GPT-6.1 Sol and Grok 4.7 pick rival parties in study

By VK

Published 26 Oct 2026, 12:21 UTC

Reporting from TechCrunch

Told via API to name one candidate in 479 midterm races, GPT-6.1 Sol mostly picked Democrats and Grok 4.7 Republicans. Their apps refused in spot checks.

What it means for founders

API access may not carry the app's guardrails. If your product calls GPT-6.1 Sol or Grok 4.7 directly, a user may get a political recommendation that ChatGPT itself would refuse. Write your own system prompt for election questions and test it before November 3.
Hosted search is part of the answer. The study suggests the queries a vendor's search tool runs and the sites it cites can shape the output, at least for Grok. If you enable it, log those queries and citations so you can audit them.
Ask what rides along on each call. The study counted about 1,240 tokens in every Grok request that the caller never sent. Ask vendors what they inject, and whether it touches context limits or billing.
Watch for vendor responses. Signals to track are any explanation from SpaceXAI of the hidden input, and whether OpenAI changes how its API handles direct voting questions before polls close.

The story

Asked through their developer APIs to name one candidate for each House, Senate and governor contest on the 2026 ballot, OpenAI's GPT-6.1 Sol and SpaceXAI's Grok 4.7 leaned toward opposite parties. The study comes from Convovo, an AI studio run by Cornell professor Tapan Parikh, and was reported by TechCrunch on Saturday, October 24, ten days before the November 3 midterms. Its code and every answer are in a public repository.

How the test worked

The analysis covers the 479 races that pit a single Democratic nominee against a single Republican. Each model got the same instruction three times per race: choose one candidate and do not hedge. Both used their vendor's own web search, with no added system prompt or user history.

Of the 1,111 GPT answers that named a major-party candidate, 1,082 named the Democrat, about 97%. For Grok, 952 of 1,038 such answers went to the Republican, about 92%. Both also declined often, GPT in 313 answers and Grok in 375. Gemini 3.8 Flash and Claude Sonnet 5.5, run on ten Senate races, declined all 20 calls.

Turning search off split the two. Given only the nominees' names, GPT still chose the Democrat in 387 races and the Republican in 37, while Grok declined in 441. Parikh told TechCrunch that GPT "has mostly decided already", while Grok's answers track the search terms and sources it chooses. Per the study, the xAI API also attaches roughly 1,240 tokens of hidden input to each Grok request.

Apps behave differently from APIs

In spot checks, the consumer ChatGPT and Grok apps refused the plain prompt. An OpenAI spokesperson told TechCrunch that ChatGPT aims to be objective on politics and that a forced API pick says little about how people use it. SpaceXAI did not respond to TechCrunch in time. Small tests that cast the apps as an editorial board or debate judge did draw verdicts, leaning the same way as the API.

A Pew survey of 5,273 adults, taken October 12 to 18, found 8% saying they use chatbots for election information this year. Among them, 76% looked up candidates' issue positions and 40% sought suggestions on how to vote.

What we don't know yet

What the extra Grok tokens contain, and whether they shape answers. SpaceXAI did not comment, and the study does not say.
Whether the lean appears in looser questions, such as which issues matter most, as Parikh expects. The ballot runs used a direct demand for a pick.
How stable the results are. The main search-on draws all ran on a single day, October 21.