The irony of the AI industry’s massive lobbying push

As market actors, AI companies are all doing something pretty similar, to similar ends, with similar needs: building models at great expense and trying to market them to users and sell them to customers. There’s a lot of uncertainty about where their collective project is going, of course, but on a basic level, they’re all trying to get access to the same resources — compute, data, and engineering talent — to climb the same curve, and to turn their models into something they can charge for.

As political actors, though, they’re drifting in different directions. In the last two years, some of them (OpenAI, Meta, and xAI) publicly aligned with the Trump administration, while others steered away (Anthropic, though a legal dispute with the Department of Defense; DeepMind, through leaders who spoke up about ICE shootings). Recently, and more intentionally, they’ve been asserting themselves as political donors and lobbyists. This spending, which has been ramping up in recent months, suggests that the people running AI firms actually want considerably different things from the political system. It also suggests that some of their own employees think they’re crazy.

It’s helpful to work back from some recent news. Last week, Anthropic CEO Dario Amodei donated $1 million of his own money to Public First, a super-PAC created to “educate Americans on key AI issues and advance an AI policy agenda supporting safeguards,” including model transparency requirements, a federal AI-governance framework, export controls on AI chips, and regulation around catastrophic AI risks. As a company, Anthropic had previously given $20 million to Public First Action, though this was purportedly earmarked for “public education,” not direct political contributions. (Public First recently spent big to support Democratic pro-AI-regulation candidate Alex Bores in an unsuccessful run for Congress in the midtown Manhattan district, but it also supports Republicans who have spoken out in support of AI regulation — or at least for export controls — including Josh Hawley and Tom Cotton.) Some of Anthropic’s employees, as well as employees at OpenAI and DeepMind, have contributed as well.

Public First is intended as a counterweight to the industry’s biggest PAC, Leading the Future. Leading the Future is a bit less forthright about its specific policy priorities, but given its donation patterns — it opposed Bores, for example — as well as its structure, backers, and staffers, it can be understood as broadly anti-regulatory and intended to get ahead of regulations that could result from the growing public backlash against data centers. It’s modeled after Fairshake, the Andreessen Horowitz–backed pro-crypto PAC, and shares some personnel. In a personal capacity, OpenAI president Greg Brockman gave $25 million to LTF around the same time he gave to a separate Trump-aligned super-PAC. Other contributors include a co-founder of Palantir.

There are other PACs — Guardrails Alliance was created to explicitly oppose Leading the Future and has attracted donations from some safety-concerned AI workers, while Meta has launched its own LTF-like PAC. These two can provide a decent sense of the shape of the industry debate: While both organizations say they’re in favor of federal AI regulation, they mean very different things. For Public First, this would mean that the federal government would become an assertive regulator with an eye on longtime AI safety concerns; for LFT, this would mean that the federal government would claim exclusive jurisdiction over AI regulation, preventing states from regulating AI companies and then presumably not doing a whole lot to regulate them itself. (This is more or less what’s happening with prediction markets now.) Public First’s position, in contrast, is that state regulation can be a force for good, at least in the absence of federal oversight. Put another way: Dario Amodei would prefer a situation in which states like New York and California set de facto national AI rules by extracting regulatory concessions from companies that wish to operate there, given that the Trump administration is running things at the federal level for the next few years; Greg Brockman and Mark Zuckerberg, on the other hand, would like the Trump administration to intervene on their behalf to prevent states from doing almost anything, under the assumption that the Trump administration won’t ask too much of them either.

The fact that there’s real ideological disagreement between industry-backed AI PACs is genuinely interesting, as are the gathering examples of employees at places like OpenAI splitting from their firms on ethical grounds. But for all the intra-industry drama — and despite the broad talk about being “pro-innovation” or concerned with “existential risk” — the political imagination on display here is fairly narrow. The anti-regulatory team, which set its policy agenda at the beginning of the second Trump term, has since seen the administration take an active interest in if not regulating AI, exactly, then intervening in the industry severely, arbitrarily, and on flexible national-security grounds. Meanwhile, figures like Amodei, while presenting themselves as realists, can seem naïve about their own prospects, and roles, within the American political system. If they’re right, and comprehensive regulation is indeed necessary to mitigate runaway AI risk or deal with employment shocks, it’s hard to imagine that mass politics will mobilize behind the less bad AI firms, or their candidates, rather than behind a politics that treats all of them as problems that need to be addressed. Just look at the data-center backlash, which is already producing state-level legislation, including in New York, that no compute-hungry AI company is going to support.

From outside the industry, in other words, these efforts aren’t quite as far apart as they might feel from the inside. The emerging range of plausible political outcomes around AI is widening by the day — there’s already bipartisan talk of public firm ownership, for example — and will only widen should the companies become more powerful. A world in which AI firms are ripping through the economy is one in which AI backlash will slip well past the control of a few PACs; similarly, a situation in which AI firms are producing genuine threats to state control, as Amodei has warned about at length, is one in which options well beyond thoughtful regulation are on the table no matter who is in charge.

Nowhere is this clearer than on the issue where the industry seems mostly aligned: China. The anti-regulatory side holds an exception for China, and in fact justifies its entire project as necessary to allow American companies to stay ahead. The pro-regulatory side justifies its support of export controls in safety terms, arguing that they’ll copy American frontier models but release them with fewer safeguards. Maybe this is true, but the goal is functionally the same: to get the United States government to step in and protect their firms’ current lead in an increasingly competitive market. The problem is that this can’t really work for long. Chinese AI labs have indeed drafted behind American research but are becoming increasingly capable on their own (last month, multiple Chinese firms released models that were competitive with American AI coding products at a fraction of the price; last week, a firm called Moonshot AI released a model, Kimi 3, that early tests suggest is near the American frontier). They’re also shifting more compute to Chinese hardware, which means that the window during which hardware export controls could even work is closing fast.

From within the AI world, the gulf between the political projects of the different labs feels enormous. Soon enough, though, these differences could easily be overwhelmed: by the trajectory of the technology itself or by a public that doesn’t want to hear from any of them.