On September 14, 2026, CNN and The Information reported that OpenAI, Anthropic, and Google have been holding closed-door working-group meetings since July 2026 to establish a joint AI standards body — modeled on FINRA, the Financial Industry Regulatory Authority that governs U.S. broker-dealers.
The proposal, as described in leaked meeting notes and subsequent reporting, would create an industry-funded body with three core functions: pre-deployment safety testing of frontier models, independent third-party audits of training and deployment practices, and a shared incident-reporting framework for model failures. Sam Altman, Dario Amodei, and Demis Hassabis have all signed off on the framework in principle. Elon Musk — whose xAI was not invited to the initial working group — publicly responded to Amodei’s September 12 essay “We Must Pace the Frontier” with two words: “Dario is right.”
This is the most serious attempt at AI industry self-regulation since the 2023 White House voluntary commitments. It is also the most likely to fail — and the reasons it might fail reveal something important about whether the AI industry can self-govern at all.
Why This Is Not 2023 Again
The 2023 White House commitments — signed by OpenAI, Anthropic, Google, Meta, Amazon, Inflection, and Stability — were the first serious attempt at AI industry self-regulation. They failed for three reasons:
- They were post-release, not pre-release. The commitments required companies to test models after training but did not require pre-deployment review by any external party. By the time a model was tested, it was already being deployed.
- They had no enforcement mechanism. A voluntary commitment, with no audit, no sanctions, and no disclosure, is a press release — not a regulatory framework.
- They did not cover open-weight models. Meta’s Llama and Mistral’s models fell outside the framework entirely, creating a structural loophole that Meta actively exploited for competitive advantage.
The 2026 proposal addresses the first two of these directly. Pre-deployment testing — with the testing body having authority to delay a release — is a structural shift from the 2023 approach. Independent audits, with the authority to publish findings regardless of company consent, are the missing enforcement mechanism. The third issue — open-weight coverage — remains unresolved and is the proposal’s most contested fault line.
The FINRA Model: What It Actually Means
The FINRA comparison is more substantive than it first appears. FINRA is not a government regulator. It is a self-regulatory organization (SRO) — an industry-funded, industry-governed body that Congress has delegated specific regulatory authority to. Its key features:
- Industry-funded, not taxpayer-funded. Member firms pay fees that finance FINRA’s operations. This insulates it from Congressional appropriations battles.
- Industry-governed, with public representation. FINRA’s board includes both industry executives and public members. The public members are a structural check on industry capture.
- Delegated government authority. FINRA can fine, suspend, and bar individuals and firms. Its decisions are reviewable by the SEC, but the SEC rarely overturns them.
- Pre-clearance authority. FINRA reviews new products before they reach the market. A broker-dealer cannot list a new structured product without FINRA sign-off.
The AI industry proposal, as described, mirrors these features closely: industry funding, mixed governance, pre-deployment review authority, and delegated government oversight. It is, in effect, an attempt to replicate the SRO model for AI — and to do so before Congress legislates a worse, more prescriptive framework.
Why It Might Work
Three things are different in 2026 that make this proposal more likely to succeed than the 2023 commitments:
The competitive pressure has shifted. In 2023, the three labs were largely aligned on safety rhetoric but diverged in practice — OpenAI shipped aggressively, Anthropic positioned itself as the safety leader, Google shipped cautiously but publicly. In 2026, the competitive dynamic is different: DeepSeek’s V4.1 Flash has made aggressive shipping the only viable competitive strategy. The three Western labs now face a common external threat that makes coordination on safety more attractive than it was when they were primarily competing with each other.
The regulatory threat is concrete. In 2023, the AI industry faced a diffuse regulatory threat — various Congressional proposals, state-level laws, and EU AI Act negotiations. In 2026, the regulatory threat is specific: the EU AI Act is in force, California’s SB 53 has been signed, and there are credible signals that federal preemption legislation is coming. An industry SRO is a defensive move to preserve industry control over the regulatory framework before Congress writes it for them.
The technical infrastructure exists. In 2023, pre-deployment review was not technically feasible — there were no standardized safety evaluations, no auditable training records, no shared incident-reporting formats. In 2026, those things exist: METR’s model evaluation framework, MLCommons’ safety benchmarks, and Anthropic’s published responsible scaling policies provide the technical scaffolding an SRO could use on day one.
