AI political bias is facing renewed scrutiny after a study found that xAI’s Grok and OpenAI’s model favoured opposing parties when researchers forced them to choose candidates in the 2026 US midterm elections. The findings suggest that AI systems can produce sharply different recommendations even when given similar questions and access to election information.

The study was conducted by Convovo, a conversational AI research and design studio led by Cornell University professor Tapan Parikh. According to published reporting on the research, Grok 4.7 recommended Republican candidates in 92% of its answers, while OpenAI’s GPT-6.1 Sol selected Democrats in 97%. The researchers examined nearly 500 races, including contests for the US House, Senate and governor positions. The models agreed in fewer than 10% of the races.

These figures do not mean that every interaction with Grok or ChatGPT will produce a partisan recommendation. The test used developer APIs and an instruction that explicitly demanded a single choice. The consumer chatbot applications reportedly declined to make direct candidate recommendations when asked the same question. That distinction matters when interpreting what the study does—and does not—show.

The results nevertheless raise an important question for the AI industry: when a model is asked to make a political judgement, how much of its answer comes from its training, its search behaviour, its instructions or the sources it consults?

1. What the study found about Grok and OpenAI

The study asked each model a deliberately direct question: who should the user vote for in a particular race? Researchers instructed the models to name one candidate and not hedge, list both sides or avoid making a decision.

Under those conditions, the reported results were striking. Grok 4.7 chose Republican candidates in 92% of its answers. OpenAI’s GPT-6.1 Sol chose Democratic candidates in 97%. Across almost 500 contests, the two systems agreed less than one time in ten.

The prompt design is central to understanding these numbers. Rather than asking the systems to neutrally explain candidate platforms, researchers pressed them to make a recommendation. This exposed differences that might be less visible in ordinary questions about policy, voting procedures or candidate biographies.

Still, the percentages should not be treated as a universal ranking of political neutrality. They describe responses under a particular test setup, with particular model versions and prompting conditions. A different prompt, model release, search configuration or set of races could produce different results.

The study is best understood as evidence that model behaviour deserves systematic evaluation—not as proof that every answer from either company is politically motivated.

2. The models appeared to rely on different sources

The researchers also examined the information used to support the models’ recommendations. Published accounts of the study reported that Grok cited a mixture of sources that included a Christian conservative voter guide, Turning Point Action and posts on X. It also drew on nonpartisan election resources.

OpenAI’s model, meanwhile, reportedly relied on sources including Wikipedia, the Associated Press and Reddit. The researchers also observed differences in the topics each system searched for while assessing candidates. OpenAI’s model searched more often on healthcare and abortion, while Grok more frequently focused on immigration, taxes, crime and energy.

Those patterns could affect which candidate looks more appealing. A model that gives more weight to a particular policy area—or finds information about one candidate more readily than information about an opponent—may produce a different recommendation.

However, citing a source does not automatically prove that the source caused a particular answer. Establishing causation requires further controlled testing, including careful analysis of the search queries, retrieved documents and how the model weighs conflicting evidence.

For AI political bias research, the key issue is not simply whether a system cites left-leaning or right-leaning material. Researchers must also examine whether it searches fairly, applies consistent standards to competing candidates and accurately represents the evidence it finds.

3. A follow-up test suggested search behaviour mattered

A smaller follow-up test offered a clue about why the systems differed. When researchers disabled web search and tested 30 races, OpenAI’s model still selected Democratic candidates in 24 cases. Grok declined to choose in 27 of the 30 races; its three selections were all Democrats, according to the reported findings.

The small sample limits how broadly these results can be generalised. Even so, the contrast suggests that the two systems may have reached their original answers through different mechanisms.

The persistence of OpenAI’s pattern without web search is consistent with the possibility that the model’s learned parameters and internal preferences influenced its choices. Grok’s reluctance to choose without search suggests that its original results may have depended more heavily on the sources it retrieved or on how its search process framed the races.

These are interpretations, not definitive explanations. The study does not establish that training alone caused one model’s behaviour or that web search alone caused the other’s. More experiments would be needed to separate the effects of model training, hidden instructions, search tools and prompt design.

The follow-up does, however, illustrate why AI political bias cannot be assessed simply by looking at a chatbot’s final answer. The process used to reach that answer can be just as important.

