Studies find AI chatbots often get election information wrong

In short: More Americans are asking AI chatbots for election information, but studies find the answers often contain mistakes and sometimes draw from partisan sources.

What's going on

Pew Research data cited by WIRED suggests AI chatbots are now common tools for voters. About half of surveyed US adults say they have used an AI chatbot at least once. Pew also found 8 percent said they used chatbots ahead of the midterms to look up information on candidates and policies.

WIRED reports that these tools can return information that is out of date, incorrect, or made up. This “made up” problem is often called a hallucination, which is when a chatbot answers confidently even though it does not actually know (like a student guessing on a test).

Outside testing has found frequent errors. Research by Forum AI found that almost 40 percent of election related answers from major chatbots included at least one factual error. WIRED highlighted examples, including Google’s Gemini giving a wrong Pennsylvania mail ballot request deadline, and Anthropic’s Claude mixing up a lawmaker with another person who had the same name.

Another concern is where chatbots get their information. NewsGuard found that mainstream chatbots often cite “pink slime” sites, which can look like local news but are funded by political groups. NewsGuard said ChatGPT cited these sites in more than 70 percent of its responses in their testing, while xAI’s Grok did so about 30 percent of the time.

What to watch

As more people use chatbots like ChatGPT, Claude, and Gemini for voting questions, researchers expect more attention on accuracy and sourcing. For now, the safer approach is to double check chatbot claims against official election websites and trusted local news.

Source: Wired