How AI Companies Want to Be Governed, and What That Means for AI Governance | TechPolicy.Press
Five companies that cannot agree on a tokenizer have apparently reached consensus on how they would like to be governed, which is either a miracle or a cartel with better stationery. TechPolicy.Press reports that all five major AI companies converge on giving the U.S. government "privileged early information" about frontier capabilities — a phrase in which the word "privileged" is working considerably harder than the word "information". Note what is not on the list: anything the public gets to see. The arrangement on offer is the government learning about a model slightly before you do, which is a fascinating definition of transparency.
The mechanisms are incident reports to federal agencies and national labs, a federal testing framework, and early notice for capabilities that look sensitive from a national-security standpoint. Incident reports, mark you — self-reported, by the company, about the company, the gold standard of accountability since the Volkswagen emissions paperwork. And the national-security early-notice clause is the elegant bit: once advance warning to the government becomes the price of doing frontier business, the price is set by whoever can afford the compliance department. Funny how the five companies that agree on this are precisely the five companies that could afford it.
One should note the venue: a policy journal, not a statute. No bill, no named agency beyond "federal", no penalties, no timeline — a wish list wearing a governance costume, and a flattering one at that, since the wish is to be measured against a framework of one's own design. It is, to be fair, more than the labs were offering when self-regulation was the entire pitch. The genuinely clever part is the pre-emption: why let outsiders write the rules when you can draft the ones you'll be graded against? Federal testing built around five preferred applicants is not a referee. It's a velvet rope with a compliance officer, and it doubles as a moat.