It is easy to imagine the moment: a voter scrolling through a feed late at night, pausing on a post that looks almost plausible. The face is familiar. The accusation is sharp. The story is false—but it spreads faster than any correction can catch it. In the final stretch before November’s US midterm elections, experts warn that social media is being flooded with an unprecedented wave of AI-driven disinformation, and much of it is probably slipping through the cracks. This is not the same as the chaotic misinformation battles of past cycles. It is faster, cheaper, and harder to trace. Frances Haugen, the former Facebook engineer turned whistleblower, put it plainly: operatives can now run information operations with a level of quality that was impossible in 2016, and they can do it for “super, super cheap.” The old playbook involved troll farms and clumsy coordination. Now a single person can generate dozens of variations of a lie, each one slightly different, each one designed to evade automated detection. “You can just keep tweaking the post a little bit each time you create them,” Haugen said, “and some of them are going to get through.” Meanwhile, the platforms themselves have pulled back. After the Cambridge Analytica scandal, Facebook and other social media giants rushed to build trust and safety teams. But a counter-backlash from the right, fueled by fears of censorship and intensified by Elon Musk’s takeover of X and Donald Trump’s return to political dominance, has led to sweeping layoffs across those teams. AI is now being used to fill the gap, but it is not the same as human judgment, and the consequences are showing up in real time, in real races, with real voters.
Take the case of James Talarico, the young Democrat running for Senate in Texas. Earlier this month, he took time off from campaigning to recover from the flu. But online, a very different story began to spread: that he was being treated at a mental health facility in Ohio. The claim was entirely fabricated, pushed by a rightwing provocateur who published a deceptively edited phone call with a hospital staffer. Talarico’s campaign released security-camera footage showing him at home in Texas. Talarico himself clarified publicly that he had never even been to Ohio. But the lie had already taken root. It was shared, stoked, and recycled for days. What happened next was even more revealing. The Digital Democracy Institute of the Americas, or DDIA, tracked how the misinformation mutated as it crossed into Spanish-language media. What began as a false story about hospitalization in Ohio became a different false story in Spanish: that Talarico had attempted suicide. The original correction, even if it ran everywhere, could never keep up with the changed version. Roberta Braga, the founder and executive director of DDIA, explained that platforms enforce their rules “more congruently” in English than in other languages. There are simply fewer guardrails for Spanish-language content, especially now that companies have rolled back their investments in human moderation. “Now that they’re using AI to do a lot of the content moderation,” she said, “even more is likely slipping through the cracks.” The Talarico case is not an isolated glitch. It is a window into how modern disinformation works: it is shape-shifting, multilingual, and nearly impossible to pin down with a single fact-check.
The same pattern is playing out in Pennsylvania, where online hit squads that once operated on the fringes now find their claims amplified by official government sources. When viral posts on X and rightwing blogs claimed that an undocumented immigrant arrested in central Pennsylvania was an active field worker and canvasser for Democrat Janelle Stelson’s congressional campaign, her staffers found themselves sprinting to catch up with a lie. They released records showing that the immigrant, Chimaobim Adedayo Igbojekwe, had no paid position on the campaign, despite an initial police incident report stating that he claimed to be a campaign worker. The claim was provably false. But as the campaign scrambled, an official statement from the Department of Homeland Security amplified the allegation. Charles Wall, the deputy director of Immigration and Customs Enforcement, declared it “unbelievable” that Stelson would employ an undocumented immigrant, let alone someone accused of indecent exposure. The Stelson campaign sent cease-and-desist letters to media outlets. A Fox News reporter took down a post about the arrest. Activists at Citizens for Responsibility and Ethics in Washington filed a complaint accusing the DHS of violating the Hatch Act, which limits the political activities of government employees. But the damage had already been done. Stelson’s opponent, Republican congressman Scott Perry, continued to run ads based on the false claim. Nora Benavidez, a civil rights and free speech attorney with the nonprofit Free Press, saw a familiar pattern. Trump has made clear he intends to interfere in elections, she said, and he has leveraged agencies across the federal government—from the Postal Service to DHS—to do it. The Stelson campaign has not directly complained to Meta, X, or YouTube, though a spokesperson said it was considering every legal avenue to fight the false claims in Perry’s ads. The situation illustrates a troubling new reality: even when campaigns catch the lie, they no longer know exactly where to turn.
