As AI reshapes work, history keeps repeating one pattern. Across industrialization, the China shock, and routine office automation, the workers who fell through the cracks were often predictable in advance: those with the least institutional support, the fewest transferable skills, and the least mobility. One lesson from history is not just who becomes vulnerable. It is whether the response reaches them in time. Read on Substack ⬇️

The Fathom Daily Dispatch: the day's top AI news, in your inbox every weekday morning. Federal, state, global, and industry, every news story sourced and linked. Sign Up Now ⬇️

Great to see Anthropic recognize that self-assessment is not enough and that a robust ecosystem of independent and empowered AI evaluators is needed. The challenge now is institutional. Independent evaluators must be more than consultants or auditors—they need to become trusted oversight organizations capable of providing credible assurance about safety, governance, and risk management. To earn that trust, they cannot simply review a developer’s own claims or testing methodologies. They must be empowered to independently stress-test and “crash test” frontier systems, probing for risks that developers may have overlooked and validating whether safeguards work in practice, and then be empowered to work with government to prevent the most dangerous systems from being released. And like every other part of the AI ecosystem, evaluators themselves must be accountable for the quality and rigor of their work. Building that ecosystem will require thoughtful policy, real investment, strong independence safeguards, and a commitment to preventing industry capture. The need is clear. The next step is building the institutions to meet it.

Much of the AI and jobs debate is about income. But work is also how people build competence, find connection, structure their days, and answer who they are. The policy toolkit on the income side is far more developed than anything addressing the meaning side. That gap is still underbuilt. More on Substack 👇

History won't hand us a playbook for AI and work. But it can stress-test the assumptions we're betting on. The risk is drifting into a default response before the picture is clear. Read more on Substack ⬇️

Google DeepMind's @OCLarter was recently on @scaling_laws w/ @KevinTFrazier and talked about "a fluid, dynamic, standards-based approach to governance" that can actually keep pace with the technology. Worth a listen (and bonus points if you catch the Ashby Workshops mention).

Sam Altman called for a New Deal-scale social contract for AI. Weeks later, he said the “jobs apocalypse” he’d braced for hadn’t arrived. That tension captures the challenge: AI’s impact on work is uncertain, but the window to prepare may be short. History won’t hand us a playbook. But it can stress-test our assumptions. New on our Substack ⬇️

Florida recently sued OpenAI and Sam Altman — the first state to target an AI company over product design and safety. We saw this coming. A thread on AI liability and what's really at stake 🧵

That predictability gap is why we call for Independent Verification Organizations: expert bodies that translate goals like "don't materially encourage self-harm" into concrete, continuously updated, testable criteria — verified on an ongoing basis, not reconstructed after a tragedy. Crucially, verified compliance becomes evidence of reasonable care that companies can point to in court. That's the incentive to actually meet the bar.