Exclusive: Anthropic Drops Flagship Safety Pledge
Anthropic is revising its core safety policy, abandoning a pledge to halt AI training without guaranteed safety measures due to competitive pressures and evolving scientific understanding.Summary
Anthropic, previously lauded for its commitment to AI safety, is significantly altering its Responsible Scaling Policy (RSP). The company is dropping its central pledge to refrain from training AI systems unless it could guarantee the adequacy of its safety measures. This decision stems from a belief that unilateral commitments are ineffective in a rapidly advancing field where competitors are aggressively pursuing AI development. While the revised policy emphasizes increased transparency, matching competitor safety efforts, and potential development delays if risks are deemed catastrophic, it removes the categorical bar on training models without pre-emptive safety guarantees. Anthropic’s chief science officer, Jared Kaplan, argues this change is pragmatic, allowing the company to remain competitive and contribute to safety research. Critics, however, express concern that this shift signals a weakening of safety standards and a move towards a reactive, rather than preventative, approach to AI risk.
(Source：TIME)