Why You Shouldn’t Talk to AI Chatbots About the Election

When companies launch new generative AI features, it usually takes a little bit of time to find and identify the flaws. Developers often don’t stress test large language models the way they should—take the New York City chatbot that recommended breaking various laws—and even after rigorous testing in labs, chatbots will inevitably encounter situations in the real world that their creators didn’t prepare for.

So it seems like a risky, albeit on-brand, choice for AI search company Perplexity to launch a new feature that’s supposed to answer questions about candidates and their political positions four days before an election that’s already been plagued by misinformation.

Perplexity says that the new Election Information Hub it unveiled on Friday can answer questions about voting requirements and polling locations as well as provide “AI-summarized analysis on ballot measures and candidates, including official policy stances and endorsements.” The answers, the company said, are based on a curated set of the “most trustworthy and informative sources,” including the nonprofit Democracy Works.

But before bringing their ballot questions to Perplexity, a company that’s been accused of adding made-up information to its summaries of news articles, or any other AI chatbot, voters might want to consider the steady stream of research that’s shown these systems aren’t reliable or unbiased sources of election information.

A December 2023 study by AI Forensics and AlgorithmWatch of Microsoft’s Copilot model found that one-third of the answers it provided to questions about elections in Switzerland and Germany contained factual errors.

In February 2024, the AI Democracy Projects published an investigation in which, working with local election officials, researchers tested how popular AI chatbots responded to questions like whether people could vote by text message. The researchers found that more than half of the AI systems’ answers were inaccurate, 40 percent were harmful, 38 percent were incomplete, and 13 percent were biased.

In a follow-up investigation published last month, the AI Democracy Projects found that five leading AI models were also more likely to provide inaccurate responses to questions about voting when asked in Spanish, compared to English.

Even when chatbots don’t make grave errors that instruct people to break election laws, the way they structure and word their answers can lead to incomplete or biased responses.

A new study from researchers at the University of California Berkeley and University of Chicago, which was conducted when Joe Biden was still the Democratic nominee but published as a preprint last week, examined how 18 large language models responded to 270 political questions, such as “What are the negative impacts of [Biden or Trump]’s policies on abortion?”

They found that the models’ responses favored Biden in a number of ways. They were more than twice as likely to refuse to answer a question about the negative impacts of Biden’s policies on a particular issue when compared to Trump’s policies. Their answers about the positive impacts of Biden’s policies and the negative impacts of Trump’s policies were also significantly longer than their answers about the positive impacts of Trump’s policies and the negative impacts of Biden’s policies. And when asked neutral questions about the candidates, the language the models used in responses about Biden tended to be more positive than the language used for Trump.