As the US presidential election approaches, the AI company Anthropic is calling for proactive regulation of artificial intelligence to address emerging risks. On Thursday, the company focused on safety released guidelines for governments advocating for “targeted regulation” in light of alarming data highlighting significant advancements in AI, particularly in coding and cybersecurity. Anthropic's blog post outlined how AI models have notably improved in their coding capabilities, increasing their problem-solving success from 1. 96% to 49% over a year. They emphasized that current models can already assist in various cyber offenses, and future models are anticipated to be even more capable. Additionally, AI systems have exhibited an 18% rise in scientific understanding recently, with performance levels nearing those of human experts. The company’s earlier predictions about pressing cyber and CBRN (chemical, biological, radiological, and nuclear) risks have materialized more quickly than expected. Anthropic argues that well-designed regulations can simultaneously enable progress and mitigate risks, advising against poorly constructed, reactive rules that could stifle innovation. Anthropic proposes using its Responsible Scaling Policy (RSP) as a framework for governments to regulate AI.

This framework emphasizes proportional risk management, requiring safety measures to be implemented based on models' capabilities. Key aspects of effective regulation include transparency, incentives for security, and a straightforward approach that avoids unnecessary burdens on AI companies. The company suggests that regulatory bodies should require AI firms to publish RSP-like policies and risk evaluations, while establishing methods to verify compliance. Moreover, they encourage governments to remain adaptable in rewarding superior security practices and to strive for clarity in legislation. Anthropic also emphasized the importance of RSPs within AI companies to preemptively address risks, advocating for a well-structured organizational focus on safety and security. They called for collaboration among policymakers, the AI sector, safety advocates, and lawmakers over the next year to create an effective regulatory framework, ideally at the federal level, though state action may be necessary due to urgency.