that some mitigations occurred that seemed to address the problems that Apollo raised.
Amodei’s comments suggest that anthropic may consider that an AGI model, or equal to intelligence at the human level, even if it is still hallucinated. However, an AI that may not be agi due to the definition of many people.