Experts warn of an AI disinformation wave before the midterms

Seven days before the US midterms, researchers say social media is carrying more AI-generated election disinformation than any previous cycle, and that platforms have cut back the teams that used to catch it. The Guardian gathered the warnings on 27 October; the vote is on 3 November.

“AI has significantly reduced the cost of information operations and made it possible to quickly create numerous versions of a single message.”

— Frances Haugen, former Facebook engineer, to the Guardian, https://www.theguardian.com/us-news/2026/oct/27/ai-disinformation-social-media-midterm-elections

Frances Haugen, the former Facebook engineer who turned over internal research in 2021, frames it as an economics problem rather than a content problem. Generating one persuasive message used to take a person. Generating two hundred variants of it takes a prompt. Each time a version gets removed, the next one goes out slightly reworded, and some of those slip past a classifier trained on the version before. The thing doing the rewriting is an LLM costing fractions of a cent per variant, which is why the supply side of this has no natural ceiling.

The enforcement side moved the other way. X replaced its internal safety review with Community Notes after Elon Musk bought it, which puts correction after publication and relies on volunteers rather than staff. Rules against false claims about voting procedure still exist on most platforms; what thinned out is the number of people checking whether anyone follows them.

$3.4m a day on one side, nothing filed on the other

Here is the asymmetry worth holding onto, and it is measurable on one side only. We pulled the FEC filings for Leading the Future, the super PAC backing pro-AI candidates. Through 14 October it had raised $75.98m, spent $52.16m and still held $23.82m in cash. Divide the remainder across the seven days left and it is $3.4m a day of disclosed political money.

Every dollar of that is filed, dated and public. The synthetic posts Haugen describes carry no filing at all, cost a rounding error of that budget to produce, and are deliberately hard to attribute to anyone. The regulated half of political persuasion is the half that was never the problem.

A week out

We wrote on 19 October about Democrats campaigning against AI and the money lining up on the other side of that argument. A week out, the question is no longer whose argument wins. It is whether anyone can tell which messages came from a campaign that has to file and which came from an operation that does not.

Informational material, not investment advice. The disinformation warnings are expert assessments rather than a measured count, the FEC figures cover the 2026 cycle through 14 October 2026, and the per-day arithmetic on them is ours.