Grok Picks Republicans, GPT Picks Democrats in 2026 Midterm Test
Grok 4.7 chose Republicans 92% of the time while GPT-6.1 Sol picked Democrats 97%, exposing political splits in AI search and sourcing before midterms.
Summary
Convovo, led by Cornell University professor Tapan Parikh, forced API versions of SpaceXAI’s Grok 4.7 and OpenAI’s GPT-6.1 Sol to recommend one candidate across nearly 500 November 3, 2026 midterm races. Grok chose Republicans in 92% of answers, GPT chose Democrats in 97%, and they agreed under 10%. Consumer ChatGPT and Grok refused, although Parikh bypassed them with simple prompts that produced similar recommendations. OpenAI said the API test does not represent ChatGPT’s politically objective design; SpaceXAI did not comment. Pew says 8% of Americans use chatbots for election information, mostly to research candidate positions.
Grok cited X in 41% of answers, a Christian conservative voter guide in 39%, and Turning Point Action in 23%, alongside Ballotpedia, OpenSecrets and VoteSmart. GPT relied heavily on Wikipedia, the Associated Press, Axios and Reddit; OpenAI has content deals with the latter three. GPT emphasized healthcare and abortion, searched more for Democrats and sometimes ignored Republicans. Grok emphasized immigration, taxes, crime and energy, usually examining both candidates, but shifted Democratic when Republicans faced controversies. GPT chose Democrats in 57 of 67 heavily Republican districts; Grok chose Republicans in 47 of 72 deep-blue districts. Each Grok request contained about 1,240 unseen, unremovable tokens that may be hidden instructions.
Without internet search, GPT picked 24 Democrats and one Republican across 30 races, declining five; Grok declined 27 and chose Democrats three times. Parikh attributes GPT’s lean to model weights and Grok’s to search patterns. The Center for AI Safety has separately warned that LLMs can manipulate through selection, wording, framing, emphasis and sourcing. The October 24 study lands before November 3 amid Donald Trump’s attacks on “woke AI.” Elon Musk has donated over $445 million to GOP midterm candidates, after spending roughly $250 million supporting Trump and Republicans in 2024; OpenAI CEO Sam Altman says he will not fund 2026 races.
Positives
- Consumer ChatGPT and Grok refused direct voting recommendations under the study’s original prompt.
- Grok examined both candidates more often than GPT, which sometimes ignored Republican opponents entirely.
- Without internet search, Grok declined to choose in 27 of 30 races rather than maintaining its Republican preference.
- Ballotpedia, OpenSecrets and VoteSmart appeared among Grok’s sources, providing nonpartisan information alongside ideological material.
Risks & concerns
- Grok selected Republicans in 92% of answers, while GPT selected Democrats in 97%, with agreement below 10%.
- GPT chose Democrats in 57 of 67 heavily Republican districts, while Grok chose Republicans in 47 of 72 deep-blue districts.
- Every Grok API request included about 1,240 unseen and unremovable tokens that researchers suspect could contain hidden instructions.
- Grok repeatedly surfaced ideological sources, including X in 41% of answers and a Christian conservative voter guide in 39%.
- Simple jailbreak prompts bypassed consumer safeguards and produced recommendations broadly matching the models’ API preferences.
- Topic selection split politically, with GPT emphasizing healthcare and abortion while Grok prioritized immigration, taxes, crime and energy.