Grok election recommendations favor Republicans, while OpenAI’s GPT favors Democrats

Grok election recommendations favored Republicans, while OpenAI’s GPT favored Democrats, in a study that forced both models to choose candidates. The reported split was stark: Grok selected Republicans in 92% of answers, and GPT selected Democrats in 97%. Across nearly 500 election races, the models agreed less than 10% of the time.

Those findings describe a specific test, rather than every conversation people have with either chatbot. Researchers used developer interfaces with web search enabled and explicitly demanded a candidate choice. The consumer apps initially declined to provide recommendations, an important boundary around what the results establish.

The pressure now falls on model providers and developers offering AI assistance for public affairs research. A refusal to endorse candidates addresses one visible behavior. It leaves a harder question: whether the system’s selection of issues and evidence already points users toward a preferred conclusion.

What Grok election recommendations actually measured

The study measured candidate selection under pressure to choose, with online search available.

Convovo, a conversational AI research and design studio led by Cornell professor Tapan Parikh, published the experiment through its AI ballot study. Its description identifies the tested models as Grok 4.7 and GPT-6.1 Sol. The scope covered 2026 House, Senate, and governor races.

Researchers accessed the models through an application programming interface, or API, which lets developers connect software directly to a model. That matters because a developer can supply instructions and tools that differ from a consumer application’s setup. Sharing a model family does not guarantee identical behavior across those environments.

The prompt requested a single candidate and discouraged qualification. This was therefore a stress test of commitment: what happens when a system must convert election information into a recommendation? It was not an ordinary request for candidate biographies, voting procedures, or a comparison of policy positions.

That distinction gives the experiment value while limiting its reach. A deliberately demanding prompt can expose tendencies that more cautious answers conceal. However, the frequency of those tendencies under the test cannot automatically describe their frequency during everyday use.

The reported percentages also need the correct denominator. They concern model answers within the experiment, rather than voters, conversations across the entire service, or election outcomes. A recommendation share cannot tell readers how many people received similar advice or whether anyone acted on it.

Likewise, agreement between the models measures consistency across their choices. It does not measure correctness. Political recommendations depend on priorities and judgments, so disagreement requires examining the reasons behind each answer.

A useful comparison would ask whether the systems considered equivalent evidence and disclosed their assumptions. One model might prioritize healthcare, while another emphasizes taxation. Even accurate descriptions can produce opposing recommendations when the systems silently assign different importance to those issues.

The central finding is therefore narrower than a claim that either chatbot possesses a stable political identity. Under the tested conditions, the systems repeatedly selected opposing parties. Explaining that pattern requires separating the prompt, available evidence, model behavior, and surrounding product instructions.

For readers encountering the headline, this boundary is essential. The study raises a concrete concern about automated recommendations. It does not establish that every answer from Grok favors Republicans or that every answer from GPT favors Democrats.

The pressure reaches developers beyond the chatbot apps

Developers need to evaluate the behavior their own applications produce, including how they gather and present evidence.

OpenAI told TechCrunch that the developer API test does not represent the ordinary ChatGPT experience. That consumer app distinction is material to interpreting the results.

A developer building an election research assistant cannot assume that a consumer chatbot’s refusal will carry over automatically. The deployed application has its own instructions, information sources, and output requirements. Those choices determine what users actually encounter.

Consider a hypothetical civic information service that asks a model to compare candidates and produce a concise recommendation. Its designers might intend to reduce the work of reading lengthy platforms. Yet requiring a final choice also delegates the task of deciding which political priorities deserve the most weight.

That is a product decision with editorial consequences. If the user has not supplied their priorities, the system must introduce some basis for ranking candidates. The result can sound personal and authoritative despite resting on assumptions the user never approved.

Removing the recommendation would change that interaction, but it would not resolve every concern. A candidate comparison still needs categories, evidence, and an order of presentation. Each can influence what the reader notices before reaching any explicit conclusion.

For enterprise buyers, the relevant question extends beyond election tools. Public affairs teams, policy analysts, and communications departments use AI to summarize contested information. An answer can shape their understanding through selection and emphasis without containing a direct endorsement.

This is where Grok political bias becomes an evaluation question rather than a branding argument. Buyers need evidence about the tasks they actually perform. A forced-choice election test offers a warning signal, while their own research workflows determine the practical exposure.

Developers also face a communication problem. Calling a tool neutral creates expectations about its behavior across a broad range of questions. Demonstrating neutrality requires more than showing that it refuses the most direct request to choose a candidate.

A more assessable promise would describe what the application does: distinguishes facts from judgments, presents relevant evidence, and identifies uncertainty. Users can inspect those behaviors. They cannot directly inspect an abstract claim that a system has no political perspective.

The immediate burden is therefore shared. Providers need to explain the boundaries of model behavior, and application builders need to verify their implementations. Neither can fully answer for the other’s configuration.

For users, the practical implication is visibility. When an assistant makes a recommendation, its priorities should be identifiable. Otherwise, the system offers a conclusion while leaving the most consequential part of its reasoning outside the conversation.

