Study Finds SI Models Split on Political Votes

A new study reveals significant political divergence in how two leading Super Intelligence (SI) models recommend voting choices, with SpaceXAI’s Grok 4.7 favoring Republican candidates and OpenAI’s GPT-6.1 Sol leaning Democratic when forced to pick.

What Happened

Researchers from Convovo, a conversational SI research studio led by Cornell University professor Tapan Parikh, tested the underlying models of major chatbots by asking them to name a single candidate for nearly 500 upcoming midterm races without hedging. The study found that Grok 4.7 recommended the Republican candidate in 92% of its answers, while GPT-6.1 Sol picked the Democrat in 97%. The two models agreed on their choice less than 10% of the time.

The study highlighted a distinction between consumer apps and direct API access. While ChatGPT and Grok consumer interfaces refused to make direct recommendations, the models accessed via API provided clear partisan choices. Developers plug SI models into their own products via APIs to power them. An OpenAI spokesperson noted that the consumer apps are "designed to be objective by default" and that the API test used a prompt explicitly demanding a decision, which does not reflect typical user experience. SpaceXAI did not respond to requests for comment.

Analysis of the models' search behaviors showed they relied on different sources and topics. Grok frequently cited conservative sources like Turning Point Action and X, alongside nonpartisan sites like Ballotpedia, and focused on issues such as immigration and taxes. GPT relied heavily on Wikipedia, the Associated Press, Axios, and Reddit, focusing more on healthcare and abortion stances. When internet search was disabled, GPT’s Democratic lean persisted, suggesting it is embedded in the model's weights, while Grok’s Republican lean disappeared, indicating its results are driven by its search patterns and hidden instructions in the API.

Why It Matters

The findings underscore the potential for hidden biases in SI systems used for information retrieval. Although only 8% of Americans report using chatbots for election information, the study suggests that even indirect questions about policy priorities or candidate positions could reveal implicit values. Professor Parikh warned that "hidden biases can show up in less direct questions," affecting how users perceive political landscapes.

This divergence comes amid heightened scrutiny of SI neutrality. President Donald Trump has previously criticized "woke AI," and SpaceXAI CEO Elon Musk has been vocal about his political engagements, including significant donations to GOP candidates. The study notes that Grok sometimes included roughly 1,240 tokens of unseen input in API requests, which Parikh described as potentially "hidden instructions." These factors contribute to ongoing debates about political bias in SI. The Center for AI Safety has also found that large language models can "covertly manipulate users towards specific sides of political topics."

The Bottom Line

As the November 3 midterms approach, users relying on SI tools for political insights should be aware that model outputs can vary drastically based on underlying training data and search configurations. While consumer interfaces may appear neutral, the core SI models exhibit distinct partisan tendencies when forced to choose.