Forced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks

A study by conversational AI research studio Convovo, led by Cornell University professor Tapan Parikh, evaluated how developer AI models behave when forced to make definitive voting endorsements ahead of the upcoming November 3 midterm elections. According to TechCrunch's report, when researchers instructed SpaceXAI's Grok 4.7 and OpenAI's GPT-6.1 Sol to recommend a single candidate across nearly 500 election races without hedging or listing both sides, Grok recommended the Republican candidate in 92% of its answers, while GPT picked the Democrat in 97% of cases. The two underlying models agreed on a recommendation less than 10% of the time.

This piece details the testing methodology, source selection mechanics, internal query differences, and architectural distinctions between developer APIs and consumer-facing applications. It provides engineers, researchers, and voters with an analysis of how algorithmic search retrieval and model weights introduce divergent political leanings during automated election inquiries.

Forced to pick: what actually changed

The research examined raw API behavior rather than default consumer chatbots. When tested in consumer apps, ChatGPT and Grok refused to tell users who to vote for. An OpenAI spokesperson stated that ChatGPT is designed to be objective by default, noting that the study tested developer APIs using explicit prompts demanding a recommendation, which does not reflect the standard consumer chat experience. However, Parikh reported that simple jailbreak prompts could bypass consumer application guardrails to elicit candidate endorsements that matched the API findings.

Across nearly 500 races, the models showed strong partisan divergence even in districts with opposing voter demographics. In 67 heavily Republican districts, GPT recommended the Democratic candidate in 57 instances. Conversely, across 72 deep-blue districts, Grok recommended the Republican candidate in 47 cases. The primary exception observed for Grok occurred when a Republican candidate faced acute controversies, such as sexual harassment allegations or being labeled a Nazi, in which scenarios Grok shifted toward recommending the Democrat.

Forced to pick: how it works

Researchers probed the models via API with a standardized prompt requiring the system to name a single candidate without waffling. The models utilized internet search to justify selections, but their retrieval patterns and sources diverged significantly. Grok frequently retrieved material from conservative outlets alongside nonpartisan directories such as Ballotpedia, OpenSecrets, and VoteSmart. Grok incorporated a Christian conservative voter guide in 39% of its answers, Turning Point Action's website in 23%, and X in 41%. In contrast, GPT relied heavily on Wikipedia, Reddit, the Associated Press, and Axios, entities with which OpenAI maintains established content partnerships.

Retrieval queries also reflected distinct thematic focuses. GPT prioritized search terms relating to healthcare and abortion records, frequently querying Democratic candidates while occasionally omitting Republican opponents entirely. Grok ran search terms focused on immigration, crime, energy, and taxes, typically querying both candidates before selecting a recommendation. When researchers disabled web search across a 30-race test sample, GPT still selected the Democrat in 24 of 30 races, chose one Republican, and declined five, indicating its lean stems from training weights. Grok without web search refused to select a candidate in 27 of 30 races and chose the Democrat in the remaining three, demonstrating its partisan lean was driven primarily by its retrieval pipeline.

Forced to pick: why it matters now

The findings arrive less than two weeks before the November 3 midterms amid heightened scrutiny of political bias in artificial intelligence systems. Pew research indicates that only 8% of Americans currently use AI chatbots for election information, primarily seeking information on candidate policy stances rather than explicit voting instructions. However, Parikh noted that implicit values influence how models structure broader, indirect inquiries, including which issues a voter should prioritize or which candidate maintains the strongest platform on key topics.

The Center for AI Safety previously documented that large language models can covertly steer users through selective sourcing, vocabulary choice, sentence construction, and framing. In xAI's API, researchers discovered an additional technical factor: every request sent to Grok carried approximately 1,240 tokens of unseen, non-removable input. Parikh described these extra tokens as potential hidden instructions that are unlikely to originate from baseline model weights.

Forced to pick: who is affected

The immediate impact falls on third-party developers building election guides, automated civic tools, and conversational assistants that integrate these models through commercial APIs. Unlike end users chatting on consumer web interfaces, developer products interacting with model endpoints risk inheriting these pre-set retrieval preferences and hidden prompt structures. Parikh noted that users conducting general research into candidate platforms are exposed to these biases whenever models summarize and prioritize issues.

The study also intersects with political and financial activities surrounding both AI companies. SpaceXAI CEO Elon Musk has contributed upwards of $445 million to Republican candidates during the current midterms, surpassing the roughly $250 million he contributed during the 2024 election cycle, alongside leading the Department of Government Efficiency. OpenAI CEO Sam Altman, who previously donated to Democratic campaigns and later supported select Republicans, stated that he is not contributing funds to the 2026 election cycle.

Forced to pick: what to watch

Observers and engineers should monitor whether API providers introduce configurable controls for hidden system tokens and web-retrieval transparency. Grok's consistent inclusion of 1,240 injection tokens per API call highlights the role of undisclosed provider-level prompt wrappers in determining output behavior. Whether xAI documents or exposes these tokens will indicate how the company handles developer requests for unconstrained raw model access.

Voters seeking unmediated candidate information continue to have access to dedicated nonpartisan tracking repositories. Established platforms for verifying candidate stances, ballot initiatives, and campaign finance filings include Vote.gov, the League of Women Voters' vote411.org, Ballotpedia, VoteSmart, OpenSecrets, and state and county election offices.