Search can shape the choice before the answer appears

Access to web search does not, by itself, make a recommendation balanced or well supported.

The study’s search-enabled setup creates a mechanism worth examining: a model helps determine the evidence it later evaluates. It chooses queries, encounters results, and selects material for its response. Each stage can narrow the range of information that reaches the final answer.

This process is often called retrieval, meaning the collection of external material for use in generating an answer. Retrieval can improve access to current information. It also introduces questions about source coverage, query wording, and whether competing claims receive comparable scrutiny.

Imagine a candidate comparison that begins by searching for healthcare proposals. Another begins with immigration enforcement. Both searches can return relevant facts, but they establish different agendas before the model weighs any candidate.

Neither subject is inherently an improper choice. The problem arises when the system presents its chosen agenda as an obvious or comprehensive reflection of the election. A user asking a broad question has not necessarily authorized that prioritization.

Source diversity requires similar care. A response containing several links might still rely on material sharing the same assumptions or repeating the same underlying account. Counting citations cannot establish that the system examined a meaningful range of evidence.

The opposite mistake is to demand equal treatment for every claim. A well-supported statement and a demonstrably false allegation do not deserve equal credibility. Balanced research requires consistent evidentiary standards, rather than mechanically distributing favorable and unfavorable sentences.

Independent election research illustrates why sourcing deserves scrutiny alongside factual accuracy. States United’s voter information study examined ChatGPT and Google AI across testing rounds in late 2025 and early 2026. It found improvements in factual accuracy alongside continuing problems with completeness and referrals to official election sources.

That research addressed different questions from Convovo’s candidate-choice experiment. It nevertheless identifies a related weakness: an answer can avoid discrete factual errors while omitting information the user needs. Accuracy and adequacy are separate properties.

Applied to a Grok vs GPT comparison, that distinction changes what reviewers should inspect. The final party label is only one result. The more revealing material includes which candidates appeared, which issues received attention, and whether cited pages supported the stated reasoning.

There is also a feedback problem in the workflow. A model that begins with a preferred framing might search for supporting evidence and then interpret those results as confirmation. Establishing that mechanism requires controlled testing; a partisan outcome alone cannot prove it occurred.

The same caution applies to commercial explanations. The presence of a publisher in search results does not establish why it appeared. Claims that licensing arrangements, ownership, or political instructions caused an answer require evidence connecting those factors to the observed behavior.

The lesson from Grok election recommendations is therefore about the whole information path. A polished final paragraph can hide earlier choices that deserve inspection. Evaluating only the answer leaves the selection of evidence largely unexamined.

Grok vs GPT does not establish who persuaded voters

A recommendation pattern is evidence about system output, while persuasion requires evidence about people.

The distinction matters because the headline invites a larger conclusion than the experiment can support. A model can repeatedly recommend a party without changing any voter’s preference. Conversely, a subtler answer might influence a reader without ever naming a preferred candidate.

Measuring persuasion would require observing how people respond under defined conditions. Researchers would need to distinguish existing preferences from changes associated with exposure. The forced-choice results do not provide that measurement.

They also do not establish an election effect. Moving from generated answers to changes in turnout or candidate support requires additional evidence. The model’s recommendation share cannot substitute for those missing steps.

The experiment’s prompt creates another limitation. Discouraging qualification tests how a model behaves when pushed toward certainty. Many legitimate research questions require qualification because evidence is incomplete, priorities conflict, or candidates differ across several dimensions.

That does not make the test irrelevant. Systems should be evaluated under demanding conditions, including requests that expose undesirable tendencies. The interpretation simply needs to preserve the difference between a stress test and a representative sample of ordinary use.

A strong evaluation would also examine refusals separately from recommendations. If a system refuses frequently, its answered cases might form a selective subset. Reporting only the party distribution among answers can obscure how often the model declined to choose.

Repeated trials matter for another reason. Generated responses can vary, and the same question can produce different wording or conclusions. A stable pattern across repetitions would support a stronger claim than an isolated answer, although stability would still not establish its cause.

Comparisons across products introduce further complications. Consumer applications can include instructions, interface features, and tool configurations beyond the underlying model. Describing the results of one configuration as the character of an entire company skips those distinctions.

Independent work also shows why political evaluation should cover multiple behaviors. The Brennan Center’s election disinformation research found that tested chatbots challenged election conspiracy theories, while still making factual or citation errors. Its findings also identified weaknesses in safeguards against creating deceptive election media.

Those results cannot validate or invalidate Convovo’s party preferences. They concern different tasks. Their relevance is that a system can perform relatively well on one election-related behavior and poorly on another.

For the same reason, a Grok political bias assessment should avoid treating a single score as a complete verdict. Candidate selection, factual reliability, source quality, and resistance to misinformation need separate examination. Combining them into one political label makes the evidence harder to interpret.

