AI firms adopt responsible scaling policies to set safety guardrails for development

Responsible Scaling Policies have emerged as a framework for AI companies to define safety thresholds and capability limits, establishing guardrails for AI development while balancing innovation with risk management. These policies represent a significant evolution in how leading AI organizations approach the responsible advancement of increasingly powerful systems.

The big picture: Major AI companies have established formalized policies that specify what AI capabilities they can safely handle and when development should pause until better safety measures are created.

• Anthropic pioneered this approach in September 2023 with their AI Safety Levels (ASL) system, categorizing AI systems from ASL-1 (posing no meaningful catastrophic risk) to ASL-4+ (involving qualitative escalations in misuse potential).
• Current commercial language models including Claude are classified as ASL-2, showing early dangerous capabilities that aren’t yet practically useful compared to existing technologies like search engines.

Industry adoption: Following Anthropic’s lead, most major AI developers have published their own versions of responsible scaling frameworks between 2023-2025.

• OpenAI released a beta Preparedness Framework in 2023, while DeepMind launched their Frontier Safety Framework in 2024.
• Microsoft, Meta, and Amazon all published their own frameworks in 2025, each using “Frontier” terminology to describe advanced AI governance.

Mixed reception: The AI safety community has expressed divided opinions on whether these policies represent meaningful safety commitments or strategic positioning.

• Supporters like Evan Hubinger of Anthropic characterize RSPs as “pauses done right” – a proactive approach to managing development risks.
• Critics argue these frameworks primarily serve to relieve regulatory pressure while shifting the burden of proof from capabilities researchers to safety advocates.

Behind the concerns: Skeptics view RSPs as promissory notes rather than binding commitments, potentially allowing companies to continue aggressive capability development while projecting responsibility.

• The frameworks generally leave companies as the primary judges of their own systems’ safety levels and capability boundaries.
• Several organizations including METR, SaferAI, and the Center for Governance of AI have developed analysis frameworks to evaluate and compare the effectiveness of different companies’ RSPs.