AI Chatbots Are Already Hallucinating Election Facts

People are actually asking AI how to vote

We noticed search traffic for voting information is shifting straight into chat interfaces instead of official sites. ChatGPT and Claude are being used to look up polling places, registration deadlines, and ballot drop-off locations. The problem is these models don't know real-time facts. They predict the next token based on training data that's months old, so they'll confidently state a deadline that passed weeks ago or invent a precinct that doesn't exist. That's not a minor bug. It's a fundamental mismatch between how these tools work and what voting requires.

Hallucinations don't get a do-over on Election Day

If an AI tells you your polling place is at a library that closed in 2022, you don't get to refresh and try again. You miss the window. Early voting deadlines are hard cutoffs, and registration rules vary by county in ways that general-purpose models simply don't track. We're looking at a scenario where someone trusts a chatbot, shows up to the wrong location, and leaves without voting because they ran out of time. That's voter suppression delivered through a friendly interface, and there's no customer support ticket you can file after the polls close.

The bias problem is quieter but more dangerous

Beyond factual errors, these models carry partisan leanings that show up in how they frame issues or describe candidates. One model might summarize a ballot measure in neutral terms while another emphasizes the opposition's talking points. A voter asking for help understanding a proposition could get a steered perspective without realizing it. Unlike a newspaper editorial, there's no byline and no disclosure. The AI presents itself as an objective tool, which makes the bias invisible to anyone who isn't already checking multiple sources.

This isn't the first time tech outran election infrastructure

We saw the same pattern in 2020 when social platforms struggled to label misinformation fast enough, and before that with robocall campaigns that targeted specific neighborhoods with false voting dates. The difference now is the interface feels personal and authoritative. A tweet looks like a tweet. A chatbot answer looks like advice from someone who knows what they're talking about. Election officials are still publishing PDFs and running phone banks while voters are getting their civic information from a model that was trained on Reddit and Wikipedia.

What this signals for the next ninety days

We're heading into a stretch where AI-generated voting guidance will only increase, and the models aren't getting more factual anytime soon. Hallucination rates haven't dropped meaningfully across the major releases, and real-time grounding still fails on niche local data. The second-order effect is that close races could be decided by thousands of people who received bad information from a tool they trusted. Nobody's building a verification layer for this, and the people most likely to rely on AI for voting help are often the ones with the least time to double-check.

What you should tell clients right now

If anyone on your team is thinking about embedding an AI assistant for civic information, tell them to stop. The liability isn't theoretical. A wrong polling location or a misstated deadline creates real-world harm that no terms of service will cover. The safe play is pointing users to state election offices and .gov domains, not letting a language model guess. We're treating this as a hard boundary: AI for drafting emails, not for deciding where people cast their ballots.