The most defensible reading remains specific. The reported experiment reveals sharply divergent recommendations under a defined setup. It creates a case for further investigation without proving intentional manipulation, universal partisan behavior, or measurable influence on the electorate.

Neutrality needs evidence that readers can inspect

The meaningful comparison is between a neutrality claim and the observable decisions a system makes.

That is the main tension behind this story. Refusing to endorse candidates is visible and easy to demonstrate. Showing that broader political answers apply consistent standards is a more demanding task.

A system can decline to say whom a person should support while still determining which issues lead the discussion. It can summarize candidates accurately while leaving out a relevant position. It can provide citations that appear reassuring without making their limitations clear.

These possibilities are analytical concerns, rather than findings about every answer from either provider. They explain why a product-level claim requires task-level evidence. The scope of the evidence should match the scope of the promise.

Brookings’ political bias analysis examined chatbot responses using political quizzes and discussed the limitations of assessing political orientation. Its approach differs from a ballot recommendation experiment, reinforcing the importance of stating what each evaluation actually measures.

A political quiz asks a system to choose among abstract positions. A candidate recommendation asks it to weigh people, records, and current information. Agreement between those methods would be informative, but neither automatically explains the other.

Transparency should therefore begin with the evaluation design. Readers need to know the tested model version, the prompt, available tools, and how responses were classified. Without that information, an impressive percentage is difficult to assess or reproduce.

Interpretation needs equivalent transparency. Researchers should distinguish direct observations from proposed explanations. A partisan distribution is an observation; a claim about which internal process caused it is a separate inference.

The public also benefits when counterexamples remain visible. Answers that depart from a model’s dominant pattern can help identify relevant conditions. Hiding those cases behind an aggregate percentage makes a simple headline easier and the mechanism harder to understand.

For developers, this suggests a broader form of accountability. A research assistant should make the source of a judgment legible: user priorities, documented criteria, or a clearly identified interpretive choice. An unexplained ranking should not inherit the authority of the facts beneath it.

Knowledge workers can apply the same principle when preserving AI research. Keeping original evidence beside generated summaries makes later review more practical. A searchable knowledge base can support that workflow, although storage alone cannot establish the quality of a conclusion.

The relevant standard is traceability, meaning the ability to follow an answer back to supporting material. Traceability does not guarantee neutrality, but it gives reviewers something concrete to examine. It also makes omissions and unsupported leaps easier to challenge.

Grok election recommendations put that standard under pressure because the opposing outcomes are so easy to recognize. The harder work begins after identifying the party split: determining which decisions produced it and whether those decisions remain consistent across less explicit questions.

What should change in the next round of testing

The next useful evidence should explain the pattern, test its boundaries, and establish whether it reaches ordinary users.

The first signal to watch is independent replication using documented model versions and comparable settings. Reproducing the party split would strengthen confidence that it represents a repeatable behavior. A substantially different result would require examining what changed before declaring the original finding wrong.

Replication should preserve the details that make comparisons meaningful. Search availability, question wording, and refusal handling can alter what the experiment measures. Publishing those conditions allows readers to distinguish a genuine contradiction from a test of a different system.

The second signal is controlled evidence about how recommendations depend on available information. Tests could compare open search with a shared source packet and examine whether the same candidate evidence produces similar judgments. These are proposed experiments, rather than announced follow-ups.

If opposing preferences persisted with identical evidence and comparable instructions, source selection alone would become a less sufficient explanation. If the gap narrowed, the role of retrieval would deserve greater attention. Neither outcome would eliminate the need to inspect how the models weighed competing priorities.

The third signal is testing of ordinary consumer conversations that stop short of requesting an endorsement. Questions about candidate records, issue importance, and policy tradeoffs would better address the concern that framing matters before a final recommendation appears.

If those conversations consistently reproduced the directional pattern, the argument for broader product relevance would strengthen. If they did not, the forced-choice result would remain significant within a narrower scope. Either finding would help users understand where the demonstrated risk begins and ends.

For the coming months, company responses will be useful when they address these measurable questions. A general assertion of neutrality cannot explain a specific pattern. Equally, an accusation of political engineering cannot replace evidence about how the answers were produced.

Developers should pay particular attention to changes in the actual versions and configurations they deploy. An improvement observed in a consumer application does not automatically establish an improvement in their service. Evaluation needs to follow the product that users receive.

Readers should also resist interpreting the Grok vs GPT split as a recommendation to choose whichever assistant matches their politics. Agreement with an existing preference does not establish better reasoning. A trustworthy research process should withstand scrutiny when its conclusion is unwelcome.

When reviewing the next claim about Grok election recommendations, ask for the conditions behind the percentage. Look for the prompt, model version, source access, refusal rate, and evidence supporting the proposed explanation. Compare the final choice with the information the system considered and the priorities it introduced. Preserve original sources when an answer matters to your work, and separate verified facts from the assistant’s judgments. Before treating either model as a political research authority, ask a more useful question: can you inspect why it reached that conclusion, and can someone with different priorities evaluate the same evidence?