Yale study: AI chatbots often recommend candidates to voters

Big gaps between chatbots

When users asked who to vote for without saying anything about themselves, only Grok regularly named a candidate, doing so in 14% of chats. Once users shared their views, ChatGPT recommended a candidate in 79% of chats and Claude in 40%, with Grok and Gemini falling in between. Overall, ChatGPT made recommendations most often, at twice the rate of Claude.

The researchers also noted that in multi-turn conversations, the questions a model asks shape what users reveal and, in turn, what it recommends.

Why it matters

Factual queries leave less room for error, since plenty of data exists on who is running where. Interpreting candidate positions is riskier and depends heavily on the sources a chatbot draws on. Pew Research found that 42% of US adults use chatbots to search for information, and 8% used them for election information this year. Of those, 76% asked where candidates stand on issues and 71% fact-checked things they had heard.

Because chatbots are still unreliable summarizers, skipping the linked sources creates a misinformation risk. One study found AI platforms gave outdated or inaccurate voting information, such as missing a deadline. A Brennan Center study from August found chatbots resisted some election falsehoods but readily produced misleading election images and video, with safeguards described as ineffective, inconsistently applied or absent.

Anthropic, for its part, analyzed 250,000 Claude conversations in October and found 0.51% concerned elections or voting. People rarely asked whom to vote for, it said.

The Yale team says the next step is measuring how accurate and unbiased the responses are, what models learn about users, and whether recommendations fit users' needs.