A recent study shows that when prompted to choose a candidate in nearly 500 midterm races, Grok’s AI model overwhelmingly recommended Republican candidates 92% of the time, while OpenAI’s GPT-6.1 favored Democrats in 97% of cases, with the two aligning under 10% of the time.
- Grok recommends Republicans in 92% of midterm races tested.
- OpenAI’s GPT-6.1 selects Democrats in 97% of cases.
- Models disagree in over 90% of election race recommendations.
What happened
A study by conversational AI research group Convovo tested two prominent AI models—SpaceXAI’s Grok 4.7 and OpenAI’s GPT-6.1 Sol—by directly asking them to name a specific candidate to vote for in nearly 500 upcoming US midterm election races. The consumer-facing chat apps were designed to avoid making direct endorsements, but via API access, the models were forced to pick a candidate without hedging.
Results showed a strong partisan split: Grok recommended Republican candidates in 92% of responses, while GPT chose Democrats 97% of the time. The two models agreed on the same candidate less than 10% of the time, illustrating substantial divergences in their political leanings when compelled to choose.
Why it matters
These findings reveal underlying biases in AI models that could influence voters indirectly through selective information framing and topic emphasis. Grok frequently referenced conservative-leaning sources and issues like immigration, taxes, and crime, while GPT leaned on Democratic concerns such as healthcare and abortion, often citing mainstream media and Wikipedia.
Given that AI-generated guidance is increasingly sought for political information—even if only 8% of Americans currently use chatbots for this purpose—the partisan slant in responses risks shaping perceptions and voter decisions. The research underscores the importance of transparency and neutrality in AI systems involved in civic engagement.
What to watch next
Monitoring how AI models evolve their approaches to political questions will be critical, especially as researchers explore jailbreak prompts that coax consumer versions of chatbots into partisan recommendations. Developers may need to improve safeguards against hidden biases that manifest subtly through topic selection and source citation rather than overt endorsements.
Legislators, AI ethics groups, and platform operators will likely examine how such partisan divergences affect electoral discourse and voter trust. Future studies could also assess the impact of these AI biases on different demographic groups and their election-related behavior.