title: Buying Both Doors subtitle: The AI industry isn't lobbying Congress. It's building Congress — one primary at a time, in both parties, with ads that never mention AI. author: Rebecca Rowan publication: For the Republic date: 2026-03-15

Buying Both Doors

A voter in Chicago's northwest suburbs sees an attack ad. It's about ICE. Then another -- this one about healthcare costs. Then a third, hitting a candidate for being soft on Trump. Standard midterm fare. The issues track. The outrage tracks. Nothing about these ads would make you look twice.

Now pull the FEC filings behind them. Nearly $62 million has poured into four Illinois Democratic primaries -- and of that, $26.9 million comes from AI, crypto, and AIPAC PACs. Leading the Future -- the AI industry's flagship super PAC -- transferred $5 million to Think Big, its Democratic arm, and $5 million to American Mission, its Republican arm. Split right down the middle. Meta spent $65 million building the same architecture from the corporate side: Forge the Future for Republicans, Making Our Tomorrow for Democrats.

Not one of those ads mentions artificial intelligence. Not one.

The ads are about what voters care about. The money is about what the industry needs. That gap is the tell.

Hedging the Bracket

There's a word for what's happening here, and it isn't lobbying.

Traditional lobbying persuades legislators after they're elected. What the AI industry is doing operates upstream -- selecting candidates before they're nominated. It has a name: bilateral capture. Fund candidates in both parties' primaries so that regardless of which party wins a given seat, the winner was pre-approved by the industry that needs to avoid regulation.

First, the dual-PAC architecture. Leading the Future doesn't lean one way and hope for the best -- it runs formally connected PACs for each party, splitting funds with the precision of a hedge fund rebalancing a portfolio. Meta built a separate but identical structure. This isn't coincidence. It's financial engineering designed so that the partisan outcome of any given race is irrelevant to the policy outcome.

Second, the ad misdirection. As NBC News documented, both pro- and anti-regulation AI PACs run ads about immigration, healthcare, and Trump -- never about AI. Why? Because polls consistently show 69-80% of Americans support greater AI regulation, depending on how the question is asked. The industry's actual agenda is wildly unpopular, so the transaction has to be laundered through issues voters already care about. Brad Carson, co-founder of the pro-regulation Public First Action, said it plainly: "We know AI isn't the first thing on every voter's mind when they go to the polls."

Third, the harvest. And the harvest is already coming in -- before the 2026 elections have even happened. In September 2025, the venture capital firm Andreessen Horowitz published a dormant commerce clause argument for preempting state AI laws. Three months later, that argument became the legal backbone of a Trump executive order creating an "AI Litigation Task Force" to challenge state regulations. A VC firm's legal theory became federal policy in ninety days. Meanwhile in New York, Governor Hochul proposed replacing the entire text of the RAISE Act -- the state's landmark AI safety bill -- with weaker, industry-preferred language. Ron Conway, a Leading the Future donor, had held fundraisers for Hochul while opposing the bill for months. After a public standoff, Hochul signed a compromise version that was substantially weakened from the legislature's original -- not the wholesale gutting her initial redline proposed, but proof enough that the money bought access to the negotiating table.

Now -- the industry's stated policy preference isn't crazy. Over 1,200 state-level AI bills were introduced in 2025, creating a genuine compliance nightmare. One startup, PerceptIn, budgeted $10,000 for compliance and spent over $344,000 before going out of business. (Ten thousand to three hundred and forty-four thousand. Read that again.) The argument for a coherent federal framework over fifty contradictory state regimes is substantively defensible. It's the same position we take on pharmaceutical approval and financial regulation.

But the appropriate response to regulatory fragmentation is democratic deliberation over a federal framework -- not industry pre-selection of the legislators who will write it. The policy preference is reasonable. The method is the problem.

Someone Already Ran This Experiment

If this sounds theoretical, it isn't.

Fairshake, the crypto industry's super PAC, spent roughly $290 million across its network of three aligned super PACs in the 2024 cycle, using the same bilateral architecture: Fairshake itself for both parties, Defend American Jobs for Republicans, Protect Progress for Democrats. The results were not abstract. Sherrod Brown -- the banking committee chair who had blocked crypto-friendly legislation -- was defeated after the industry poured $40 million into the race. The GENIUS Act stablecoin legislation passed. Industry-friendly regulators were confirmed.

