Anthropic Drops Flagship Safety Pledge

Anthropic, the AI startup founded on principles of safety and co-founded by former OpenAI researchers, has made a pivotal strategic shift by formally abandoning its 2023 commitment to pause the training of advanced AI models if safety evaluations fell behind capability gains. This 'Responsible Scaling Policy' was a cornerstone of Anthropic's public identity, distinguishing it from competitors by promising a self-imposed 'off-switch' for development in the face of uncontrolled risks. The change, approved by CEO Dario Amodei and reported by TIME, signals a maturation of the company's internal safety frameworks but also reflects the intense commercial and technological pressures of the generative AI arms race, where pausing development could mean ceding a decisive lead.

The revised policy, detailed in an internal document, replaces the binary 'halt training' pledge with a more nuanced, tiered system of safety thresholds (ASL-2, ASL-3) that triggers intensified safety research and mitigation efforts rather than an automatic stop. This move follows Anthropic's recent launch of the powerful Claude 3.5 Sonnet model and comes amid reports of the company seeking billions in new funding to compete with OpenAI's GPT-4o and Google's Gemini. The decision underscores the practical challenges of implementing theoretical safety guardrails while operating a for-profit enterprise in a hyper-competitive market, potentially recalibrating the entire industry's approach to voluntary safety commitments.