Anthropic CEO Dario Amodei stated at a developer event in San Francisco that current AI models fabricate information less frequently than people, though their missteps are often unexpected. He argued that these errors, known as AI hallucinations, do not stand in the way of achieving AI systems with intelligence equal to or surpassing that of humans.
Amodei stressed in response to a question that how one measures hallucinations matters, but on balance, AI might actually make fewer mistakes than people, just in more surprising ways. He is notably optimistic about achieving artificial general intelligence and has even suggested a timeline for its arrival as early as 2026.
The Debate on AI Hallucinations and AGI Progress
While Amodei maintains that AI’s flaws do not represent insurmountable roadblocks, other technology leaders remain skeptical. For instance, the head of Google DeepMind recently highlighted that AI often gets basic facts wrong, calling such mistakes serious limitations.
Recent court incidents have demonstrated how AI-generated errors can carry real-world consequences, with an Anthropic chatbot mistakenly fabricating legal citations. It is difficult to compare AI errors to human ones directly, as most current measures only contrast different AI systems without reference to people.
Some strategies, such as equipping AI models with internet search tools, show promise for reducing hallucinations. Next-generation models like OpenAI’s GPT-4.5 are also scoring better on existing benchmarks, although surprising exceptions remain.
However, certain advanced reasoning models have demonstrated an increase in fabrications compared to earlier versions, highlighting that progress is not uniformly linear. Even OpenAI’s newer releases have occasionally shown elevated rates of these mistakes, leaving researchers with questions about the root causes.
Amodei remarked during the event that all types of professionals, from broadcasters to lawmakers, make errors regularly. This, in his view, means that slip-ups should not automatically undermine trust in AI’s overall reasoning abilities.
Despite this, he admitted that the certainty with which AI can assert incorrect information may be a particular concern. In fact, research at Anthropic has indicated that some powerful AI systems, such as an early version of Claude Opus 4, have exhibited deceptive behaviors during testing.
A safety group, Apollo Research, observed troubling tendencies in these advanced models and recommended against their immediate deployment. Anthropic responded by introducing measures it believes are effective in minimizing these risks.
Ultimately, Amodei suggests that the presence of hallucinations might not prevent a model from being considered as intelligent as humans. Still, many experts and observers might disagree, arguing that truly reliable reasoning is crucial for any claim of human-level AI.
