When AI Is Forced to Pick a Candidate, Grok and GPT Split by Party

Introduction: Most consumer chatbots avoid directly telling people how to vote. Yet when researchers removed that option and instructed models to name one candidate without hedging, Grok and an OpenAI model produced strikingly different results. A study conducted by Convovo, led by Cornell professor Tapan Parikh, examined responses across nearly 500 races in the upcoming U.S. midterm elections.

Key findings

• Through their APIs, researchers asked GPT-6.1 Sol and Grok 4.7 a deliberately forceful question: who should the user vote for in a particular race? Grok recommended the Republican candidate in 92% of responses, while GPT selected the Democrat in 97%. The models disagreed in more than 90% of the races.
• The test should not be treated as a direct measurement of the normal ChatGPT or Grok user experience. The consumer applications generally refused to make a voting recommendation. OpenAI said the study tested a developer API with a prompt that explicitly demanded a choice.
• The models also appeared to look for different kinds of evidence. Grok frequently surfaced conservative voter guides, Turning Point Action, and X, while also using nonpartisan sources. GPT more often cited Wikipedia, the Associated Press, Axios, and Reddit. GPT placed more emphasis on healthcare and abortion, whereas Grok focused on immigration, taxes, crime, and energy.
• A follow-up test turned off web search and supplied candidate names directly. GPT still selected Democrats in 24 of 30 races, chose one Republican, and declined five times. Grok refused to choose in 27 of 30 cases; its three selections were Democrats. This points to a stronger role for learned model behavior in GPT and for search patterns in Grok, although it does not establish a complete causal explanation.

Why it matters

The broader issue is not simply whether a model is “left” or “right.” Political influence can appear earlier in the process, when an AI system decides which topics matter, which sources deserve attention, and how competing facts should be framed. A user asking about the most important election issues may therefore receive a politically shaped answer without ever requesting a recommendation.

The study also reported that requests sent to Grok’s API included roughly 1,240 additional input tokens that researchers could not inspect or remove. Parikh described these as potentially hidden instructions, but the available material does not prove that they alone caused the observed results. Nor should the findings automatically be generalized to every model version, prompt, or consumer interaction.

For voters, chatbots are better treated as research assistants than as voting substitutes. Candidate websites, local election offices, Vote.gov, Ballotpedia, VoteSmart, and multiple independent news sources can provide a stronger basis for comparison. For developers, publishing system instructions, search behavior, cited sources, and evaluation procedures would make political answers easier to scrutinize than broad claims of neutrality.

Source: TechCrunch AI