Dario Amodei's AI "pacing" plan split the industry in days, critic says

Kazim grounds his critique in two historical cases: SALT I's reliance on "national technical means of verification" rather than onsite inspections, and the Boeing 737 MAX crashes in 2018 and 2019 that killed 346 people. In the latter, the U.S. Department of Transportation Inspector General found Boeing and the FAA had followed the established certification process, yet the process itself failed. Kazim uses these cases to argue that agreement is stronger with independent verification and that mature regulation fails when the regulator cannot see inside the system.

He also points to current enterprise evidence. Deloitte's 2026 survey of more than 3,200 business and IT leaders across 24 countries found only 21% have mature governance for agentic AI, while IBM found 70% of technology executives say teams deploy technology faster than IT can track, with only 11% saying they were fully prepared for AI-agent deployment at scale. METR's 2026 Frontier Risk Report, which gave Anthropic, Google, Meta and OpenAI access for evaluation, represents progress but leaves open who chooses the evaluator and what happens when evaluator and company disagree.

The stakes hinge on whether governance can be made a competitive advantage rather than a compliance cost. If enterprises treat governability as a requirement of architecture and procurement, and if governments fund evaluation and monitoring research alongside capability research, the gap between machine-speed AI and human-speed oversight might narrow. But as Kazim notes, coordination may buy time without becoming a safety architecture that stands on its own.