OpenAI Staff Fund a PAC Against Brockman's AI Group

OpenAI president Greg Brockman backed a political network seeking to shape AI rules. Seven people still employed by OpenAI gave money to a PAC opposing it. OpenAI says Brockman’s support was personal.

Those positions can all be true. A company need not fund a political organization for its president’s affiliation to matter to the people building its models. In July, Wired reported that the seven employees had given more than $215,000 to Guardrails Alliance, including $200,000 from one alignment researcher. The donations establish dissent by seven workers, not a company-wide vote.

AI policy and regulation now runs through organizations that do more than translate technical concerns for government. They raise election money, press for particular legal boundaries, propose standards, and confer credibility on the people drafting them. AI policy advocacy groups have become a contested governance layer: as their influence over frontier-model rules grows, executives’ outside alliances can create obligations and opposition inside the labs whose interests those alliances claim to advance.

The first industry forums were built to share practice, not win elections

Early industry coalitions answered a coordination problem. Amazon, Facebook, Google, IBM, and Microsoft launched the Partnership on AI to develop best practices; its founders said the nonprofit was not intended to lobby government. In 2023, Anthropic, Google, Microsoft, and OpenAI established the Frontier Model Forum to advance safety practices for powerful models. Neither arrangement made the members politically neutral, but each offered a way to work on common technical questions while competing elsewhere.

That design depended on a distinction between sharing expertise and deciding the rules under which members would operate. Critics raised the concern at the Frontier Model Forum’s launch: an industry body might help its members head off regulation. The concern was not proof that its safety work was empty. It identified a limit in the design. Once lawmakers needed technical standards and model evaluations, the organizations able to supply them could also influence what counted as adequate oversight.

The Center for AI Safety and the Future of Life Institute occupied other parts of this field. OpenAI signed a Center for AI Safety letter expressing concern about AI risks; the Future of Life Institute later drew signatories as different as Steve Bannon, Susan Rice, Ralph Nader, and Daron Acemoglu to its Pro-Human AI Declaration. Those groups are not a demonstrated coordinated network with the frontier labs or their PACs. Their different constituencies show why “AI safety” cannot stand in for a single policy position.

Political money turned the boundary into a ballot question

Leading the Future launched in 2025 with more than $100 million in backing from Andreessen Horowitz, Brockman, and others, explicitly looking to crypto political network Fairshake as a model. By January 2026, Axios reported that it had raised more than $125 million to influence the midterms and federal AI policy. Leading the Future pushed for a uniform national approach over a patchwork of state laws and sought to limit more restrictive federal regulation. It also ran attack ads against Alex Bores, a congressional candidate identified with AI regulation. A technical-policy argument had acquired a candidate and an ad budget.

Leading the Future was not entering an empty room. More than 500 organizations lobbied the White House and Congress on AI in the first half of 2025; reporting put OpenAI’s lobbying spending at $1.8 million, up from $380,000 in 2023. In the coverage record, Leading the Future appeared in one article in 2025’s third quarter and 14 in 2026’s second quarter. That is a measure of coverage, not of the group’s effectiveness. It does show how quickly a newly launched political vehicle became a recurring part of the policy fight.

Rival funders chose different legal boundaries. Anthropic opposed a congressional measure that would bar states from regulating AI. It also gave $20 million to Public First Action and later committed another $20 million; Public First Action called itself a counterweight to Leading the Future. The disagreement reaches beyond the word “safety.” A federal rule that overrides state measures, a state’s authority to impose safeguards, and an independent evaluator’s power to stop deployment would place decisions in different hands.

Workers used political voice before they had an internal vote

Guardrails Alliance received more than $215,000 from seven current OpenAI employees who wanted stricter AI safeguards and opposed the Brockman-backed group. Their personal donations did not give them formal authority over OpenAI’s political posture. Their technical roles did give their objection a particular weight: they could argue about the risks of systems they helped build, rather than only about the company’s public reputation.

Employees had challenged frontier-lab governance before this PAC dispute. In 2024, current and former OpenAI and Google DeepMind staff warned of recklessness and secrecy at frontier AI companies and called for stronger ways to raise concerns. Political counter-funding added another channel. Workers could challenge a policy position inside the company and support its opponents outside it without first leaving their jobs.

In June 2026, OpenAI employees pushed back in Slack against Brockman’s Leading the Future ties. Reporting connected that pushback to his decision not to proceed with a planned $25 million donation. Brockman, who had already donated $25 million, subsequently said he had no plans to give more because the PAC had become a distraction for OpenAI employees. Neither the seven employees’ donations nor the Slack pushback establishes an employee veto or workforce-wide opposition. It does document an executive changing course after his outside affiliation became an internal issue.

Private safety machinery still needs someone outside the room

Labs have also built less electoral forms of governance. OpenAI adopted a structured safety-case framework for frontier reinforcement-learning training, drawing on aviation and nuclear practice. A safety case records why an operator believes a high-consequence activity can proceed; it does not, by itself, decide who may reject that argument. OpenAI, Anthropic, and Google have explored an industry-led safety standards body without government oversight, with a launch reported as a plan for late 2026 or 2027, not an accomplished fact.

Private coordination has produced concrete work. OpenAI, Anthropic, and Google used the Frontier Model Forum to share information about adversarial attempts to distill their models. That exchange can help the companies identify prohibited activity, even as it protects competitive and security boundaries. Political disagreement among the labs does not make every shared standard a bid to capture regulation. It does make the scope of a standard—and who gets to inspect its application—important.

Public evaluation has not disappeared. OpenAI and Anthropic agreed to give the US AI Safety Institute early access to major models for testing, and OpenAI has called for mandatory government-led cyber-risk evaluations of advanced systems. Anthropic’s Dario Amodei has called for third-party tests with authority to block or revoke deployments posing catastrophic risk. In the emerging regime of US frontier-model oversight, a lab’s documentation and a government evaluator’s judgment are different checks, not interchangeable names for the same one.

“Personal” describes the donor, not every consequence

OpenAI says it has not donated to any super PAC, does not operate an employee-funded PAC, and neither directs nor has visibility into Leading the Future’s operations. It describes Brockman’s involvement as personal. That distinction matters: a president’s donation should not be reported as a company payment, and seven employees’ donations should not be reported as company policy.

But a legal boundary around the money does not settle who bears the organizational cost. When an executive funds a group seeking rules that will affect his employer’s models, employees and the board still have reason to ask what the affiliation represents, what has been disclosed, and where staff can contest it. OpenAI has said its board can hold back a model release even if leadership considers the model safe. Political advocacy raises a different decision: who, if anyone, can authorize a senior executive to lend the company’s credibility to a campaign while calling the campaign personal?

OpenAI put “personal” on its policy page. Brockman put “distraction” in an employee Slack message beside a $25 million limit.