Anthropic CEO claims AI models hallucinate less than humans
TechCrunch
Press briefing at Code with Claude, Anthropic's developer event, San Francisco
“It really depends how you measure it, but I suspect that AI models probably hallucinate less than humans, but they hallucinate in more surprising ways”
Dario Amodei22 May 2025- What you'll lose0/5
Nothing taken.
- Condescension3/5
Humans make mistakes too.
- Tone-deafness3/5
The model's errors, defended by pointing at yours.
- Doom0/5
None.
- Hypocrisy0/5
None.
Full time6 pitchforksof a possible 25
Think the referee got a count wrong?
Call received. VAR is having a look: the referee reviews every call, and the score stands until then.
The occasion
Press briefing at Code with Claude, Anthropic's developer event, San Francisco, 22 May 2025.
Context
Responding to TechCrunch's question at a press briefing, Amodei argued that hallucinations are not a barrier on Anthropic's path to AGI, because humans also make confident mistakes. He added that how you measure it matters.
Crowd reaction
Critical commentary
Mild. TechCrunch's own report was sceptical: it called the claim 'difficult to verify' because most benchmarks compare models with each other and not with humans, noted evidence that hallucinations are 'getting worse in advanced reasoning AI models', and said the remark suggests Anthropic may call a model AGI even if it still hallucinates, which 'may fall short of AGI by many people's definition'. Business Today repeated it with 'but there's a catch'. No social-media backlash was documented in the pages read.