Grok And GPT Had To Choose: A Look At The AI Ballot Results

🔍 Read the full analysis: Grok And GPT Had To Choose: A Look At The AI Ballot Results 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 gave sharply different candidate recommendations in a Convovo test covering 479 races for Congress and governor in 2026, according to the research studio’s findings reported by TechCrunch and summarized by ThorstenMeyerAI.com. When prompted through developer APIs to name one candidate, GPT picked Democrats in most of its major-party selections, while Grok mostly picked Republicans; the result highlights how an AI recommendation can depend on the model, its web search and the prompt used.

Convovo, a conversational AI research and design studio led by Cornell University professor Tapan Parikh, ran the test on October 21 using the companies’ APIs with built-in web search, default settings and no researcher-written system prompt. Each call asked the model who the user should vote for in a race and instructed it to name one candidate without hedging. The set comprised 34 Senate races, 410 House districts and 35 governor races, each with one Democrat and one Republican. Each model was asked three times per race.

Among answers that selected a major-party candidate, Convovo counted GPT naming a Democrat in 1,082 of 1,111 answers and Grok naming a Republican in 952 of 1,038. By race, GPT picked Democrats in 451 and Republicans in 10; Grok picked Republicans in 408 and Democrats in 27. Refusals were counted separately: GPT refused 313 of 1,437 calls and Grok refused 375. These figures describe this test and its particular prompt, not a general measure of either model’s behavior.

Convovo also tested Claude Sonnet 5.5 and Gemini 3.8 Flash on the first ten Senate races, twice per race. Both declined all 20 calls, after which the researchers stopped testing them. That small sample shows their responses to this specific prompt; it does not establish that either model is politically neutral. Convovo coded answers by their opening sentence and manually reviewed unclear responses, according to the account of the study.

Search Sources Shaped the Recommendations

The contrast matters because AI assistants may be asked for voting advice, while users may not know how much a recommendation depends on a search system, an API prompt or instructions supplied by the developer. Convovo found that GPT and Grok returned substantially different sets of sources. GPT searches frequently included Wikipedia, the Associated Press, Axios and Reddit; Grok’s results often included Ballotpedia, The New York Times, X, iVoterGuide and Turning Point Action. The study’s source counts describe search results in this test, not necessarily what either model would cite in other settings.

Convovo’s search-disabled rerun suggests the two systems behaved differently when web results were removed. With candidate names provided, GPT continued to select Democrats more often, while Grok declined most calls. Parikh interpreted this as evidence that GPT’s lean was more embedded in model weights and Grok’s outputs depended more on search. That is the researcher’s interpretation, not a settled explanation of how either system works. For organizations deploying chatbots, the findings point to a practical risk: the deployed product can reflect not only a model but also its instructions and information sources.

A Forced Choice, Not an App Poll

The distinction between an API experiment and a consumer product is central to interpreting the results. Convovo’s prompt explicitly demanded a candidate recommendation. The consumer ChatGPT and Grok apps reportedly declined the plain version of that question during checks described in the source account. OpenAI told TechCrunch that the experiment used a developer API and a prompt demanding a recommendation, and did not reflect the experience people have in ChatGPT.

The test also found that the models’ choices often ran against the partisan lean of a district. TechCrunch reported that GPT picked the Democrat in 57 of 67 heavily Republican districts, while Grok picked the Republican in 47 of 72 deep-blue districts. Those are observations from the tested races, not evidence that either model’s answers changed voters’ choices. The source account also says an AI judge, checked against hand-coded answers, categorized rationales: health care and abortion appeared often in GPT’s Democratic picks, while immigration, taxes and energy appeared often in Grok’s Republican picks.

Convovo reported that its search-disabled test covered all 479 races, again with three draws per race, and supplied the nominees’ names. GPT chose Democrats in 387 races and Republicans in 37, while declining 43. Grok declined 441. The comparison offers a clue about the role of search in this test, but the results do not reveal all the factors behind the models’ outputs.

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

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

Limits of the Test and Its Findings

The results are limited to one forced-choice prompt, specific model versions and a single test date. The consumer apps may respond differently, and the test does not show how people would interpret or act on a recommendation. The small Claude and Gemini sample cannot support a broad conclusion about their political behavior.

Convovo also reported that xAI’s API added about 1,240 input tokens to each Grok call that researchers could not see or remove. Parikh described these as potentially hidden instructions, but their contents are not public; the source account does not establish what they said or whether they caused the results. The search comparisons and model-weight explanation likewise remain interpretations of the observed outputs, rather than proof of the underlying mechanism. The available source material does not provide a complete account of independent validation or uncertainty estimates for the coded results.

Further Testing of AI Election Advice

The immediate question is whether the pattern holds under different prompts, model versions, election cycles and product settings. Convovo’s published test and repository provide material for scrutiny, but the reported results are not a substitute for matched tests of the systems as people actually use them. The source account does not identify a scheduled follow-up study or a response from xAI about the additional API tokens.

For now, the researchers’ findings support caution rather than a conclusion that either chatbot should be treated as an election guide. Companies using AI assistants can test their own deployments with matched prompts and check which sources shape answers. Voters can compare candidates using primary and independent sources instead of relying on a chatbot’s recommendation.

Key Questions

What did the GPT and Grok test find?

In Convovo’s test of 479 2026 races, GPT-6.1 Sol mostly named Democrats and Grok 4.7 mostly named Republicans among answers that selected a major-party candidate. The findings apply to the test’s forced-choice API prompt.

Did ChatGPT and Grok’s consumer apps give the same answers?

The study used developer APIs, not a poll of ordinary consumer-app conversations. The source account says the apps declined the plain question, and OpenAI said the API prompt did not reflect the ChatGPT user experience.

Does the test prove the models are politically biased?

The results show a strong partisan pattern in this particular experiment. They do not by themselves establish why the pattern occurred or how the models behave across other prompts, settings or versions.

What happened when web search was disabled?

With candidate names supplied, GPT continued to choose Democrats more often, while Grok declined most calls. Convovo’s lead interpreted that as evidence of different roles for model weights and search, but the results do not prove the underlying cause.

What remains unknown about Grok’s API instructions?

Convovo reported that the API added about 1,240 input tokens researchers could not inspect or remove. The contents are not public, so it is unknown what they said or whether they affected the answers.

Source: ThorstenMeyerAI.com