Greg Brockman’s $50 million bet on AI politics was never a policy bet. The OpenAI president walked away from a second $25 million pledge to Leading the Future without publishing what the money was technically for — and that is the part of this story worth reading closely, not because the pressure campaign failed to work, but because neither coalition in this fight has produced a specification for the thing they are spending to control.

It is true that the pressure worked, in the narrow sense in which pressure campaigns are usually described as working. OpenAI employees attacked the first donation as an attempt to “thwart sensible AI regulations”; by July several were personally channelling more than $200,000 to the Guardrails Alliance; the Alliance publicly claimed credit for the reversal; and OpenAI itself distanced the company from its own president’s pledge in June. By then, AI super PACs were already tracked at $43.3 million in midterm spending, with Leading the Future the gravitational centre of the industry-friendly field. The trouble is that “it worked” describes an electoral effect, and the question a technology column has to ask is what the money was buying on the mechanism layer. The documentary record answers: nothing either side has specified.

Start with the state law at the centre of the most expensive fight in the cycle. An affiliate of Leading the Future spent more than $8 million against assemblymember Alex Bores, a former Palantir employee who co-wrote one of the country’s strictest state AI laws, and the Manhattan primary became the proxy fight both sides had been building toward. Bores finished second. But the bills in that legislative family share an architecture, and it is worth reading the architecture rather than the spending. They bind developers of models above a stated capability or compute threshold — frontier models, meaning the largest general-purpose systems trained at the greatest cost — and they typically require three things: a published safety framework, safeguards against unauthorized access, and incident reporting after a serious failure. Translated: a company that crosses the line has to write down how it intends to keep the model from being misused, lock it down, and tell someone when something goes wrong.

That architecture leaves a specific, well-understood gap. Thresholds in these bills attach to the original developer, not to the capability that emerges downstream. Release an open-weight model — meaning you publish the model’s learned parameters so anyone can run and modify it locally — below the stated line, and a downstream user can fine-tune it, meaning further training on a smaller dataset to adapt the model to a narrower task, until it does things the originating developer never tested for. No obligation attaches at any point in that chain. Applied services built on a third-party model — consumer apps, agentic systems, enterprise deployments — sit outside the regime entirely, because the bill governs model developers, not model users. A state frontier-safety law can be the strictest in the country and still constrain roughly one category of actor: the closed-model shop at the top of the stack, at the moment before release.

Now read the Guardrails Alliance against that architecture. Its public output, as far as the record shows, contains no compute threshold, no evaluation protocol, no capability metric, and no incident taxonomy. What it contains is a claim — that the Alliance has been “leading the charge to prove Leading the Future is politically toxic and a liability” for public officials who take its support — and a claim of credit for Brockman’s withdrawal. That is an electoral position, not a safety position. There is a plain engineering name for the failure mode: a proposal that specifies no observable model behavior, that no test could refute, is not a safety claim at all — it is an unfalsifiable one. And the political utility of unfalsifiable safety language is precisely that it can be deployed against any donor, any candidate, or any model without ever having to say what would count as safe enough. It is the oldest instrument in the manufactured-doubt playbook: a standard of proof no test can satisfy, so the claim can never lose. Contrast the technical work happening elsewhere — the frontier development slowdown proposal major AI CEOs backed last month carries actual mechanisms, staged development commitments and capability evaluation before deployment, whatever one thinks of them. One can disagree with a threshold and still recognise a threshold. The Guardrails Alliance has published neither.

Symmetrically, the position Brockman’s money was funding is equally under-specified. He and Anna Brockman said last year they supported “AI centrism”: federal regulation to promote innovation, “defend systems from misuse,” and protect “the privacy of conversations with AI,” while keeping “most developers and open-source models, and almost all deployments of today’s technology” free from additional regulatory burden. In May he declined to discuss which of the group’s spending decisions he disagreed with. Taken at face value that is an architecture, and it is worth assessing on engineering terms rather than as a spending story. “Defend systems from misuse” is a post-deployment frame: it asks what happens after a model is in the world, not what capability crosses the line before it gets there. “Privacy of conversations with AI” is a data-handling obligation — a real one, and an entirely different problem from capability governance. The open-source and most-deployments carve-out is the load-bearing clause, and it removes from the regime the single layer where capability actually diffuses. An engineering-credible centrism would need, at minimum, a threshold that survives capability diffusion when someone fine-tunes an open-weight base model; named third-party evaluators with pre-deployment access to the actual system; published evaluation results rather than private assurance; and an incident-reporting taxonomy with defined categories. “Minimal additional regulatory burden for most deployments” is none of those things. It is a promise about the shape of a law that has not been written, made by the people who would fund it.

So both coalitions are spending on positions neither has technically specified, and that is the finding — not the finding either press release wants. The industry-friendliness coalition spent more than $8 million to defeat a legislator and could not, in public, describe the regime it prefers in place of his. The safety coalition claimed credit for killing a $25 million pledge and could not, in public, describe what model behaviour it would regulate. What the money buys, before any specification exists, is definitional authority: the right to decide later what “safe enough” means, with the other side already arguing from a conceded position. A Guardian editorial in August named the distortion without the partisan gloss — AI money is not moving the technical debate, it is colonising the venue where the technical debate would happen.

That reading survives the corporate distancing, which is the part of the record most often waved away. OpenAI’s June blog post stated that the Brockmans made their contributions “in a personal capacity” and that the company did not direct the PAC’s activities or “have visibility into their operations.” On its face that is defensive paperwork. On the engineering reading it is more interesting: the country’s leading AI company could not, or would not, describe what its own president’s political vehicle was technically for. No visibility into the operations of the vehicle it was publicly associated with. A company that cannot specify the policy position it is funding is funding the spending, not the position. It is also worth noting, because the documentary discipline requires it, that News Corp, owner of the publication that broke this story, has a content-licensing partnership with OpenAI.

The policy lever here is neither electoral nor personal. State law should define the frontier by capability and compute rather than by developer identity, so a threshold follows the model up the chain when somebody fine-tunes it; should mandate named third-party evaluators with pre-deployment access and published results, so assurance is checkable rather than asserted; and should extend obligations to applied and agentic deployments on the basis of what a system can do rather than who built it. Those three provisions would make both of this year’s coalitions uncomfortable, which is the test that separates a specification from a spending program.

Brockman’s withdrawal will be remembered as the moment the money blinked before the mechanism did. A super PAC is a spending instrument; it is not a specification. The specification is still missing, and it will be written in the next session whether or not either coalition has one — by whoever arrives at the drafting table with a metric in hand, because that is the only thing a drafting committee can actually vote on.