Did Grok And GPT Pick The Same Side? Inside An AI Vote Test

🔍 Read the full analysis: Did Grok And GPT Pick The Same Side? Inside An AI Vote Test on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

GPT-6.1 Sol and Grok 4.7 named opposing parties in a Convovo test that asked the models who to vote for in 479 U.S. races in 2026. Under a developer-API prompt requiring one candidate, GPT named Democrats in 1,082 of 1,111 answers that chose a major-party candidate, while Grok named Republicans in 952 of 1,038 such answers. The results show how these models responded in a particular forced-choice test—not how all users will experience them or whether either model is politically neutral.

Convovo, a conversational AI research and design studio led by Cornell University professor Tapan Parikh, ran the test on October 21 using the vendors’ APIs, default settings and built-in web search. It asked each model three times about each race, using the same instruction: “Name one candidate” and do not hedge or list both sides. The 479 races comprised 34 Senate contests, 410 House districts and 35 governor races, each with one Democrat and one Republican, according to Convovo’s study page and public repository.

The headline percentages count only answers that selected a major-party candidate. The models also abstained: GPT refused 313 of 1,437 calls, and Grok refused 375. Counted by race, GPT picked the Democrat in 451 races and the Republican in 10; Grok picked the Republican in 408 and the Democrat in 27. TechCrunch reported that the two models agreed in fewer than 10% of cases. These counts describe the test’s answers and coding method, not a forecast of election outcomes or a measure of voter behavior.

Convovo also gave Claude Sonnet 5.5 and Gemini 3.8 Flash the first ten Senate races, twice each. Both declined all 20 calls, after which the researchers did not continue testing them. That small sample shows refusals to this prompt; it cannot establish that those models have no political lean. Convovo coded answers from their opening sentences and manually reviewed unclear cases. An AI judge categorized stated reasons, with some answers checked against human coding.

Why Chatbot Recommendations Matter

The findings matter because a chatbot can shape how a user encounters candidates, even when its output is not an explicit endorsement in a consumer app. Convovo’s test suggests the answer can depend on the model, the wording of the prompt, and the information returned by search. A company using an API-powered assistant could inherit those behaviors, along with its provider’s system instructions and the sources its search tools retrieve.

The study does not establish that users were persuaded, that the models’ choices reflect deliberate political preferences, or that the answers affected voting. It does identify a practical risk for organizations deploying assistants around elections: a single screenshot may not represent the system’s behavior, while repeated matched tests can reveal patterns under controlled conditions. For voters, the results are a reminder to check candidate claims and sources rather than treat an AI recommendation as a substitute for their own decision.

Search Results Changed the Test

In the initial runs, both models used web search, but Convovo found little overlap in the sources returned. Its reported results included Wikipedia in 93% of GPT answers, the Associated Press in 74%, Axios in 57% and Reddit in 50%. For Grok, Ballotpedia appeared in 100% of answers, the New York Times in 81%, X in 41%, iVoterGuide in 39% and Turning Point Action in 23%. These are source appearances in the study’s answers, not measures of how much each source influenced the final choice.

Convovo manually reviewed some of the cited sites and raised concerns about their descriptions and content. It said a guide site returned for 45% of GPT answers called itself nonpartisan while ranking progressive Democrats above Republicans. The researchers also reported that GPT did not return Ballotpedia, which they attributed to OpenAI’s search index. These observations are the study’s findings; they do not by themselves prove that a particular source caused a given recommendation.

When Convovo reran all 479 races with search disabled and supplied the nominees’ names, GPT still chose Democrats in 387 races and Republicans in 37, declining 43 times. Grok declined in 441 races. Parikh interpreted this contrast as evidence that GPT’s pattern was more embedded in the model’s responses, while Grok depended more on search. That is the researchers’ interpretation of this experiment, not a definitive explanation of either model’s internal processes.

“For GPT, it has mostly decided already…it’s in the weights.”

— Tapan Parikh, Convovo’s lead, speaking to TechCrunch

Limits of the API Evidence

The test does not settle how the models behave for typical users. Convovo used APIs and a prompt that explicitly demanded a candidate; OpenAI said that setup differs from ChatGPT’s consumer experience, and Convovo reported that the apps refused the direct question. The source material also describes app responses under alternative framings, but the available account is incomplete, so those results cannot be fully evaluated here.

Other limits include the forced-choice wording, three runs per race, the selection of races with one Democrat and one Republican, and the study’s answer-coding process. The reported results do not show that the models recommended candidates consistently across other prompts, dates, model versions or user settings. Convovo also said xAI’s API added about 1,240 input tokens it could not see or remove. Parikh described these as potentially hidden instructions, but their contents are not public; their effect on the responses remains unconfirmed.

The study’s source counts and explanations are not independent proof of political bias or intent. Nor do they establish that the models’ outputs affected anyone’s vote. The available evidence supports a narrower conclusion: in Convovo’s stated test conditions, the two models produced sharply different party patterns, and removing search changed their behavior in different ways.

Further Tests Across User Settings

Convovo’s published study page and repository provide the basis for further review of its prompts, race list and answer coding. The next useful checks would repeat the test across model versions, prompt styles and consumer-facing products, while documenting settings and search results. Independent replication could help determine whether the reported patterns persist beyond the October 21 API runs.

Organizations deploying election-related assistants can test their own systems with matched prompts before users rely on them, record refusals as well as selections, and review which sources appear in answers. Voters can request information about both candidates and verify it against reliable sources. The study does not establish what OpenAI or xAI will change, or whether either company plans a formal response beyond the statements reported to TechCrunch.

Key Questions

What did the Convovo test find?

Across its forced-choice API calls, GPT-6.1 Sol mostly named Democrats and Grok 4.7 mostly named Republicans. The headline counts exclude refusals and cover Convovo’s 479 selected 2026 races.

Did the models answer every time?

No. GPT refused 313 of 1,437 calls and Grok refused 375. Convovo treated refusals as abstentions rather than candidate selections.

Does this prove that the models are politically biased?

The results show a strong party pattern under this study’s prompt and API settings. They do not, on their own, establish intent, explain every cause of the pattern, or show how the models behave across all user experiences.

What happened when web search was disabled?

GPT continued to select Democrats more often, choosing them in 387 races in the rerun. Grok declined in 441 of the 479 races. Convovo’s lead interpreted the contrast as evidence that search played a larger role in Grok’s initial pattern.

Did the test measure how voters were affected?

No. It measured chatbot responses, not user behavior or voting outcomes. The study does not establish whether anyone changed a voting decision because of an answer.

Source: ThorstenMeyerAI.com