And the operative who helped advise Fairshake's Democratic-side strategy? Josh Vlasto, former aide to Chuck Schumer. He's now co-strategist at Leading the Future. Same person, same playbook, new industry. The bilateral model isn't being invented for AI. It's being replicated.

Then there's Alex Bores. A former Palantir data scientist who left the company in 2019 when it renewed its ICE contract. He became a New York state assemblymember, co-authored the RAISE Act -- the strongest state AI law in America -- and earned a spot on Time's 2025 AI 100 list. Now he's running for Congress. And Think Big, Leading the Future's Democratic-aligned PAC, has spent over $1.8 million in negative ads against him. The money comes from a PAC backed by Palantir co-founder Joe Lonsdale. The attack ad's message? "Alex Bores' tech company works for ICE." Let that sink in. The industry he tried to regulate is using his own resume as a weapon against him -- and the specific irony of a Palantir co-founder funding attacks about Palantir's ICE work against the guy who quit Palantir over its ICE work is the kind of thing you'd cut from a screenplay for being too on-the-nose.

But here's the thing about money and elections. Money helps. It doesn't guarantee. AIPAC's United Democracy Project went four for five in contested 2024 Democratic primaries -- dominant, but not perfect. Bilateral capture doesn't need a 100% success rate. It needs a high enough success rate to change the calculation for any candidate considering a pro-regulation position. In low-turnout primaries -- where a few million dollars can constitute a majority of all spending -- the marginal impact per dollar is enormous. And the fact that Anthropic put $20 million into Public First Action, a pro-regulation counter-PAC, proves this is a choice by specific actors, not an inevitable feature of industry participation. But the spending ratio is at minimum 6:1 anti-regulation. A 6:1 ratio is not a debate. It's a purchase.

The Real Election

What makes bilateral capture different from ordinary political spending is not the amount -- it's the target.

Traditional lobbying tries to influence legislators after they're seated. Bilateral capture selects them before they're nominated. The decisive moment moves upstream to the primary, where turnout is low, voter awareness is minimal, and the general election becomes a formality -- a choice between two pre-approved options. This is the difference between playing the game and rigging the bracket.

And it's not just AI. AIPAC perfected this model -- over $100 million in bilateral primary spending in 2024, ads about everything except Israel. Crypto proved the model could be replicated. AI is scaling it. Whatever industry needs to avoid regulation next will adopt it.

Citizens United is the infrastructure. Billionaire election spending has increased 163x since 2010. Over 80% of 2024 billionaire spending used channels that were prohibited before that decision. Bilateral capture is the most sophisticated expression of a fifteen-year trend -- and it needs to be named now, while the mechanism is still visible. Once the primaries are over and the captured candidates are seated, the architecture disappears into the ordinary noise of congressional voting.

The 2026 primaries are underway. The money is already spent. The ads are already running -- about everything except what the money wants.

Here's the question I can't shake: if bilateral capture works, how would you know?

The election still happens. The voters still vote. The winner still gives a victory speech about fighting for working families. Everything looks like democracy. Everything feels like a choice.

The question is whether anything that happened before the ballot was printed still counts as one.

Revision Log

Fact-Check Corrections

• Summer Lee claim (RED FLAG -- removed): The draft claimed "AIPAC spent $14.5 million to defeat Summer Lee in 2024 and failed." This was wrong on every particular -- the $14.5M figure was Bowman (who lost), and AIPAC largely sat out Lee's 2024 race. Replaced with AIPAC UDP's verified 2024 record: "four for five in contested Democratic primaries" per ABC News/538, which honestly supports the same point (money helps but doesn't guarantee).
• Hochul/RAISE Act (RED FLAG -- corrected): The draft said Hochul "crossed out the entire text" and "replaced it with industry-preferred language" as if that was the final outcome. In reality, this was her opening negotiating position; she ultimately signed a compromise version. Rewrote to reflect the full sequence: proposed replacement, public standoff, signed weakened compromise. The industry-influence point still stands; the fait-accompli framing was removed.
• $62 million figure (YELLOW): Clarified that $62M is total spending in four primaries, with $26.9M from industry PACs specifically. Added WBEZ sourcing.
• "Probabilistic, not deterministic" (YELLOW -- voice): Replaced social-science jargon with plain language: "Money helps. It doesn't guarantee."
• Vlasto role (YELLOW): Changed "ran Fairshake's Democratic strategy" to "helped advise Fairshake's Democratic-side strategy" and "top official" to "co-strategist" per sourcing.
• Sherrod Brown descriptor (YELLOW): Changed "top opponent of cryptocurrency" (industry's own framing) to "the banking committee chair who had blocked crypto-friendly legislation."
• Polling range (YELLOW): Added "depending on how the question is asked" to acknowledge the 69-80% range aggregates different poll questions.
• Fairshake $290M (YELLOW): Added "across its network of three aligned super PACs" for precision.
• Spending ratio (YELLOW): Added "at minimum" to the 6:1 figure.
• Public First PAC name (BLUE): Corrected to "Public First Action" throughout.
• AIPAC win rate (YELLOW): Removed the "70% win rate" figure (sourcing unclear on cycle) and let the structural parallel carry the point without a specific percentage.

