AI hallucinates, but humans suffer more

Anthropic 's newly released Claude 4 is smarter, but still hallucinating. Photo: Anthropic .

At Anthropic’s first developer event in San Francisco, CEO Dario Amodei said current AI models can create “hallucinations” at a lower rate than humans. “Hallucinations” here are understood as the phenomenon of AI creating fictional content but presenting it as real.

Amodei stressed that the fact that AI hallucinates doesn't hinder Anthropic's goal of achieving AGI (Artificial General Intelligence). "It depends on your yardstick, but I think the frequency of hallucinations in AI models is probably lower than in humans, it's just that the way they make mistakes is more surprising," he said.

Amodei has been one of the most vocal proponents of AGI in the industry. “People are always trying to find an ‘upper limit’ for AI,” he says. “Right now, we don’t see any such upper limit.”

Anthropic CEO is one of the advocates of "AI optimism". Photo: Anthropic.

Not everyone agrees, however. Google DeepMind CEO Demis Hassabis pointed out this week that current AI models are “full of holes,” even getting some basic questions wrong.

There are also signs that some newer models have a higher incidence of problems when it comes to complex inference tasks. OpenAI’s o3 and o4-mini versions have a higher incidence of hallucinations than previous inference models. Even OpenAI itself doesn’t understand why.

Amodei also mentioned that humans make mistakes all the time. Therefore, he said that just because AI makes mistakes does not mean it is "not smart enough". However, the Anthropic leader also admitted that they can easily cause problems when they give out false information with extremely high reliability.

Amodei’s remarks have sparked debate in the AI community about the true capabilities of current models. Some experts agree that AI is making significant progress in reducing illusions. Others warn of the potential risks of AI overconfidence in inaccurate information.

The discussion reflects a major challenge facing the AI industry. Companies are looking to develop systems that are both intelligent and trustworthy, able to handle uncertainty in their responses.

Source: https://znews.vn/ai-ao-giac-nhung-con-nguoi-bi-con-nhieu-hon-post1557486.html