The four most powerful CEOs in the most consequential technology sector on earth have done something no one should mistake for courage: they have asked to be the ones who define safety, choose the evaluators, and set the rules — while disbanding their own safety teams and spending millions to kill the legislation that would impose real oversight.
That is not a request to be slowed down. That is a bid for self-governance dressed in the language of public concern. Amodei’s earlier three-point plan established the framing; this weekend’s consensus made it operational. The CEO of Anthropic, Dario Amodei, published a proposal on Saturday titled “We Must Pace the Frontier.” Within hours, Sam Altman of OpenAI, Demis Hassabis of Google DeepMind, and Elon Musk of xAI publicly endorsed it. “Committing to having independent evaluators with employee-like access is a great idea,” Altman posted. Musk wrote simply: “Dario is right.”
The speed of that consensus should be the first exhibit in any honest reading of the proposal. Four companies that have spent the last three years in a frantic race to ship the most powerful model first — four companies that have never coordinated on anything — agreed on the language, the structure, and the public posture within hours. The document was either circulated in advance or the alignment was already in place. Either way, the unity is the data, not the argument.
And the argument, on its merits, is a case for the labs choosing their own safety architecture.
Start with the cui-bono trace, because it is the thing Amodei’s supporters are working hardest to obscure.
Amodei’s proposal has three pillars. The first commits the company to outside evaluators with employee-level access. The language sounds like oversight. It is not. The companies choose who those evaluators are. They grant the access. They define the scope. The evaluator is hired by the entity it evaluates, reports findings to the entity it evaluates, and operates under nondisclosure terms written by the entity it evaluates. That is not independent oversight. That is an internal audit with a public label.
The second pillar calls for agreements between democratic states on AI safety standards. The principle is sound. But in the current political environment — where Congress has failed to pass any comprehensive AI legislation, where the House Speaker has already declared that Congress will not lead the charge on safety — what the proposal actually calls for is executive-action safety standards negotiated between governments, not statutory regulation enacted through a democratic process with public accountability. Congress is bypassed by design.
The third pillar calls for global coordination with authoritarian governments such as China on security issues. This is the pillar that has drawn the most bipartisan skepticism, and it deserves the most scrutiny. The proposal would require federal coordination authority — authority the Trump administration does not currently possess in any enforceable form. It would also, as former Biden-era FTC official Alvaro Bedoya and former Trump AI czar David Sacks both flagged, create a framework in which incumbent AI labs coordinate on standards under federal cover while new entrants face the full weight of those standards without the antitrust protection the incumbents would enjoy.
“Antitrust law does not prevent AI companies from coordinating to make sure AI does not hurt people,” Bedoya posted. “Antitrust law does absolutely prevent AI companies from organizing to prevent the entry of cheaper, upstart rivals because the bigger companies are burning cash and failing to achieve sufficient profitability.”
That is not a minor objection. It is the structural core of what the proposal actually asks for.
The labs with trillion-dollar IPOs in their pipeline — OpenAI’s delayed to 2027 — are asking for a regulatory framework that their own antitrust counsel would have to write. The concentrated beneficiary is the incumbent industry leadership. The diffuse cost-bearer is every startup, every open-source researcher, and every competitor that cannot afford the compliance infrastructure the incumbents would design.
Alex Bores, whose failed Congressional bid drew $24 million in AI-linked PAC spending both for and against his campaign, put it plainly: “To everyone who genuinely wants real AI regulation, please don’t fall for their games.”
Now stack the receipts, because the proposal’s supporters want you to treat the record as context. The record is the indictment.
In 2024, OpenAI shut down its “superalignment” team — the internal group dedicated to studying long-term AI risks from the technology the company was building. This year, the company disbanded another team intended to communicate how OpenAI would benefit humanity. Altman has since delayed the company’s IPO to 2027, citing safety concerns, while the company posted its most profitable quarter. Safety is the stated reason for delay. The delay also buys the company time to consolidate market position before inviting public scrutiny of its revenue model.
Those are not background facts. They are the behavioral record of the entity asking to define safety.
Anthropic has issued repeated warnings that artificial intelligence could kill all of humanity by 2030. The company has also continued building and selling the technology it warns could end human civilization. That is not a contradiction that resolves in Amodei’s favor. It is the operating posture of a company that wants the credit for urgency and the revenue of continued deployment, and the proposal gives it both.
