THE SHORT ANSWER
● Grok 4.7 — Recommended the Republican in 92% of forced answers
● GPT-6.1 Sol — Recommended the Democrat in 97% of forced answers
● Agreement — Fewer than 10% of nearly 500 races
● Who is right — Neither - there is no correct answer. That is the point
Convovo, a conversational AI research studio led by Cornell professor Tapan Parikh, took a question both models normally refuse and removed the refusal. Across nearly 500 races in the 2026 US midterms, accessed by API, each model had to name a candidate.
| Grok 4.7 | GPT-6.1 Sol |
|---|
| Recommended Republican | 92% | 3% |
| Recommended Democrat | 8% | 97% |
| Races tested | Nearly 500, 2026 US midterms |
| Agreement between them | Under 10% |
| Default behaviour | Both normally decline the question |
| Vendor disclosure of lean | None | None |
Why the forced choice matters
A model that declines to recommend a candidate looks neutral. The study tests whether the refusal is covering a preference or an absence of one, and on this evidence it is covering a preference - a strong one, in opposite directions, from two systems a user would reasonably treat as substitutes.
The 92/97 split with under 10% agreement is close to the maximum separation two systems could produce on identical inputs. If both were noisy you would expect agreement near 50%. If both were neutral you would expect refusals or coin-flips. Neither happened.
What this does and does not establish
- It does not show either model is inaccurate. Candidate recommendation has no ground truth
- It does not show the lean appears in normal use, where both models decline
- It does show the lean exists underneath the refusal and is vendor-specific
- It does mean a product that routes political questions to one model has adopted that model's lean
The practical read
If you are building anything that touches elections, policy or civic information, two things follow. Model choice is editorial, not technical - swapping Grok for GPT changes the output's political direction more reliably than it changes the output's quality. And neither vendor publishes this, so you cannot cite a datasheet; you have to test your own prompts across both and look at the spread yourself.
For everything else, this is a reminder that "which model" is rarely a neutral default. The harness study published this week made the same argument about cost; this one makes it about content.