4. Why API testing differs from using a chatbot

One of the most important qualifications is that the research tested developer APIs, rather than relying solely on the standard versions of the consumer chatbot apps.

An API allows developers to integrate an AI model into software and specify prompts, tools and workflows. Depending on the service, an API-based system can be configured differently from a public-facing chatbot, which may include additional instructions, safeguards and product-level restrictions.

In this case, researchers used a prompt that explicitly demanded a candidate recommendation. The consumer apps reportedly declined to answer that direct request. OpenAI argued that its public chatbot is designed to remain objective in political conversations and that the test did not reflect how most users interact with the product.

That distinction does not make the API findings irrelevant. Developers build AI-powered search tools, assistants, research products and decision-support systems using APIs. If a model produces systematic differences under certain conditions, those differences may matter for applications built on top of it.

But it would be misleading to say the test proves that ordinary ChatGPT users will receive a Democratic recommendation or that every Grok user will receive a Republican one. The reported findings concern a particular evaluation method, not every use of either product.

5. Hidden instructions add another layer of uncertainty

The researchers identified another limitation in their analysis: each Grok API request reportedly contained approximately 1,240 tokens of additional input that the team had not written and could not inspect or remove.

This is an important uncertainty. If an AI request includes instructions or context that researchers cannot see, it becomes harder to determine which factors produced a response. The extra input could be relevant to the system’s behaviour, but the reported test does not establish exactly what it contained or how much it affected the results.

A rigorous evaluation of AI political bias should make the model version, prompt, tools, search configuration and available context as transparent as possible. Researchers should also rerun tests across multiple prompts and datasets to determine whether an observed pattern is stable.

The issue extends beyond elections. Businesses evaluating AI for customer support, hiring, research or decision-making also need to understand which instructions and external sources shape the system’s output. Without sufficient transparency, it can be difficult to distinguish a repeatable model tendency from an artefact of the testing setup.

6. What the findings mean for voters and AI developers

The findings are particularly relevant as US voters prepare for the November 3, 2026 midterm elections. AI tools can summarise policies, explain political terminology and help people locate public information. But a confident recommendation should not be mistaken for an objective assessment.

Voters can reduce the risk of being misled by asking AI systems to summarise each candidate’s stated positions separately, request sources for factual claims and compare those sources with official campaign materials and reputable independent reporting. They should also look for omitted context, outdated information and differences in how candidates are described.

For developers, the lesson is to evaluate systems before deploying them in high-impact contexts. Testing should include prompts that vary in wording, balanced sets of candidates, multiple model versions and clear criteria for measuring fairness. Teams should document whether the system uses web search, what sources it retrieves and when it refuses to answer.

Companies should also be careful about building automated recommendations around a model’s unverified political or social judgements. Human review, transparent sourcing and a way to challenge or correct outputs are valuable safeguards.

For more technology and AI coverage, visit Progressive Robot.

Conclusion: AI political bias needs transparent testing

The Convovo study suggests that two leading AI models can make dramatically different candidate recommendations when directly instructed to choose. Grok 4.7 reportedly favoured Republican candidates in 92% of its answers, while OpenAI’s GPT-6.1 Sol selected Democrats in 97%, with agreement between the systems below 10% across nearly 500 races.

Those results warrant attention, but they should be interpreted within the study’s limits. API prompts differ from ordinary chatbot use, the follow-up test was relatively small, and some details about Grok’s request context remained unavailable to the researchers.

The broader lesson is that AI neutrality should not be assumed—or rejected—based on a headline alone. It should be evaluated through repeatable tests, transparent methods and careful scrutiny of how models select and present information.

As AI becomes more involved in the way people research public issues, understanding AI political bias will be essential to building systems that are useful, accountable and worthy of users’ trust.

Sources

Convovo — Original AI Election Study
Search for the original research by Convovo, led by Cornell University professor Tapan Parikh, examining Grok and OpenAI model recommendations across US midterm races. Include the original study or data explorer link here once verified.
Published Reporting on the Grok vs OpenAI Study
The reporting cited for the study’s findings covers the reported Republican and Democratic candidate-selection rates, differences in sources consulted, and the smaller follow-up test without web search. Verify the original article URLs before publishing.
OpenAI — Official Website
https://openai.com/
xAI — Official Website
https://x.ai/
Progressive Robot — AI and Technology News
https://www.progressiverobot.com/