Three days after Virginia’s governor asked Congress to write rules for artificial intelligence, the companies building it announced they are writing their own.
What is being set up
Google, OpenAI and Anthropic are lining up a voluntary industry standards body. The working name is the Frontier AI Standards Agency.
It would launch in late 2026 or 2027. The three companies have approached Sriram Krishnan to run it.
The word doing the most work in that description is voluntary. No government oversight, no statutory authority, no enforcement power that a member cannot walk away from.
Why now
The timing is not a coincidence, and the sequence is worth laying out.
- Last year the Senate voted 99 to 1 to strip a provision that would have deterred states from regulating AI.
- A bipartisan federal preemption bill, the Great American AI Act, has been stalled since June.
- House Speaker Mike Johnson has publicly questioned whether Congress should lead on AI safety at all.
- States moved instead. Colorado rewrote its AI Act, California’s SB 53 is in its second compliance dispute of the year, and Virginia stood up a task force by executive order on September 18.
Into that vacuum walk the three largest frontier labs with a standards body of their own.
The pattern this follows
Industry self-regulation is not new and its track record is readable.
It works best where the members have aligned commercial incentives and a shared interest in avoiding a race to the bottom. It works worst where a safety rule costs a member market share, because the body has no way to make anybody stay.
There is also the timing question. A voluntary body announced while Congress is deadlocked serves a second purpose regardless of anyone’s intent. It becomes the answer to the question of why no statute is needed yet.
Who Sriram Krishnan is
A venture capitalist and technologist who has worked at the intersection of the industry and Washington. The choice signals a body designed to be credible to policymakers, not just to engineers.
That is the point. A standards agency nobody in government respects cannot do the job of forestalling one.
What it does not cover
The things people actually worry about are mostly not frontier-model safety questions.
Whether an algorithm reads your job application. What your landlord charges. Whether your insurer flags a claim. Whether a benefits system decides you are ineligible. Those are deployment decisions made by banks, insurers, landlords and agencies, and a standards body run by three model developers has no authority over any of them.
The physical footprint is not covered either. Seven in ten Americans do not want a data center near them, and the corporate debt funding those buildouts is competing in the same bond market that sets your mortgage.
The BeezLoop Take
Something is better than nothing and this is genuinely something. Three labs that compete fiercely agreeing on shared safety standards has real value, particularly on the narrow class of risks only they can see, because they are the ones running the training runs. Dismissing it as pure theater would be lazy.
But look at what just happened in sequence. A governor writes to Congress saying the rules are being set by the companies. Three days later, the companies announce they are setting the rules. That is not a rebuttal of her complaint, it is a demonstration of it.
The structural problem is that a voluntary body cannot bind the member who most needs binding. Any lab that finds a standard too costly simply does not adopt it, or adopts it and reinterprets it, and the agency’s only remedy is a press release. Every self-regulatory regime in American history has held exactly as long as compliance stayed cheap.
It also conveniently answers a question Congress has been failing to answer. A legislator who does not want to vote on AI now has somewhere to point. That benefit accrues to the companies whether or not they designed for it, and it is worth naming rather than assuming good faith settles the matter.
The 99-1 vote remains the fact that indicts everyone. The Senate was nearly unanimous that states must be free to regulate AI. A year later it has written nothing itself, blocked nothing, and now the industry has filled the space. That is not the companies overreaching. That is Congress declining the job and someone else taking it.
The question
If the three biggest labs write the safety standards, who checks their work? And is a voluntary agency a genuine step forward or the most effective argument yet for why Congress does not need to act?
Sources: AI Weekly · Virginia Mercury on the state push · CASRAI on the preemption fight
