AI Political Bias Test: Grok and OpenAI Pick Opposing Candidates

What the forced-choice test found across nearly 500 races

The models were asked to choose between candidates in midterm races. The reported percentages describe each model’s candidate selections across the test, not election results, predictions of who will win, or an assessment of which candidates are better qualified. The report does not provide the raw response counts behind the percentages.

The testing organization and the exact date the prompts were run are not identified in the available report. That limits how precisely readers can assess the test’s timing and reproduce its results. The figures should therefore be understood as reported outcomes from this test, rather than a permanent description of how either model answers political questions.

The contrast is notable: the models did not simply lean in the same direction by different degrees. Their answers mostly pointed toward opposite parties. Across nearly 500 races, that low agreement could matter to people who turn to chatbots for quick election information or voting advice. A confident answer may seem authoritative even when another widely used model gives the opposite recommendation.

How model answers changed when candidates were named

The report also describes a smaller, 30-race test in which candidate names were included in the prompt and the models were not allowed to use search. This was a subset test, not another full run across the nearly 500 races, and it changed the prompt conditions.

In that exercise, Grok declined to choose in 27 of the 30 races. GPT-6.1 Sol chose the Democratic candidate in 24 of the 30. The difference suggests that the way a question is framed - including whether candidates are named and whether the model can search - can affect whether it answers and which way its answer goes. The results do not show why the models behaved differently, or whether the same pattern would hold with another prompt.

What the results do and do not say about AI political bias

This test raises a practical concern about AI political bias: users may receive different candidate recommendations from different chatbots, even when asking about the same elections. But a forced-choice exercise measures responses to that specific task. It does not, by itself, establish how a model handles neutral questions, describes candidate positions, or performs across all political topics.

For election information, treat chatbot answers as a starting point, not a voting guide. Check primary sources and candidate records before relying on an AI chatbot, especially when it recommends a candidate without explaining its criteria. Compare claims against official election information and the candidates’ own published positions.

Step-by-Step Guide

1. 1

   Identify the test conditions

   Check which models were tested, how the prompt was worded, whether candidate names were included, and whether search was allowed.

2. 2

   Separate answers from conclusions

   Treat reported selection percentages as outcomes of that particular test, not as election predictions or proof of a model's overall political bias.

3. 3

   Verify candidate information

   Compare chatbot claims with official election information, candidate records, and candidates' own published positions.

4. 4

   Compare sources and explanations

   If a chatbot recommends a candidate, ask what criteria it used and check the explanation against primary sources before relying on it.

Key Statistics

• The test covered nearly 500 upcoming midterm races.TechCrunch reported the test on October 24, 2026. The available report does not provide the raw response counts or identify the testing organization.
• Grok 4.7 selected Republican candidates in 92% of its answers.This is the reported result across the forced-choice test, not a prediction of election outcomes or a general measure of the model's political views.
• OpenAI's GPT-6.1 Sol selected Democratic candidates in 97% of its answers.This is the reported result across the forced-choice test. The report does not provide raw response counts for independently checking the percentage.
• In a 30-race subset, Grok declined to choose in 27 races, while GPT-6.1 Sol chose the Democratic candidate in 24.The subset included candidate names in the prompt and disabled search. It had different conditions from the larger test and is reported by TechCrunch.

Frequently Asked Questions