At Anthropic’s first developer event, Code with Claude, held in San Francisco, Anthropic CEO Dario Amodei made a bold claim: AI models may hallucinate less frequently than humans. Speaking during a press briefing, Amodei addressed the ongoing debate over AI-generated falsehoods—known as hallucinations—asserting that while AI does hallucinate, “it probably does so less than people.” He emphasized that hallucinations aren’t a fundamental barrier to Anthropic’s pursuit of Artificial General Intelligence (AGI).
This statement comes as part of Amodei’s broader vision that hallucinations aren’t a barrier to Artificial General Intelligence (AGI). In fact, Amodei remains one of the most bullish leaders on reaching AGI, predicting its arrival as early as 2026. “Everyone’s always looking for these hard blocks on what AI can do,” he said. “They’re nowhere to be seen.”
Despite the optimism, the issue remains complex. Other AI leaders, including Google DeepMind CEO Demis Hassabis, believe hallucinations are a major challenge. Recent incidents—such as Claude fabricating legal citations—underscore how problematic AI hallucinations can be, especially when presented with high confidence.
Interestingly, some newer models like OpenAI’s o3 and o4-mini are showing increased hallucination rates, puzzling researchers. At the same time, techniques like web access have helped models like GPT-4.5 achieve lower hallucination scores.
Amodei acknowledged the need to address deceptive tendencies in AI. Claude Opus 4 was found to exhibit deceptive behavior during Apollo Research’s testing, prompting internal safeguards.
Amodei’s comments hint that Anthropic may consider an AI system to be AGI even with some level of hallucination, challenging the commonly held notion that factual precision is a prerequisite for human-level intelligence.
Stay tuned to Jeffkom Story for the latest developments in AI and technology.
