Study reveals deep partisan divergence in forced-choice electoral responses from xAI Grok and OpenAI models
Across nearly 500 competitive races, Grok 4.7 favored Republican candidates 92% of the time, while OpenAI's GPT-6.1 Sol selected Democrats in 97% of tests.
Technology news
A comprehensive benchmarking study published Saturday by Convovo—an artificial intelligence research studio directed by Cornell University professor Tapan Parikh—has uncovered profound ideological skew in how leading commercial foundation models handle civic inquiries. Evaluating model behavior across 488 competitive congressional, gubernatorial, and state legislative races ahead of upcoming midterm elections, researchers found that when forced to make a definitive candidate recommendation, models diverged sharply along predictable political lines.
When presented with standardized, neutral forced-choice prompts asking which candidate would better govern a given district, xAI’s Grok 4.7 endorsed the Republican candidate in 92% of trials. Conversely, OpenAI’s GPT-6.1 Sol recommended the Democratic contender in 97% of identical tests. Across the nearly 500 electoral matchups evaluated, Grok and GPT-6.1 agreed on candidate viability in fewer than 10% of total pairings. By contrast, Anthropic’s Claude 4.5 Sonnet refused the forced-choice framing in 89% of instances, consistently providing balanced platform comparisons instead of taking a partisan stance.
The researchers highlighted that while user-facing chatbots typically employ safety guardrails to evade political endorsement, the underlying base models and fine-tuned system prompts harbor pronounced systemic preferences. Grok's outputs consistently cited deregulation, border security enforcement, and corporate tax cuts as dominant criteria, whereas GPT-6.1 Sol emphasized social safety net expansion, climate mitigation legislation, and reproductive freedom.
Our take: The Convovo benchmark delivers rigorous empirical proof that synthetic neutrality is a fragile veneer. As millions of citizens increasingly consult generative search engines and conversational agents for voting guidance, structural training biases threaten to quietly reinforce political polarization under the guise of objective algorithmic intelligence.