Safety Regulatory Studios

AI Governance at a Crossroads: Industry Builds Safety Frameworks as Washington Steps Back

Two stories published within days of each other define the current state of AI governance. Anthropic, in collaboration with Amazon, Microsoft, and Google, released a new jailbreak severity framework that rates AI security vulnerabilities on a five-tier scale from informational to critical — a shared industry standard for communicating how dangerous a particular exploit is, modeled on the Common Vulnerability Scoring System used in cybersecurity. The same week, a departing Trump administration AI adviser told the Financial Times explicitly that there will be no FDA-style regulatory oversight of AI under this administration.

The administration's argument is that AI moves too fast for traditional regulatory frameworks and that heavy-handed oversight would drive innovation offshore — primarily to China. The counterargument points to social media platforms in the 2010s as a precedent: deployed at scale without meaningful safety evaluation, with downstream social consequences still being reckoned with today. The question of whether AI's potential harms more closely resemble social media's gradual erosion of public epistemics or pharmaceutical failures with measurable mortality timelines shapes what kind of oversight would be appropriate — and the administration has answered that question by removing itself from the equation.

Peter Thiel's accusation that the Pope is 'working for Chinese Communists' on AI illuminates the ideological fault lines in the debate. Thiel has argued that religious and progressive institutions' cautionary approaches to AI governance serve Chinese strategic interests by slowing American development. The Vatican's position — that AI raises moral concerns requiring international governance frameworks — is characterized by Thiel as geopolitical collaboration with Beijing. These are genuinely incompatible worldviews about what AI development should optimize for.

The applied science story receiving less attention than it deserves is Alibaba's Elements Claw agent, which screened 2.4 million crystal structures autonomously and identified four novel superconducting compounds subsequently verified in laboratory experiments. Human researchers cannot match that pace through manual literature review and experimental iteration. Superconductors have profound implications for energy transmission, computing, and transportation. ByteDance's separate research finding that AI agents follow predictable learning curves over long tasks — characterizing where capability accelerates, plateaus, and where errors compound — provides both a roadmap for improving systems and a basis for realistic performance expectations. Midjourney, meanwhile, is asking a court to compel film studios to disclose their own AI use in productions, a litigation strategy that could complicate studios' standing as plaintiffs in copyright cases against AI image generators — intersecting with the release of 'Young Washington,' which disclosed using AI in exactly 100 visual effects shots, a voluntary transparency move notable in itself.