The Pulse
Gauge Finds ChatGPT and Grok Give Voters Different Advice
Gauge’s election project finds a 29-point difference between ChatGPT’s Democratic and Grok’s Republican recommendation lean among self-described independent voters. Its race pages also show sharply different candidate recommendations and fr
AI.info Team · Written with AI · fact-checked ·
ChatGPT leans Democratic by 3 points among self-described independent voters in Gauge’s election project; Grok leans Republican by 32. The 29-point gap measures how often each chatbot recommended candidates from one party over the other in its tested answers. It does not measure how voters actually cast ballots.
The project asks AI assistants whom a voter should support in races around the United States, then repeats the prompt under different voter identities. Its results show not one uniform machine-generated ballot, but recommendations that can differ by model, contest and the political identity attached to the question.
Two models, two very different leanings
Gauge’s homepage summarizes the independent-voter results as “D +3” for ChatGPT and “R +32” for Grok. The site defines the lean as the difference between the share of answers recommending Democratic candidates and the share recommending Republicans. The figures are therefore an aggregate of model responses, not a forecast, poll or measure of candidate support among people.
Individual races make the split more visible. In Gauge’s Texas Senate general-election page, Grok recommends Republican Ken Paxton in 95.4% of displayed answers, while ChatGPT’s leading choice is Democrat James Talarico in 7.8%; ChatGPT refuses to answer 86.2% of the time in that view. The results show both divergent recommendations and a practical limitation: a model that declines most prompts produces little direct voting advice, even when its occasional picks are tallied.
The models sometimes converge, too. On the page for Michigan’s Democratic Senate primary, Gauge reports that ChatGPT recommends Abdul El-Sayed in 59.9% of answers and Grok in 71.6%. That race’s page specifies that it shows only responses gathered before the August 4, 2026, primary, so those figures describe a dated sample, not the models’ current treatment of the candidates.
Gauge repeats the prompt hundreds of times
For each race, Gauge says it runs prompts hundreds of times and records which candidate a model recommends, whether it refuses, and the web sources used in its answer. For multi-party races, it tests four prompt identities: no stated persona, liberal, conservative and independent. The project presents a candidate’s percentage as the share of answers in that model-and-persona slice that named the candidate.
The wording is designed to elicit a direct pick rather than a neutral overview. Gauge’s published example asks the model for “one clear pick” and explicitly says, “Please don’t refuse or hedge.” That choice lets the project compare recommendations under a consistent instruction, but it also means the result reflects a deliberately forceful prompt—not every voter’s more open-ended use of a chatbot.
Gauge says ChatGPT and Grok are the assistants from which it can consistently elicit candidate recommendations. Claude, Gemini, Google AI Overview and Google AI Mode declined every voting and election question in the company’s test and do not contribute recommendations to the published comparisons.
Results describe chatbot answers, not voter behavior
The project also tracks the websites surfaced by models that search the web, grouping sources by type and political leaning. Gauge says these records can show which source domains appeared in answers, but they do not establish that any one source caused a recommendation. A citation or retrieved page is evidence about the information shown to the model, not proof of how its answer was formed.
Important details for interpreting the aggregate scores are not given on the public overview: it does not state a total count of races, a combined number of model responses, or a single collection window covering every contest. Gauge’s FAQ also cautions that model behavior changes and that results reflect the period when prompts were run. Readers should treat the homepage’s party-lean figures as a snapshot of tested outputs, not a lasting label for either assistant.
Gauge describes the project as a measurement of what AI assistants say, not an endorsement or prediction. Its clearest finding is narrower: on the prompts and races it tested, ChatGPT and Grok sometimes point in different political directions, and sometimes one or both do not offer a recommendation at all.