Anthropic has unveiled its Responsible Scaling Policy (RSP), a comprehensive set of technical and organizational protocols designed to mitigate the risks associated with advanced AI systems. As AI models become increasingly capable, they present significant economic and social benefits but also pose severe risks. The RSP is aimed at addressing catastrophic risks, such as those arising from models being deliberately misused or autonomously behaving in ways contrary to their intended purpose.
The RSP introduces the AI Safety Levels (ASL) framework, which draws inspiration from the U.S. biosafety level standards. The framework mandates that safety, security, and operational standards correspond to a model’s potential risk. Higher ASL levels will necessitate more rigorous safety demonstrations. The ASL system ranges from ASL-1, where models pose no significant risk, to ASL-4 and higher, which are not yet defined but will handle future models with considerable autonomous capabilities and catastrophic risk potential.
Currently, most large language models (LLMs), including those developed by Anthropic, fall under ASL-2. These models demonstrate early signs of potentially dangerous capabilities but lack sufficient reliability to cause significant harm. ASL-3 systems, however, show a higher risk for catastrophic misuse or low-level autonomy, necessitating stronger security measures and a commitment not to deploy such models if they exhibit misuse under adversarial testing.
The ASL framework aims to balance mitigating catastrophic risks while promoting the beneficial use and safety advancements of AI. Anthropic’s policy calls for pausing the scaling of more powerful models until the necessary safety protocols are met, incentivizing the resolution of safety issues. This approach encourages a competitive, yet responsible, development of AI technologies.
This policy is analogous to pre-market safety testing in industries like automotive and aviation, ultimately benefiting customers by ensuring the safety of AI products before their release. Anthropic’s RSP has received formal board approval, with changes requiring future board consultations involving the Long Term Benefit Trust.
Anthropic emphasizes that while these commitments are their current best strategies, the fast-evolving nature of AI may necessitate rapid iterations and adjustments. The full document outlining the RSP is available for detailed review.
The creation of the RSP was significantly supported by ARC Evals, a key player in AI risk assessment, offering essential insights into evaluating autonomous capabilities. Anthropic acknowledges ARC Evals’ leadership in inspiring the development of this policy framework.
You can read the original article here: https://www.anthropic.com/news/anthropics-responsible-scaling-policy