Why It Might Fail
The reasons the proposal might fail are more interesting — and more specific to the AI industry’s structure:
The Open-Weight Loophole
The proposal does not, as currently described, cover open-weight models. Meta’s Llama, Mistral’s models, and DeepSeek’s V4.1 Flash (released as open weights on Hugging Face) would all fall outside the SRO’s jurisdiction. This is the same loophole that crippled the 2023 commitments — and it is a bigger loophole now than it was then, because open-weight frontier models are more capable in 2026 than they were in 2023.
The labs’ argument for excluding open weights is that they cannot control what downstream users do with open models. That argument is technically correct but strategically self-serving: the labs that ship open weights (Meta, DeepSeek, Mistral) gain a competitive advantage over the labs that ship closed weights (OpenAI, Anthropic, Google) — because the open-weight labs do not bear the compliance costs the SRO would impose.
Until this loophole is closed, the SRO will function as a cartel enforcement mechanism — imposing costs on closed-weight labs that open-weight competitors do not bear. That is not self-regulation. It is competitive positioning dressed up as safety governance.
The Enforcement Problem
FINRA works because its sanctions are credible: a broker-dealer that loses FINRA membership cannot operate. The AI industry’s equivalent sanction — revoking the right to deploy frontier models — has no legal basis. The labs cannot grant each other the authority to bar competitors from deploying models; that would be a per se antitrust violation.
Without a credible sanction, the SRO devolves into a testing service: it can publish findings, but it cannot stop a release. And a testing service with no enforcement authority is, from a regulatory perspective, indistinguishable from a trade association. The labs know this. The regulators know this. The question is whether the labs are willing to accept legally binding pre-deployment review — which would require Congressional delegation of authority — or whether they will settle for voluntary pre-deployment review that is structurally weaker than FINRA.
The Geopolitical Asymmetry
The SRO, as proposed, is a Western SRO. DeepSeek, Kimi, and the Chinese AI labs are not invited. This means the SRO’s pre-deployment review would apply only to Western frontier models — while the most aggressively shipped frontier model of 2026 (DeepSeek’s V4.1 Flash) is Chinese and open-weight.
This is not a minor defect. It is a structural asymmetry that will, over time, erode the SRO’s legitimacy. A safety framework that applies only to the labs that are already shipping cautiously — while exempting the labs that are shipping aggressively — is not a safety framework. It is a uncompetitive tax on the labs that are playing by the rules.
What Comes Next
The working group is expected to publish a framework document by the end of 2026, with operational launch targeted for mid-2027. Between now and then, three decisions will determine whether the SRO succeeds or fails:
-
Does Congress delegate authority? Without statutory authority to enforce pre-deployment decisions, the SRO is a testing service, not a regulator. The labs’ lobbying strategy — give us self-regulation before you give us legislation — depends on Congress being willing to delegate. If Congress is not, the SRO is dead on arrival.
-
Does the framework cover open weights? If Meta, Mistral, and DeepSeek are excluded, the SRO is a competitive cartel, not a regulatory framework. The labs that want to ship aggressively will simply ship open weights and bypass the SRO entirely.
-
Does DeepSeek participate? A Western-only SRO, in a world where DeepSeek is setting the pace of frontier releases, is structurally irrelevant. DeepSeek’s participation would require either a parallel Chinese SRO (which does not currently exist) or a bilateral framework that bridges U.S. and Chinese regulatory systems (which is politically impossible given current tensions).
The Real Question
The proposal, for all its limitations, is the most credible attempt at AI self-regulation to date. It is being driven by labs that have the technical capability to make pre-deployment review meaningful, and it is being designed with awareness of the failures of the 2023 commitments. It may work. It is more likely to fail than to succeed.
But the real question the proposal raises is not whether the SRO will succeed. It is whether the AI industry is capable of self-governance at all — whether a group of competitors, each racing to ship the next frontier model, can collectively impose costs on themselves that their individual incentives push them to avoid. The FINRA model works for broker-dealers because they operate in a mature industry with stable competitive dynamics. The AI industry has neither of those things. It is an industry in active technological disruption, where a six-month delay can mean the difference between market leadership and irrelevance.
The labs’ answer to that question, in the form of this proposal, is: yes, we can self-govern, because the alternative is worse. That is the same answer every industry has given when facing regulatory pressure. It has sometimes been true. It has often not been. The AI industry is about to find out which one it is.