The platforms themselves have rules, but those rules are narrow, and the enforcement is often shallow. Meta, which owns Facebook and Instagram, requires advertisers to disclose the use of synthetic images and AI-generated content in election ads, and it uses automated labeling for AI-generated images identified through the Coalition for Content Provenance and Authenticity metadata. It also bans posts and ads that misrepresent voting dates, locations, or legal requirements, and it prohibits threats of violence. But Meta does little active moderation of political commentary beyond excluding unsolicited political content from recommended posts. An analysis by Free Press noted that in 2023, Meta narrowed its historical restrictions to allow debate around the legitimacy of past election cycles. In other words, a falsehood about a past election is now protected speech on the platform. YouTube, meanwhile, says all users are subject to the same community guidelines and have access to the same reporting tools. But its definition of election misinformation is specific and limited: content that misleads voters about the time, place, means, or eligibility requirements for voting is banned, as are false claims that could materially discourage voting. Claims that a candidate is somehow ineligible, or calls to hack a polling site’s computers, are also forbidden. But a direct lie that falls outside those lines can propagate without intervention. In 2023, YouTube also stopped removing content that falsely claimed the 2020 election or other past US presidential elections were marred by widespread fraud, errors, or glitches. The company said the removals had curbed misinformation but also had a chilling effect on political speech. Google, which owns YouTube, has a likeness detection system that allows civic leaders to request removal of videos when AI-generated images falsely depict their words or deeds. But rather than relying on takedowns, Google prefers to amplify authoritative news outlets and election boards in search results and underneath high-stakes civic videos, using contextual information panels. This is a softer approach, and it reflects a broader industry shift away from direct enforcement and toward containment.
Nowhere is that shift more visible than on X, the platform formerly known as Twitter. After Elon Musk acquired the company, he abandoned internal safety reviews and replaced them with the crowdsourced Community Notes program. The platform still has a civic integrity policy that bars false information about how to vote and suppression schemes. It also supposedly restricts synthetic media designed to mislead. But experts argue that without real enforcement, these policies are little more than window dressing. Community Notes relies on users to add context to misleading posts, and while it can be useful, it is not the same as proactive moderation. It also tends to be slower than the viral spread of a lie, and it can be gamed. Haugen, the whistleblower, says the problem is not just the content itself but the systems that distribute it. “When you stop being able to focus on content because the AI will just slightly tweak it until it gets through,” she said, “you have to focus on what are the systems of distribution. What are the biases? What are the ways we have feedback systems, cycles, that allow the community to provide feedback, that allow it to be damped down?” Her point is crucial: in a world where AI can generate endless variations of a lie, relying on content-by-content review is hopeless. The platforms need to think about design, reach, and algorithmic amplification. But so far, there is little sign they are willing to do that. Instead, they are cutting staff, narrowing policy, and hoping the problem stays just out of view.
The real cost of all this is borne not by the platforms but by ordinary people: by voters trying to make sense of a confusing information environment, and by candidates trying to defend their reputations against attacks that never stop evolving. Talarico had to produce security footage to prove he was not in Ohio. Stelson had to send legal letters to media outlets to rebut a claim that should never have been made. Both of them are campaigning in a world where a single false story can dominate the news cycle, mutate across languages, and be amplified by government officials. The experts who study this problem say the trend is not accidental. It is the result of deliberate choices made by social media companies to prioritize speech, or profit, or both, over safety. It is the result of a political climate that punishes moderation as censorship. It is the result of artificial intelligence being deployed without adequate human oversight. And it is the result of a public that is exhausted, skeptical, and increasingly unsure what to believe. There is no easy fix. But there are warning signs, and there are people paying attention. The question is whether the platforms, the regulators, and the voting public will look up from their feeds in time to do something about it. For now, the flood continues, and the cracks are wider than ever.