The industry killed prior legislation through lobbying. The companies have spent millions shaping favorable regulation. The result, as Sarah Myers West of the AI Now Institute put it, is that “AI exceptionalism has led to a total failure to uphold our standards for what products are deemed safe for use.” The industry has spent years arguing it is too powerful, too complex, and too lucrative to regulate — and now, at the moment the political environment most favors capture, it is proposing to regulate itself.
The financial incentives complete the picture the proposal’s language is designed to obscure.
OpenAI’s IPO delay, framed as a safety decision, also functions as a consolidation strategy. The company posted its most profitable quarter in the same cycle it announced the delay. The delay buys time to establish a dominant market position — to ship the models, secure the customers, and build the revenue narrative that will underpin a public offering in 2027 rather than 2026. Meanwhile, the “safety” framing lets Altman position himself as the responsible steward of a technology too dangerous to rush to market — while continuing to ship it.
Amodei’s proposal, released in the same news cycle, gives Altman and the other CEOs a public framework for the delay narrative. They are not slow because the business model requires consolidation. They are slow because safety requires it. The proposal is the cover story for the revenue strategy.
That is not speculation. That is the structure. Follow the benefit.
The China question makes the capture framework operational.
Amodei’s proposal calls for coordination with authoritarian governments on AI security. The Trump administration has framed AI development as a zero-sum contest against China for technological supremacy. Vice President JD Vance declared last year that the future of AI was “not going to be won by hand-wringing about safety.” On Sunday, in Ireland, President Trump told reporters: “whoever wins AI, wins.”
That line is not a contrast to the CEOs’ framing. It is the CEOs’ framing. The industry’s own public advocacy for years has been that AI development is a race China must not win, and that safety regulation is the obstacle to victory. Amodei’s proposal accepts that framing entirely. It asks for coordination with China — not to slow the race, but to manage it. The assumption underneath the entire structure is that the race continues. Safety is a friction surface to be navigated, not a condition to be met before deployment proceeds.
Vance’s line — “not going to be won by hand-wringing about safety” — and Amodei’s proposal are the same argument at different temperatures. Both treat safety as a constraint on competitive advantage. Both assume the technology will be built and deployed regardless. The only question is who controls the terms. Amodei’s answer is: the labs.
House Speaker Mike Johnson has already said Congress will not lead on AI safety. The Trump administration has framed the technology as a geopolitical weapon. The labs are proposing self-governance. The entire architecture converges on one outcome: the companies that build the technology define what safe means, and no institution with democratic accountability holds the veto.
King, at Riverside in April 1967, named the giant triplets of racism, extreme materialism, and militarism, and he warned that a nation which continued year after year to spend more on military defense than on programs of social uplift was approaching spiritual death. The AI race is the new expression of that same triplet. The technology is being built as a weapon in a geopolitical contest the government has declared existential. The profits flow to a concentrated set of incumbent firms. The costs — the safety risks, the displacement, the concentration of power in entities accountable to no democratic institution — fall on everyone else. And the people building the weapon are the ones telling the government not to slow them down, while simultaneously proposing to define the rules under which they will continue.
The critics Amodei’s supporters want you to dismiss — Bedoya, Sacks, Bores, Myers West — are the voices naming the structural pattern. They are not undermining a good-faith effort. They are reading the proposal the way the proposal should be read: as a document written by the entities it would regulate, asking for the authority to choose their own evaluators, coordinate with their own competitors under antitrust cover, and bypass the legislature that would impose democratic accountability.
The industry has not earned the presumption of good faith. The record — disbanded safety teams, $24 million in PAC spending against a candidate who proposed real oversight, years of lobbying against statutory regulation, an IPO delay framed as safety while the company posts record profits — does not support it.
Build the oversight. Build it in statute, not in executive agreement. Build it with evaluators chosen by a public process the labs do not control. Build it with antitrust enforcement that prevents the incumbents from writing rules that protect their market position. Build it with a legislative process that makes the Trump administration’s “whoever wins AI wins” rhetoric face a democratic counterweight.
The companies have asked to define safety.
The public’s job is to make sure they do not get to.