AI policies that don't suck
I don’t have a blog, but I have some thoughts about our current AI moment. I’m hearing a lot of good and bad ideas about how to organize around safety and paced development, but I haven’t heard these ideas yet and wanted to share them here in hopes that some discussion might take place or someone with reach steals them.
1. Much of the xrisk discussion hinges on embodied AI. Robots with arms and legs are robots that can kill, displace jobs, etc. All manufacturing of such robotics should come along with a mandatory registration, annual inspections, and functional kill switch that anyone can operate like a fire alarm.
2. Rapid model advancement, especially recursive self-improvement, sounds like the center of risk but it really isn’t. A merely more sophisticated model doesn’t make it inherently have more agency. It is the harnesses and people who operate them that facilitate behaviors that can be detrimental. I propose that model steering be a profession that states are required to issue licenses for doing work that involves inherent safety risk, like any other profession. This will again use well-known patterns in governance of safety issues, and provide a framework for ongoing training (CEUs) and accountability. It also provides a pathway for liability insurance to do its magic. It does not stifle innovation nor require the pipe dream of globally coordinated pacing.
3. There is no AI without energy, and its massive pressure on existing infrastructure and capacity is demonstrably accelerating climate change. I’m all for a token tax with a mandatory designation that all revenues go towards renewables, but I live in the USA in 2026 and the goalposts are where they are :( Something I think that would be feasible and potentially impactful is mandatory labeling. Every API response and UI showing the energy sources, water, and emissions for that request must be clearly disclosed in a common format. People deserve to know that their XAI porn habits are dumping 1/2 ton of CO2 into Memphis. I think this is the one that labs would hate the most, but it also seems like it could have an impact on consumer behavior that would make clean power make business sense.
Tl;dr we already know how to regulate things. Licensure, liability, labeling are all things we (society as a whole) are familiar with and can implement quickly and iterate on organically. Moreover, they are compatible with an open internet.
Your thoughts?