AI Developers Scale Back Safety Engagements

Leading AI developers, including Google DeepMind, OpenAI, Anthropic, and Meta, have reportedly weakened their voluntary safety commitments. This finding comes from a study by the Center for AI Safety (CAIS), which assessed company alignment with safety best practices.

The study analysed public statements and policy changes across these firms. Researchers observed a shift over the past year, with several companies reducing their engagement with safety protocols they previously endorsed. This retraction of commitments raises concerns among some researchers. CAIS developed a 'safety index' to track these developments. The index evaluates factors such as responsible scaling policies and external auditing.

OpenAI, for example, reportedly removed its 'Preparedness Team Charter' from its website. This document previously outlined a framework for monitoring and mitigating risks from advanced AI systems. Sources close to the company indicate the team itself remains, but its public charter is no longer accessible. Furthermore, OpenAI is said to have discontinued its chief scientist's tradition of publishing an annual AI safety letter.

Other companies also show similar trends. Google DeepMind reportedly altered its 'responsible scaling' policies. These policies previously mandated specific risk assessments before deploying more powerful models. Now, the language is less stringent, suggesting a reduced emphasis on predefined thresholds. Meta, meanwhile, has reportedly scaled back specific safety-related initiatives within its open-source model development. Anthropic also reportedly revised its public stance on certain alignment research priorities.

The CAIS study suggests these changes coincide with increasing pressure to deploy more capable models quickly. The weakening of these voluntary measures occurs amid the continued rapid advancement of AI capabilities. The study prompts a re-evaluation of industry-led safety initiatives. What might this shift mean for future AI development and regulation?