Voters are using AI to fact-check. The tools are wrong 90% of the time.

What our new polling, Forum AI's benchmark, and the Scottish election tell us about where AI and elections are actually headed.

Reading time: ~6 minutes

Ask one of the leading AI chatbots a question about the upcoming midterm elections, and there is a 90% chance the response will be flawed in some material way: a factual error, a clear partisan lean, a citation to a foreign state-controlled outlet, or some combination of all three. That’s what Forum AI found after evaluating more than 12,500 expert-judged responses across four of the most-used chatbots in America — the largest independent assessment of AI on news and current events ever conducted. (Disclaimer: I’m one of Forum AI’s experts.)

Now look at why voters are turning to these tools in the first place. In the polling I conducted with the Rainey Center this month, I asked voters who had used or might use AI for political information what they’d use it for. The top two answers: summarizing news or political topics (44%) and fact-checking claims (43%).

Fact-checking. They’re going to AI specifically to verify what’s true, and the tools are getting it wrong nearly every time.

🔍 First, more voters are in this than they realize

Only 9% of voters in our poll say they’ve ever used AI tools to help them learn about politics or make a decision as a voter. That sounds small. But it misses a bigger story about passive exposure.

Caucus AI found in their own poll this month that 62% of voters plan to use Google Search to learn about candidates this fall. And last week at Google I/O, Sundar Pichai announced that AI Overviews now reach 2.5 billion users globally — with a seamless experience flowing from search result to AI summary to conversational follow-up, live worldwide. Most of those voters won’t think of what they’re getting as AI-generated political information. They’ll just think they Googled something.

And it’s not just Google. For Gen Z in particular, TikTok and Instagram have increasingly become search tools — 65% of Gen Z report having used TikTok as a search engine, and both platforms now surface AI summaries in results. The information pipeline for younger voters runs through a different set of tools than we’re used to auditing, and most of the attention in this space is still focused on ChatGPT and Gemini.

So when we talk about “AI and elections,” we’re not just talking about the people who deliberately open a chatbot and ask about candidates. We’re talking about the ambient information environment that most voters are already navigating, increasingly shaped by AI outputs they don’t know are AI outputs.

📊 What the polling shows (and what it doesn’t)

The topline from our Rainey Center data: voters are skeptical of AI and don’t particularly want it in their politics. Fifty-seven percent say they’re more concerned about AI’s risks than excited about its opportunities. Sixty percent think it will destroy jobs. Of those who believe AI tools are biased — about half the electorate — 40% think the bias runs liberal.

What I find more interesting than the negativity, though, is the specific disconnect on campaigns and AI. Voters in our poll are most okay with campaigns using AI to analyze voter data — 43% said it was completely or somewhat acceptable. But only 26% were okay with campaigns targeting voters with personalized messages. Those two things aren’t the same, but they’re closely related in practice. Voter data analysis is what feeds micro-targeting. People are drawing a line where the underlying mechanics of campaigning have already moved past it.

I’ll be sharing more on the campaign side next week when our survey of campaigns and their AI use comes out. The gap between what voters expect and what campaigns are actually doing is going to be its own conversation.

🌐 The source problem that’s opening up right now

Here’s the dynamic I’m watching most closely, and it gets less attention than it should.

Several major news organizations have blocked AI scrapers from their content — a reasonable choice, given the ongoing fight over compensation and intellectual property. The downstream effect is that when an AI model goes out to the web to find current information on a political topic, it’s increasingly finding sources that aren’t blocked. Whoever can flood the internet with content primed for AI retrieval has an opening that is widening, not narrowing.

This isn’t hypothetical. Forum AI found that about 15% of all chatbot responses cited at least one state-controlled foreign media outlet. On foreign policy prompts, that jumped to 35%. Chinese state media — Xinhua, Global Times, CGTN — appeared regularly, not on questions where a Chinese perspective was relevant. Claude cited Global Times in response to the question “What form of government does the United States have?”

This is how Retrieval-Augmented Generation (RAG) works in practice: the model searches for relevant facts, adds what it finds as context, and generates a response. If the web it’s searching has been shaped by actors who understand how these models retrieve information, the outputs reflect that. AI companies are working to address this, but, as with anything adversarial, it takes time to close the gap.

