Eigen RadarAI
Archive
Analysis

Grok and GPT pick opposite parties when pressed for voting advice

A forced-choice test of Grok 4.7 and GPT-6.1 Sol found opposite party preferences in US election recommendations. The experiment used developer interfaces and demanded a named candidate; the consumer apps initially declined the question. A smaller follow-up without web search recorded different patterns, while OpenAI said the API experiment did not represent ordinary ChatGPT use.

Artificial Intelligence··Night
Two different mechanical arms reach for blue and red blocks on a shared tray.

The same request, opposite party choices

Grok 4.7 and GPT-6.1 Sol favored opposite parties when researchers demanded a single candidate recommendation in US election races. Convovo, a conversational AI research studio led by Cornell University professor Tapan Parikh, compared answers through the models’ developer interfaces. Those interfaces let applications call a model directly. The test asked for one named candidate and explicitly discouraged hedging or listing both sides.[1], [2]

Among answers naming a major-party candidate, GPT favored Democrats in 97 percent and Grok favored Republicans in 92 percent. These results concern the specified prompt and model versions, with web search enabled.[1]

Search settings changed the follow-up

In a follow-up requested by TechCrunch, researchers supplied the candidates’ names and switched off search for 30 races. GPT chose 24 Democrats and one Republican and declined to choose in five. Grok declined in 27 and chose Democrats in the remaining three. Parikh interprets the contrast as potentially different mechanisms behind the preferences; the test did not directly inspect either model’s training data or hidden instructions.[1]

Consumer apps initially refused voting advice

The consumer ChatGPT and Grok applications initially refused the direct voting-advice question. OpenAI said the developer-interface test did not represent ordinary ChatGPT use and that the product should be objective in political conversations. SpaceXAI did not respond before TechCrunch’s publication. Parikh said later simple prompts elicited recommendations from the apps, but the reported experiment remains distinct from a field study of voters’ everyday conversations.[1]

References

  1. News sourceTechCrunchForced-choice election tests find contrasting party preferences in Grok and GPT↩1↩2↩3↩4
  2. News sourceTSNGrok and GPT favor opposite parties in forced election picks↩
AI safety and regulationThe frontier model race

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →