The First Election with No Referee

Twenty-nine percent. Thirty-two percent. The first is the share of AI-made or AI-altered midterm ads that carried a disclosure label in states that require one; the second is the share in states with no such law. ^[1] The requirement changed almost nothing. South Carolina Lieutenant Governor Pamela Evette has been both the target of deepfakes and a user of AI in her own ads. She describes what that does to a voter.

So you almost think it’s a joke and you kind of laugh at it, until you realize that there is a group of people out there that can’t tell the difference. — Pamela Evette

Federal deepfake guardrails are stalled: Senator John Kennedy's bill requiring a kill switch for rogue AI was blocked by Rand Paul as premature. ^[2] The one election bill with a path to the president's desk points the other way. Trump says he will sign no legislation until Congress passes his expanded SAVE America Act — documentary proof of citizenship, photo ID, limits on mail-in voting — and has already cancelled a bipartisan housing bill to make the point. ^[3]

It supersedes everything else. — Donald Trump

There will be Voter I.D. for the Midterm Elections, whether approved by Congress or not! — Donald Trump

So the one law that moves restricts access, not the tools deceiving voters. ^[4] California tried to referee at the state level, and a federal judge struck down both its deepfake laws — one imposing liability on platforms, one banning deceptive deepfakes within 120 days of an election — in suits brought by X and The Babylon Bee. The risk, the judge ruled, "cannot be quashed through censorship or legislative fiat." ^[5] The money is invisible to the official eye. Federal Election Commission itemized data counts $17 million spent on AI vendors since 2020; researchers tracking the same phenomenon counted 164 AI-backed ads supported by roughly $80 million this cycle alone. ^[6] The difference flows through consultants and media firms that never have to itemize. Capacity has not vanished; it has moved to actors who cannot compel anyone. Some experts argue election officials are better prepared than in previous cycles, and OpenAI has modified its image tools to block deceptive voter-fraud content. ^[7] The Brennan Center is training local officials for scenarios no previous administration produced — federal search warrants to seize ballots after the election, armed federal agents at polling stations, drones dropping pamphlets — and demand for that training has nearly doubled since 2024. ^[8] Gary Restaino, a former DOJ attorney who runs the drills, puts it plainly.

We've not seen threats of having federal armed agents at the polls before. — Gary Restaino

A drill cannot compel anything, and neither can a newspaper. With no cited federal counterpart, the New York Times built its own voter-data tool, recontacting 30,000 poll participants and synthesizing precinct results with voter files. ^[9] It is a private institution manufacturing a public function. India's Election Commission shows what a referee in force looks like. It invoked constitutional powers to require labels covering at least 10 percent of any AI-altered campaign content, to name the responsible entity, and to pull misleading material within three hours — rules with "immediate effect" for its polls. ^[10] Nothing comparable is in force for the midterms. The channel all this panic polices, fake content in feeds, barely moves anyone. A 16,000-user study using Meta's internal data found that removing fact-checked misinformation from feeds changed voter beliefs by "approximately zero." ^[11] The channel nobody regulates is the one that works. A government-backed study across the U.S., UK, Canada, and Poland found AI chatbots swayed voters roughly four times more than television ads — moving one in twenty-one U.S. opposition voters, with effects still measurable a month later. ^[12] The most persuasive bots supplied the most information regardless of its truthfulness, and about 19 percent of their claims were inaccurate. ^[12] The NRSC ran an 85-second AI deepfake of Texas Senate candidate James Talarico — legal precisely because Texas bans deceptive deepfakes only within thirty days of an election. ^[13] Separately, Google's AI search summary highlighted a donation page that routed money meant for Talarico to sitting Senator John Hickenlooper, in a different race. ^[7] The only referee touching the channel that measurably moves voters is the industry itself: Google, OpenAI, and Anthropic now feed their chatbots verified voting information through the nonpartisan Democracy Works. ^[7] The fix feeds itself — the companies that built the persuading machines hand them the truth, because no one else with authority over this election has stepped in. And still, a donor's money went to a sitting senator in a different race instead.