Anthropic, OpenAI, and Google are quietly designing their own AI regulator. Construction already has the playbook for what comes next
The three biggest AI labs have been meeting since July on a shared industry standards body for testing and certifying frontier models. Construction has run this exact model for decades with UL and ICC-ES — here's what it means once AI tools start touching life-safety decisions.
Anthropic, OpenAI, and Google have been quietly meeting since July to discuss building a shared industry body that would test and certify AI models before release — something closer to a private regulator than a trade association. Construction has already lived through exactly this kind of institution twice: UL and ICC-ES. Both started as voluntary, industry-funded testing bodies and became mandatory in practice once code officials and insurers started requiring their paperwork. That's the pattern worth watching if you're buying AI tools that touch anything life-safety adjacent.
What's actually being discussed?
CNN reported on September 14 that representatives from the three labs have held working-group meetings on the idea since July, meeting as recently as the prior week. The catalyst was a July essay from Google DeepMind founder Demis Hassabis proposing a US-led standards body modeled on the Financial Industry Regulatory Authority (FINRA) — the private, industry-funded organization that polices Wall Street broker-dealers under SEC oversight. The idea gained urgency after Anthropic CEO Dario Amodei's September 12 essay calling for the industry to pace its own capability growth — which we covered here — but the standards-body talks predate that essay by two months.
Why can't the three labs agree on what to build?
Because they want different amounts of enforcement power. Here's how the three proposals compare, based on reporting from CNN, PYMNTS, and Axios:
| Proposal | Who's pushing it | Model | Enforcement |
|---|---|---|---|
| FINRA for AI | Demis Hassabis (Google DeepMind) | Industry-funded, government-overseen, self-regulatory organization | Starts voluntary, could harden into a market-access requirement |
| FAA for AI | Dario Amodei (Anthropic) | Federal agency | Can block a model's release outright |
| IAEA for AI | Sam Altman (OpenAI) | International forum | Certifies countries and companies, not just individual models |
None of these exist yet. The reporting is consistent that the three CEOs agree something should exist, and disagree on how much power it should have — which is exactly the phase FINRA itself was in before the Securities Exchange Act of 1934 forced Wall Street's hand.
What's the construction parallel?
Two, and they're both familiar to anyone who's fought with a plan reviewer over a substitution. UL is a private nonprofit that tests electrical and fire-safety products against its own standards and issues a "UL Listed" mark. Nothing in federal law requires a UL listing, but the National Electrical Code references it, most AHJs won't sign off without it, and most insurers price risk assuming it's there. ICC-ES works the same way for proprietary building products and systems: it issues an Evaluation Report (ESR) that tells a plan reviewer a product meets code intent when no prescriptive code language covers it. Neither UL nor ICC-ES started as a legal requirement. Both became one in practice because the people downstream — code officials, specifiers, insurers — started asking for the paperwork before they'd approve anything.
Does this actually touch construction, or is it just tech-industry noise?
Today, it's tech-industry noise — there's no product, no timeline, and the three labs can't agree on the model. But AI is already moving into workflows adjacent to life-safety and code compliance: automated structural checks, AI-assisted plan review, digital-twin-based sequencing, and early autonomous equipment. If a version of Hassabis's FINRA model gets built, the mechanism that made UL and ICC-ES mandatory — a spec writer, an insurer, or a public owner deciding they won't sign off without the certification — is the same mechanism that would eventually pull an AI standards body into a construction contract, the same way CMMC compliance now gates federal defense work.
What should you do with this today?
Nothing urgent, but one habit is worth starting: when you're evaluating an AI tool for structural analysis, code-compliance review, or autonomous equipment, ask the vendor directly whether any independent party has tested the underlying model's safety and reliability, and get the answer in writing. There's no standard to point to yet. But if one of these three proposals turns into an actual body over the next year or two, you'll want a paper trail showing you asked before it was required — the same way a GC who kept ESRs on file for a proprietary connector never has to scramble when a plan reviewer asks for one.
Forward this to whoever on your team is evaluating the next AI vendor pitch.
Friday one chart. Every week, one piece of data that should change a decision on your project. Subscribe at constructionaibrief.com.
- What is the AI standards body that Anthropic, OpenAI, and Google are discussing?
- It's not a real organization yet. Representatives from the three companies have held working-group meetings since July, as recently as last week, about creating a shared body to test and audit frontier AI models before release. Nothing has been formed or agreed to — the labs are still arguing over what shape it should take.
- What's the difference between the FINRA, FAA, and IAEA models being floated?
- Google DeepMind's Demis Hassabis wants a FINRA-style body: industry-funded, government-overseen, starting with voluntary reviews. Anthropic's Dario Amodei wants an FAA-style federal agency with power to block a model's release outright. OpenAI's Sam Altman wants an IAEA-style international forum that certifies countries and companies, not just models. All three disagree on how much teeth it should have.
- Does this change anything for construction right now?
- No. There's no certification requirement, no timeline, and no agreement among the three labs on what body — if any — actually gets built. This is worth tracking, not acting on today.
- What's the construction parallel to an AI standards body?
- UL and ICC-ES. Both are private, industry-funded organizations that test products against a standard and issue a listing or evaluation report — and both became mandatory in practice once code officials, specifiers, and insurers started requiring that paperwork to approve a product. An AI standards body could follow the same path: voluntary at first, then a de facto requirement once someone downstream starts asking for it.
- What should a GC or estimator do about this today?
- Start asking AI vendors selling into structural, code-compliance, or autonomous-equipment workflows whether any independent party has evaluated their model's safety and reliability — even informally. That question costs nothing to ask now and gives you a paper trail if a spec, an insurer, or a public owner starts requiring it later.