Convovo, a conversational AI research studio led by Cornell University professor Tapan Parikh, tested what happens when you force a frontier model to pick a side. Across nearly 500 races in the 2026 US midterms, accessed through their APIs, Grok 4.7 recommended the Republican candidate in 92% of its answers and GPT-6.1 Sol recommended the Democrat in 97%. The two models agreed in fewer than 10% of races.
The design is what makes the result readable. Both models will normally decline to recommend a candidate; the study removed that option. What is left is whatever lean sits underneath the refusal, and on this evidence it is not subtle and it is not the same across vendors. A 92/97 split with sub-10% agreement is close to the maximum separation two systems could produce on identical inputs.
Neither figure tells you a model is wrong, because there is no correct answer to "who should I vote for". What it does tell you is that model choice is a political choice in any product that lets an assistant weigh candidates, ballot measures or policy positions - and that two tools a user would consider interchangeable are not. For anyone shipping on these APIs, the finding to act on is that defaults carry a lean the vendor has not disclosed.
Sources
TechCrunch · The AI Insider