AI chatbots provide incorrect voting information for US elections

Errors in voter guides

As the United States approaches key elections, researchers have identified a significant issue with how AI chatbots handle political information. Reports from the Washington Post indicate that top models are providing users with flawed and inaccurate voting guides. These errors occur even when the prompts are specific to certain states or local jurisdictions.

In one specific test involving Missouri, a chatbot included a Senate seat in a voter guide that was not actually up for re-election. Another test for a resident in Wyoming resulted in the AI claiming that local politicians in municipal elections were actually running for federal seats. These examples demonstrate a lack of precision in how the models process geographical and political data.

The scale of these errors is notable because they appear across multiple platforms and queries. The data suggests that the models are not merely making small typos but are constructing entirely incorrect narratives about the electoral process. This creates a risk for voters who rely on these tools for quick, convenient summaries of their local ballots.

The cycle of AI slop

A major factor in these inaccuracies is a phenomenon where AI models consume and reproduce content generated by other AI systems. The Washington Post analyzed 3,000 AI-generated voting guides and found that much of the false information traced back to other AI-generated websites. This creates a loop where the models are essentially 'eating their own slop' and presenting it as authoritative fact.

When an AI model pulls data from a website that was originally populated by another AI, the errors can compound. This makes it increasingly difficult for the software to distinguish between verified human reporting and synthetic misinformation. The result is a degradation of information quality as the content circulates through different layers of machine learning.

This cycle is particularly concerning for election integrity because it can make misinformation appear more credible. Because the AI presents this data with a high degree of confidence, a user might not realize the information is synthetic. The persistence of these errors suggests that the underlying data sources are becoming increasingly polluted with machine-generated content.

District identification failures

The research highlighted a specific technical failure in how chatbots map physical addresses to political boundaries. The study found that chatbots failed to identify which congressional district an exact address was located in 48 percent of the time. This is a fundamental requirement for any accurate voter guide, as it determines which representatives a citizen can actually vote for.

Even in the instances where the chatbots were successful in finding a district, the accuracy was still significantly compromised. The paper noted that those that did manage to find a district got it wrong 22 percent of the time. These numbers suggest that the models struggle with the granular, localized data required for accurate civic participation.

For a user, these errors could mean being directed to the wrong candidates or being told they are eligible to vote in races that do not apply to them. The high failure rate underscores the limitations of current large language models when tasked with high-stakes, fact-dependent queries. It highlights a gap between the models' ability to generate fluent text and their ability to maintain factual accuracy.

Current user adoption

Despite the prevalence of these errors, the number of people currently using AI to research elections appears to be relatively small. A survey conducted by the Pew Research Center in late October found that a small share of US voters have used a chatbot like ChatGPT, Copilot, or Gemini for election information. This suggests that while the technology is capable of spreading misinformation, it has not yet become a primary source for the majority of the electorate.

The low adoption rate provides a small buffer against widespread misinformation for the current election cycle. However, the presence of these errors in the models' outputs remains a concern for future use. As more people turn to AI for quick answers, the risk of these hallucinations impacting public perception of the democratic process increases.

The goal for any nation would be for zero voters to receive AI-generated misinformation regarding their right to vote. Given the current state of the technology, the report suggests that we may have to manage these inaccuracies as they persist. The challenge lies in ensuring that users remain aware of the limitations of these tools when seeking critical civic information.

The path forward

The findings from the Washington Post serve as a warning about the reliability of AI-generated content in sensitive contexts. As models continue to be integrated into search engines and personal assistants, the risk of 'slop' entering the information ecosystem grows. Developers face the challenge of creating better filters to prevent models from learning from and repeating synthetic errors.

For the general public, these results emphasize the need for skepticism when using AI for local news or government services. While the chatbots are convenient for summarizing long texts or brainstorming ideas, they are currently unreliable for specific data points like election dates or candidate names. Users should prioritize official government websites for any information regarding their specific voting requirements.

Looking ahead, the interaction between AI and political information will likely become a major focus for researchers and policymakers. The goal will be to break the cycle of AI-generated misinformation by improving the quality of the data these models are trained on. Until then, the accuracy of AI-generated voting guides remains a significant hurdle for the technology's role in civic life.