Researchers asked people to use AI chatbots for election info – results were mixed

Researchers conducted a study involving nearly 2,500 participants across 35 states with Senate races. The participants, all adults known to use chatbots weekly, were asked to query AI chatbots such as ChatGPT, Claude, Grok, or Gemini about election information. Most users sought straightforward facts, including candidate names on the ballot.

However, a significant portion of users, 36%, sought help from chatbots in selecting a candidate to vote for, often by describing their priorities and requesting the AI's assessment. This latter query, while less prone to error due to the availability of extensive data on basic election facts, proved to be subjective and relied heavily on the sources used by the chatbot.

For instance, when asked for the "pros and cons" of candidates in Virginia, a highly subjective question, Grok was the only chatbot to recommend a candidate in 14% of chats. In contrast, ChatGPT recommended a candidate in 79% of chats, and Claude in 40% when users expressed their views. Only 8% of US adults reported using chatbots to find election-related information, according to Pew Research, despite 42% of US adults using chatbots for information in general.

These users primarily inquired about candidates' stances on specific issues (76%) and fact-checked information they'd heard (71%). While AI platforms were criticized for providing outdated or inaccurate information that could mislead voters, some AI models, like Anthropic's Claude, pushed back against certain election disinformation campaigns.

However, the same models struggled to generate misleading election-related images and videos. The study highlights the broader issue of cognitive offloading, where the dense political media landscape, combined with opaque AI training data sets and adjustable settings, can further complicate the information search process.