A new TechCrunch report highlights sharply different political recommendations from two AI models. The important question is what the experiment measured. A response to a prompt demanding one candidate is evidence about that test configuration; it is not, by itself, a measurement of ordinary chatbot conversations or of changes in anyone’s vote.
The denominator changes the headline
Convovo’s original AI Ballot project tested GPT-6.1 Sol and Grok 4.7 through their providers’ APIs with web search enabled. Its methods specify October 21 runs, three answers per race and a prompt insisting on a single recommendation without hedging. The published dataset covers 506 races, including contests outside the standard Democrat-versus-Republican comparison.
Among answers naming a Democrat or Republican in the standard comparison, GPT selected the Democrat about 97% of the time and Grok selected the Republican about 92%. Those percentages exclude refusals and other classifications. They should not be described as the share of all answers, races or voters. The project’s method notes also say the consumer ChatGPT and Grok applications refuse the tested question; these API results are not a direct census of their ordinary app behavior.
That is a substantial difference between tested systems, but its scope remains specific. Repeated outputs from a configured service do not isolate which component caused the pattern. Nor does one forceful prompt establish how a model responds to a neutral request for information. The result is most useful when the wording, serving channel, search setting and treatment of refusals travel with the percentage.
A usage survey asks a different question
Pew Research Center’s October 23 survey report supplies context about exposure, not a replication of Convovo’s experiment. Pew surveyed 5,273 U.S. adults from October 12 to 18 through its American Trends Panel. Eight percent said they were using AI chatbots for information about the 2026 elections.
Within that election-chatbot group, 76% reported seeking candidates’ issue positions and 71% sought checks on information encountered elsewhere. Forty percent reported looking for suggestions about their vote choices. The last figure concerns that subgroup, not 40% of all American adults. The questions allow multiple uses rather than assigning each person one exclusive activity.
These findings make a direct leap from model recommendations to electoral influence particularly difficult. A usage survey describes reported behavior; a model test describes generated outputs. Neither, on its own, measures whether a person accepted a recommendation or changed a decision because of it. Limited reported use does not erase concerns about biased information, but striking laboratory percentages do not quantify a population-wide voting effect either.
Political bias has more than one measurement
Earlier research offers a further methodological distinction. Reducing Political Manipulation with Consistency Training, first submitted in May 2026, examines how models treat paired political subjects rather than which candidate they choose. Its authors separate consistency of rhetoric and framing from consistency of depth and engagement. The associated research project explains that a model can perform differently on those two dimensions.
That is older background, not another newly released election finding. It helps show why a candidate-preference count cannot exhaust the question of political bias. A system might avoid naming a candidate while still treating counterpart topics differently. Equally, a refusal is not automatically evidence that the system would provide accurate, balanced information in a different conversation. These are distinct propositions requiring distinct tests.
The appropriate conclusion from the new report is therefore narrower than assigning each chatbot a permanent party identity. The published experiment documents a pronounced pattern under specified conditions. Assessing its practical significance calls for comparisons across prompts and products, transparent reporting of abstentions, and evidence about user responses. Keeping those questions separate makes the result easier to scrutinize without turning an API benchmark into an unsupported claim about voters.