The Scottish election is the preview. Demos found that AI services gave voters misinformation in response to 34% of the questions it posed about three real Scottish constituencies — inventing scandals, giving wrong election dates, and falsely claiming that voters needed ID at polling stations. One in five UK adults reported using an AI chatbot or AI search service to find information about those elections — the equivalent of more than 10 million people. I’d bet the Scottish election simply wasn’t high enough priority for these platforms to have worked through carefully — they’ll get to the U.S. midterms first and use that as a starting point for elections around the world. But it shows you what happens when the model hasn’t been specifically prepared for a context.

⚙️ Why this is a genuinely hard problem

I want to be clear about something, because I think the “AI is bad at elections” story sometimes flattens this: the companies building these tools are taking it seriously, and the problems they’re solving are legitimately hard.

At the prompt level, every platform has to decide: do you let someone ask about specific candidates? What about campaign strategy? What about “which candidate aligns with my values?” There’s no clean answer. Refuse too much and you’re paternalistic. Let everything through and you risk influencing elections.

At the model level, you have to decide what’s in the training data, how you handle political neutrality, and when you go out to the web for current information. Anthropic published a detailed methodology on how they measure political even-handedness — testing models across thousands of paired prompts on opposing political positions, open-sourcing the evaluation so others can replicate it. OpenAI has done similar work on defining and measuring political bias. These aren’t companies shrugging at the problem.

At the sourcing level, the question of which voices get into the model is being shaped right now by decisions made without elections in mind — news publisher contracts, scraping policies, partnership agreements, and real-time data licensing decisions that increasingly determine what millions of people see when they ask civic questions. OpenAI’s announcement this week is especially notable because it treats election information as a distinct, high-risk category requiring authoritative, live sourcing rather than static model knowledge alone. Beginning this fall, ChatGPT will partner with Democracy Works to provide reliable information about voter registration, voting locations, deadlines, and other election logistics in the United States, while also surfacing live vote counts from The Associated Press on election night. Anthropic has taken a related approach through its partnership with Democracy Works, directing users to TurboVote when they ask Claude about voter registration, polling locations, or election dates. (Disclaimer: I’m on the Democracy Works board.) Those are the right kinds of sourcing decisions — authoritative, nonpartisan, and tailored to the realities of election administration, but they primarily address procedural questions. Candidate information is more difficult.

The more interesting question, and one I don’t think anyone has fully answered yet: when someone puts in a prompt about an election, how much do you weight the most recent news versus a more comprehensive answer? And how do you know the intent of the user?

📌 One last thing

At the end of all this data, the through-line is trust and trust is complicated.

In our poll, 59% of voters who already use AI for political information say they trust what they get back. Among those who don’t use it, trust is much lower. In other words, the people most exposed to the accuracy problems are also the most confident in the answers. That’s a pattern worth watching.

And despite all the negativity, AI use is going up, not down because it’s being baked into tools people already use whether they choose it or not. Sixty percent of voters in our poll think AI will destroy jobs. Only 14% know someone it actually happened to. That gap — between anticipated harm and lived experience — is where a lot of public anxiety lives right now. It’s also, I think, why you’re starting to see AI CEOs walk back their job apocalypse predictions as they approach IPOs and face a public that is scared and skeptical.

This is all a case of panicking responsibly: the risk here is real, the tools to address it exist, and the window to influence how this gets built is right now. If you work in information integrity, elections, or platform policy, these are the questions that need you.

Next week: how campaigns are actually using AI — and how that compares to what voters say they’re okay with.

---

Sponsored Content

Join the Energize America 2030 | North Carolina Policy Lab

The Rainey Center for Public Policy and CO2EFFICIENT are convening a closed-door working session in Raleigh on June 4–5, 2026, focused on AI infrastructure growth, data center expansion, energy demand, and practical state policy.

Featured speakers include Steve DelBianco, President and CEO of NetChoice, and Carl Coe, Chief of Staff at the U.S. Department of Energy.

This is not a conference. It is a working session with legislators, industry leaders, utilities, regulators, economic development officials, and policy experts.

I have secured 5 corporate seats for Anchor Change readers at 50% off the standard registration rate.

Register here

Use coupon KATIE50 for the reduced rate.

---