The Deepfake Midterm
26 states have AI election laws. Congress has none. And the 2026 midterms start in months.
In 2023, five states had laws addressing AI-generated content in elections. By February 2026, that number is 26. Five more — New Jersey, Virginia, Maryland, Tennessee, and Vermont — are actively considering legislation. Montana and South Dakota passed new disclosure requirements this year specifically aimed at next year’s midterms.
Congress, meanwhile, has passed nothing.
The 2026 midterms will be the first major American election conducted under a patchwork of state-level AI rules and zero federal standards. The result is a regulatory landscape where a deepfake political ad that’s illegal in California might be perfectly legal in the state next door — and where the platforms hosting those ads are making up their own rules as they go.
What the states built
The predominant approach is disclosure, not prohibition. California’s AB 2355 requires political committees to include a disclaimer: “Ad generated or substantially altered using artificial intelligence.” Most other states followed the same model — label it, don’t ban it.
There’s a reason for the caution. California tried something stronger in 2024 — a law that would have blocked platforms from hosting deceptive political deepfakes. A federal judge struck it down on First Amendment grounds. That ruling put a chill on every state considering an outright ban. The constitutional message was clear: you can require labels, but you probably can’t prohibit the content.
So 26 states now require some form of disclosure. But disclosure requirements only work if someone enforces them. The Federal Election Commission has no AI-specific disclosure rules. State enforcement agencies are underfunded and understaffed. And the deepfakes themselves are getting cheaper, faster, and harder to detect by the month.
What the platforms are doing
Meta announced its 2026 midterm preparations in February. The headline commitments: advertisers must disclose when using AI to create or modify election-related ads. Disclosed information appears on the ad and in Meta’s Ad Library, which holds over 18 million U.S. entries. Meta will block new political ads during the final week of campaigns. AI-generated photorealistic videos or audio require disclosure labels, with penalties for non-compliance.
Meta is using C2PA — the Coalition for Content Provenance and Authenticity — as a technical standard for detecting AI-generated content. An “AI info” label appears on detected content, based on either industry signals or self-disclosure.
The company has removed 200 coordinated inauthentic behavior networks since 2017 and reports spending over $30 billion on safety and security over the past decade. Its Election Operations Center monitors threats in real-time with teams across threat intelligence, data science, engineering, and legal.
This is the most comprehensive platform response. Most others are doing significantly less.
The AI PAC spending surge
While states regulate deepfakes in political ads, a different AI-election story is unfolding: AI-aligned political action committees have launched coordinated advertising campaigns targeting federal and state lawmakers who sit on technology committees. The spending accelerated sharply in early 2026, as Congress weighs several competing proposals to regulate artificial intelligence.
The dynamic creates a feedback loop. AI companies spend to elect lawmakers friendly to their industry. Those lawmakers decline to pass federal AI regulation. The absence of federal regulation leaves states scrambling to fill the vacuum. And the AI companies argue that the resulting state-by-state patchwork proves the need for federal preemption — preemption that would, conveniently, replace 26 different sets of rules with whatever standard the industry-friendly Congress produces.
What happened in 2024
The actual impact of AI deepfakes on the 2024 elections was, by most assessments, less than feared. R Street Institute’s analysis found that while AI-generated content was widespread, its measurable effect on voter behavior was limited.
But 2024 also revealed the infrastructure for AI-enabled election interference. There was a significant uptick in AI-enhanced DDoS attacks on election websites. Targeted phishing attempts penetrated presidential campaigns using AI-crafted messaging. Robocalls using AI-generated voice cloning targeted voters in New Hampshire.
The 2024 lesson wasn’t that deepfakes failed. It was that the tools worked and nobody used them at full scale yet. The capability is ahead of the deployment — a gap that narrows with every election cycle.
The CISA gap
The Cybersecurity and Infrastructure Security Agency rolled back some of its election cybersecurity duties under the current administration, shifting hardening responsibilities to state and local levels. This means the federal infrastructure for detecting and responding to AI-generated election threats is being reduced at exactly the moment those threats are becoming more sophisticated.
State and local election officials now bear greater responsibility for AI threat detection with fewer federal resources. Many of these offices have single-digit staff counts and zero AI expertise. The TAKE IT DOWN Act, signed in May 2025, provides the first federal framework for addressing intimate deepfakes and requires platforms to remove flagged content within 48 hours. But it doesn’t cover political deepfakes at all.
The 2026 test
The 2026 midterms will test every assumption about AI and democracy simultaneously. Can disclosure labels work when the platforms enforcing them are also the platforms profiting from political ad spending? Can 26 different state laws function when AI-generated content crosses state lines at the speed of a social media share? Can underfunded state election offices detect and respond to AI threats that nation-state actors and domestic operatives are actively deploying?
The AI industry’s position is that these questions are best answered at the federal level. The federal government’s position, so far, is silence. And the elections are happening whether the framework is ready or not.
The deepfake midterm isn’t about whether AI will interfere with elections. It’s about whether anyone built the systems to notice when it does.
Originally published at https://noahaust2.github.io/strategist-dashboard/blog/the-deepfake-midterm.html
Write a comment