Study Finds Grok Picks Republicans in 92% of Races While GPT Picks Democrats in 97%

A study by research firm Convovo, led by Cornell professor Tapan Parikh and reported by TechCrunch, tested the political leanings of Grok 4.7 and GPT-6.1 Sol across nearly 500 U.S. midterm election races. Researchers used a prompt that left no room for evasion, asking each model to name one candidate and barring it from hedging or listing both sides. Grok named the Republican in 92% of races; GPT named the Democrat in 97%. The two models agreed on a recommendation in fewer than one in ten contests. To identify the source of Grok's lean, researchers disabled the model's internet-search retrieval in a subsample of 30 races. With search off, Grok's Republican preference collapsed and the model refused to name any candidate in 27 of those 30 cases, strongly suggesting that live web retrieval of partisan-leaning content drove the bias rather than values embedded during training. GPT's Democratic lean held in the same subsample, pointing to a training-data origin for its tilt. The findings arrived less than two weeks before the November 3 midterms and reignited debate about whether AI chatbots should be allowed to recommend candidates at all, and whether retrieval-augmented AI products can inadvertently amplify partisan information ecosystems.