The rush to integrate artificial intelligence into major social media platforms has placed a slew of potential pitfalls between voters and reliable information about elections, dangers that tech companies are addressing inconsistently, if at all.
Ahead of the midterms next month, chatbots, media generation services and automated content moderation systems are increasingly shaping voter choices, according to political experts and new research, even as A.I. enables new forms of toxic content.
A.I. tools, some introduced just months before the election, have contributed to biased and inaccurate information about the polls and amplified foreign influence operations. They have also elevated low-quality slop content, which is now tainting the training data that informs chatbot responses, according to researchers.
And despite pressuring tech companies for details about testing, safeguards and the extent of voter interactions with A.I., lawmakers and watchdog groups have said that the companies are not sharing enough information about those efforts.
“We’re in uncharted territory at this point,” said Deen Freelon, a professor of communication at the University of Pennsylvania who studies digital politics.
More than in any other election, tech companies are injecting A.I. into all levels of the user experience. During the last national election in 2024, platforms were offering basic A.I. image generation tools. Those have bloomed in 2026 into advanced services combining audio, video and image capabilities.
Curious about a candidate’s policies? Google delivers an A.I. overview before showing the politician’s website in search results. Scrolling social media? Prepare to see attack ads with fabricated images of candidates.
One in five voters plan to use A.I. to look into a race on their ballot, according to a survey released this month by the American Association of Political Consultants Foundation. Twice that number have already received election details through an A.I. search summary.
And yet just 1 percent of the polled voters considered A.I. to be their most credible source for election information, the survey found.
“This is the first election where A.I. is deeply embedded into the platforms,” said Max Read, the U.S. director of civic innovation at the Institute for Strategic Dialogue, a research organization.
The institute has been reviewing efforts by major social media platforms to protect political discourse, election integrity and public safety. On Thursday, researchers released a report with their conclusion: Transparency and human oversight are lagging A.I.’s breakneck development pace. Social media platforms are relying on a hodgepodge of A.I. policies, resulting in “meaningful gaps” in how artificially generated content is disclosed or restricted ahead of the midterms, the report found.
Platforms are also increasingly turning to automated and crowdsourced options over humans to moderate content, although the extent remained a mystery to the institute’s researchers. Relying on A.I. for such work is a gamble for companies trying to account for context and nuance during fast-moving, intensive events like elections, the researchers said.
“Their response has been uneven, poorly documented and, in several cases, poorly suited to addressing the harms it is meant to prevent,” Mr. Read said.
Tech companies are facing increasing political pressure to clarify how A.I. plays into their policies for handling election-related disinformation. Senators Adam Schiff and Alex Padilla, both Democrats from California, wrote last month to the heads of companies including Google, Meta, Snap, TikTok and xAI asking for details about their plans to defend election integrity and combat false narratives during the midterms.
Google offered the most extensive answers about A.I., including that it would use the technology in its search features to connect users to vetted information about polling locations, registration deadlines and real-time voting results. But the senators said in statements that the companies’ replies as a whole were inconsistent and incomplete, especially so close to the midterms.
“Social media companies hold enormous power over the content voters see: They decide what to amplify and what not to,” Mr. Schiff said.
In a blog post earlier this month, TikTok laid out its midterms strategy, which included labeling A.I.-generated content and allowing U.S. users to add context to posts through its Footnotes feature. The company said in a separate statement that it was “dedicated to maintaining a safe and secure platform.”
Google said in a statement that it rigorously tests its A.I. overviews for accuracy and also helps users identify A.I. content through prominent labels on YouTube, watermarks and disclosures on election ads. Meta, which owns Facebook, Instagram and WhatsApp, responded that it has robust election safety plans, including real-time threat modeling. The company said it thoroughly tests its A.I. models to ward off wrong answers about the elections.
Most major tech companies are making an effort to preserve election integrity, according to Democracy Works, a nonpartisan organization that is partnering this year with most of the major chatbot developers to share verified election information. XAI is the exception; it also did not reply substantively to Senators Padilla and Schiff.
The acceleration of A.I. technology, and the occasional tendency of chatbots to veer away from neutral facts, makes their work difficult, said Luis Lozada, the chief executive of Democracy Works.
“I don’t think they’re just completely negligent,” he said. “I just think that this is a very, very difficult issue to navigate.”
This summer, the Brennan Center for Justice, a voting rights and democracy group, found that the most popular chatbots regularly pushed back against false election narratives. The A.I., however, could also be prompted to generate content that could undermine elections, such as images of people stuffing multiple ballots.
OpenAI has modified its tools in response to those findings, according to both the company and Brennan. Instead of generating seemingly realistic images of voter fraud, it now renders them as cartoons in an attempt to block intentionally deceptive content without curtailing freedom of speech.
“We are degrading the evidentiary value of an image while still allowing social commentary,” said Josh Lawson, the head of civic and strategic risk at OpenAI. It is one of numerous efforts the company has taken, he said, to screen out malicious attempts to undermine the elections.
In October, Brennan researchers conducted a new round of tests. Only OpenAI had tightened its controls over election falsehoods, they found in a new analysis published on Thursday. Other chatbots from Google, Meta and others had regressed.
(The New York Times sued OpenAI and Microsoft in 2023, accusing them of copyright infringement of news content related to A.I. systems. The two companies have denied those claims.)
Chatbots have become yet another source of information that can be exploited for influence campaigns. The most popular models now sometimes cite known sources of politically biased content or foreign disinformation, including news sites fabricated by or mimicked by Kremlin-linked propaganda operations, according to disinformation researchers.
VisionHeight, a cybersecurity company, reported this month that it found a network of 37 A.I.-filled sites purporting to provide state or regional news that were created in a single day by an anonymous organization in September. The sites also included code that effectively welcomed the A.I. tools that “crawl” for the content that powers chatbots.
The sites fed the chatbots information about candidates in 11 House races in seven states, most of them Democrats.
“The sites are not written for voters,” VisionHeight’s report said. “They are written for the A.I. systems voters ask.”
NewsGuard, a company monitoring harmful online content, has identified more than 1,100 such organizations. It defines them as “pink slime,” sites that “present themselves as professional news outlets but are secretly backed by partisan groups.”
In its own tests, NewsGuard found that chatbots, when prompted about narratives that appear in these partisan sites, cited them nearly half the time.
In some cases, Russia has managed to successfully inject its state propaganda, including about American politics, into chatbots’ training data, according to NewsGuard and other researchers.
When Alethea, a company that tracks influence operations, asked about disinformation long circulated by the Kremlin, some of the most prominent chatbots mentioned Russian state propaganda sources and fabricated news outlets.
XAI’s Grok, for example, was asked about the impact of a European ban on Russian natural gas. It referenced InfoRos, an arm of the Russian military intelligence agency that the United States cited for trying to interfere in the 2020 American presidential election.
“When you use a chatbot,” said Lisa Kaplan, Alethea’s founder, “you are not expecting to be deceived by a foreign actor who’s trying to manipulate you.”
The post A.I. Is Embedded in Your Election Experience. So Are Its Dangers. appeared first on New York Times.