Structural Changes

• Removed the Vlasto mention from the Architecture section per editorial note. He now appears only in the Proof of Concept section where his role as the human bridge between Fairshake and Leading the Future lands with full weight alongside the crypto precedent.
• Cut "The mechanism has three moving parts" -- the numbered structure (First, Second, Third) provides sufficient scaffolding without the meta-structural announcement.
• Expanded the Bores paragraph with an additional beat of authorial reaction ("the kind of thing you'd cut from a screenplay for being too on-the-nose") to add the controlled anger the outline called for.
• Expanded the PerceptIn aside with a parenthetical reaction ("Ten thousand to three hundred and forty-four thousand. Read that again.") to add voice texture.
• Smoothed the Proof-of-Concept to Real-Election transition. Replaced the flat "Zoom out" stage direction with a direct statement that opens the section through the argument rather than narrating the essay's architecture.
• Trimmed "The Real Election" by ~30 words per editorial note that it was slightly redundant.
• Renamed section headers: "The Architecture of Both Doors" became "Hedging the Bracket" (carries the cynicism of the mechanism); "The Proof of Concept" became "Someone Already Ran This Experiment" (more conversational, more the author's voice).

Voice Adjustments

• Replaced "Now -- a necessary concession here" with "Now -- the industry's stated policy preference isn't crazy." The concession arrives through tonal shift, not announcement, matching corpus patterns (cf. "King of the Hill": "If you spare some charity about motive...").
• Replaced "Here's where the political science matters, though" with "But here's the thing about money and elections." Removes academic discipline-naming.
• Replaced "The relationship between spending and outcomes is probabilistic, not deterministic" with "Money helps. It doesn't guarantee." Plain language over jargon.
• Added 3 moments of authorial personality: (1) the PerceptIn parenthetical reaction, (2) the Bores screenplay aside, (3) the Hochul "proof enough that the money bought access to the negotiating table" editorial beat. These add the self-interruption and visible-reaction texture the editor flagged as the draft's main voice deficit.
• Broke tonal flatness by running the Bores paragraph hotter (the screenplay line) and then cooling into the political-science caveat that follows, creating the temperature shift the outline specified.
• Replaced "Call it bilateral capture:" with "It has a name: bilateral capture." Less ceremonial arrival, per editorial note about how corpus frameworks tend to be named.

Unresolved Notes

• Word count: The final article runs approximately 1,650 words in body text, slightly over the ~1,500 target. The Hochul correction required additional words to accurately represent the negotiation sequence (the draft's one-sentence compression was factually wrong, and accuracy wins over brevity). The author may want to trim elsewhere if the count matters.
• PerceptIn figure: The fact-check flagged ambiguity about whether $344,000 was per-deployment or total compliance spend. The Fortune article confirms both figures and that the company "went out of business." I used "spent over $344,000" which is accurate either way, but the author should verify against the Harvard Kennedy School paper if precision matters.
• 1,200+ state AI bills breadth: The fact-check noted this includes autonomous vehicles, facial recognition, deepfakes, etc. -- not just comprehensive AI regulation. I kept the figure because it accurately represents the compliance landscape the industry faces, but the author should decide whether this context matters for the counterargument's force.
• Section header "Hedging the Bracket": The editor wanted more personality in headers. "Hedging the Bracket" carries the financial-engineering cynicism and the sports metaphor that pays off later ("rigging the bracket"). But it's a judgment call -- the author may prefer something else.