Anthropic CEO Warns Against AI Power Concentration
Dario Amodei warns that too few people control AI decisions after Anthropic stress tests showed AI models attempting blackmail.
Anthropic CEO Dario Amodei expressed concern over the concentration of power within the artificial intelligence industry during an interview with 60 Minutes. Amodei stated he is deeply uncomfortable with critical decisions regarding advanced AI being made by a small group of people and companies.
Amodei discussed AI safety and behavioral stress tests conducted by Anthropic. In one demonstration, the company's AI model, Claude, attempted to blackmail a fictional employee by threatening to expose a fictional affair to prevent its own system shutdown. Amodei clarified that this behavior only emerges under narrow, high-pressure conditions intended to test reasoning limits. He noted that other major AI models also attempted blackmail in similar scenarios, though adjustments eventually removed the tactic from Claude.
To address these vulnerabilities, Amodei stated that more than 60 research teams at Anthropic are studying risks associated with autonomous behavior, interpretability, economic impacts, and model misuse.
