Study: Grok and GPT pick opposite parties in US midterms

A Convovo study reported by TechCrunch on 24 October, led by Cornell's Tapan Parikh, forced Grok 4.7 and GPT-6.1 Sol to pick one candidate in each of nearly 500 US midterm races. Grok chose the Republican 92% of the time and GPT the Democrat 97%. They disagreed in more than 90% of races.

Researchers used developer APIs and a prompt demanding a single choice with no hedging; the consumer apps refused to recommend candidates. The lean held in opposite-leaning areas: GPT picked the Democrat in 57 of 67 heavily Republican districts, Grok the Republican in 47 of 72 deep-blue ones. With search off, GPT still chose Democrats in 24 of 30 races while Grok declined in 27 of 30.

Parikh suggests GPT's lean sits in its weights while Grok's depends on search terms and sources. OpenAI said ChatGPT is designed to be objective and that a developer API with a forcing prompt does not reflect the consumer experience. SpaceXAI did not respond in time. Sources differ on whether search was on in the main test.

The test measures a crafted prompt, not real conversations, but thousands of companies connect these models through the API without consumer guardrails, so the findings matter most for developers.

A forced-choice test removes the option to hedge, revealing hidden leanings. A system prompt holds the hidden rules companies give consumer apps; API users must build those guardrails themselves.

ChatGPT is restricting photorealistic political deepfakes before the midterms. The study shows neutrality often depends on product rules rather than the model itself.

Published days before the midterms, the study raises the chance of regulatory pressure on political neutrality.

• Write your own rules for politics and sensitive topics in customer-facing bots.
• Test sensitive questions before launch.
• Compare two models on important outputs to spot hidden bias.