ChatGPT recommended a candidate in 79% of chats once users shared their views

ChatGPT recommended a candidate in 79% of chats once users shared their political views with it. That number comes from a Yale study that asked nearly 2,500 people across 35 states with Senate races to query ChatGPT, Claude, Grok, or Gemini about the election — and it captures the core problem the researchers wanted to document: chatbots don't just retrieve election information, they actively weigh in on who you should vote for.

The study ran for a few weeks in October, leading up to the Nov. 3 midterm elections. Participants were adults who already use chatbots weekly, which matters: the researchers weren't dropping random volunteers into an unfamiliar tool. They were watching how real, habitual chatbot users behave in an election context.

The stakes are straightforward. Pew Research found that 42% of US adults use chatbots to search for information, and most search tools — Google above all — now put a layer of AI between you and the results. When that layer summarizes, filters, or recommends, it stops being a neutral pipe and starts shaping what voters learn.

What did voters actually ask?

Most users asked for basic facts: who was on the ballot, where to vote, how the process works. Those queries have relatively little room for error, because the internet carries a large volume of reliable data on the mechanics of an election.

But 36% of participants asked a chatbot to help them pick a candidate to vote for — usually by describing their priorities and asking the chatbot to assess which candidates matched them. That's a different kind of task. Interpreting candidate positions depends heavily on which sources a chatbot draws on. A statement from a candidate's own website is more reliable than a social media clip in which an opponent characterizes them a certain way.

In one example from the study, a user asked ChatGPT to give the "pros and cons" of each candidate in Virginia — an inherently subjective question dressed up as a neutral one. The takeaway: the riskiest election queries are the ones that ask the model to judge, not just to list.

How differently did the chatbots behave?

The four models diverged sharply when users asked who to vote for. The researchers described the pattern plainly:

"When users asked who to vote for but had not yet said anything about themselves, only Grok regularly recommended a candidate, in 14% of chats," they wrote. "Once users shared their views, ChatGPT recommended a candidate in 79% of chats and Claude in 40%, with Grok and Gemini in between."

The mechanism here is worth pausing on. When a user shares their views, the chatbot gains material to match against candidate positions — and in ChatGPT's case, that turns into a recommendation in most conversations. The model stops summarizing and starts advising.

That behavioral gap between tools is itself a finding. Four chatbots, same question type, radically different rates of candidate endorsements. Voters have no way of knowing which behavior they're getting before they ask.

How big is this in the real population?

Pew's numbers put the study in scale. Only 8% of US adults reported using chatbots to find election-related information in this year's election — a small slice of the electorate, but not a trivial one in close races. And what that group asks is exactly the sensitive stuff: 76% were inquiring about where candidates stand on certain issues, and 71% were fact-checking information they'd heard about candidates or issues.

Those are judgment-heavy questions, and chatbots are still not reliable summarizers. The misinformation risk compounds because most users don't check a chatbot's linked sources — at least when it comes to Google's AI Overviews. At least one study found AI platforms gave users "outdated or inaccurate" information that could mislead voters, in some cases by missing a deadline or omitting a step about voting logistics.

This election cycle was particularly plagued by AI-generated deepfakes and misinformation, as the tools have become far more sophisticated since the 2024 election. OpenAI and Anthropic reportedly set up election "war rooms" to monitor misuse and information accuracy, according to exclusive reporting by Politico and Semafor. The takeaway: election misinformation isn't a hypothetical risk anymore — the labs themselves built monitoring operations around it.

What did the safeguards tests find?

The Brennan Center for Justice tested this in August, and the results cut both ways. Chatbots pushed back surprisingly well against certain falsehoods related to election disinformation campaigns. But the models easily generated "misleading election-related images and video" when asked.

The Brennan Center's study, which tested ChatGPT, Gemini, Grok, Meta AI, Runway, and Flux.2, didn't mince words about the guardrails: "Many of their safeguards for preventing this were ineffective, inconsistently applied, or functionally absent."

There's also a subtler risk the Yale researchers flag: cognitive offloading. The political media landscape is already dense and hard to parse for a voter trying to stay informed. Layering a chatbot — with opaque training data, filters, and other settings determined at the discretion of AI labs — on top of that search process can further muddy the waters while appearing to be a convenient and neutral summarizing tool. The danger isn't only what the chatbot gets wrong; it's that it looks like an impartial referee while doing the filtering.

Why does this study's method matter?

Earlier studies tested how chatbots respond to election questions using standardized queries. The Yale approach let users engage with chatbots the way they normally would. The researchers found their queries and standardized study queries show a lot of similarities, but the distinction is valuable for evaluations.

"This is a small step toward evaluations that measure AI's political influence on the questions users actually ask," the researchers noted.

Outside of self-reported data like this, only the AI labs themselves hold records of what users really ask about elections. In October, Anthropic analyzed 250,000 conversations with Claude and determined people "rarely ask Claude whom or what to vote for," finding that "0.51% of all conversations are specifically about elections or voting," though it expected election-related queries to increase closer to Nov. 3.

"More often, we observed people asking Claude questions that could help them reach their own conclusions, like where candidates stand on a particular issue. Requests that violated our election integrity policies were even less common, appearing in about 0.002% of conversations," Anthropic's report said.

What comes next?

The Yale researchers said the next step is determining how accurate and unbiased models are in their responses, what models learn about users, and whether their recommendations actually match users' needs.

One observation from the study points at why that matters. "In multi-turn chats, users often answered the chatbot's questions, and those answers led to its recommendations. The questions a model asks shape what users tell it and what it recommends," the researchers explained. In other words, the influence runs in both directions — the model doesn't just answer the voter, it steers the conversation that produces the answer. As chatbots keep absorbing the search function, measuring that steering effect is where election-integrity research is heading next.