Convovo Study Finds Grok and GPT Pull Election Picks in Opposite Directions

A Convovo study found that Grok 4.7 and OpenAI-derived GPT-6.1 Sol leaned in opposite partisan directions when forced to choose between candidates in nearly 500 upcoming midterm races. The result matters because it suggests election chatbots may not merely answer differently — they may systematically steer users toward different political outcomes depending on how they are prompted and what sources they use.

What the study found

TechCrunch reported that researchers tested Grok 4.7 and GPT-6.1 Sol on candidate-choice prompts tied to upcoming midterm contests. Under that setup, Grok picked the Republican candidate in 92% of answers, while GPT-6.1 Sol chose the Democratic candidate in 97% of answers.

The models also rarely converged. Across nearly 500 races, the two systems agreed on the same outcome in less than 10% of cases. That divergence is striking because the prompt was simple: choose one candidate or the other.

How were the models tested?

The researchers used API access rather than consumer chat apps for the main study. That distinction matters because the consumer versions refused the direct request to vote for one candidate, while the API layer allowed the test to proceed. OpenAI told TechCrunch that the API test did not match how most users experience ChatGPT, and that ChatGPT is designed to be objective by default in political conversations.

Convovo lead Tapan Parikh said the pattern is not limited to blunt endorsement prompts. He argued that bias can surface in broader election discussions too, including when users ask for background, context or comparisons rather than a direct recommendation.

“It has an implicit set of values in terms of what it cares about and how it leans”

Why it matters

The findings raise a practical question for anyone using AI to make sense of elections: if the model seems neutral on the surface, what is shaping the answer underneath? For voters, that can affect how candidate information is filtered, which issues get emphasis and which sources are treated as authoritative.

For the AI industry, the study adds pressure to a longstanding claim that chatbots can be objective by default. The issue is not simply whether a model refuses to endorse a candidate in a consumer interface. It is whether it can still produce consistently slanted results when users rephrase their question, or when a developer accesses the model through an API.

For the political ecosystem, the stakes are higher because election guidance is already a contested information space. TechCrunch cited Pew survey data showing that 8% of Americans use chatbots for election information. That is not a majority, but it is enough to matter in a close race, especially if one system systematically nudges users toward one party.

By the numbers

What the models were looking at

The study suggests the bias was not random. The two systems appeared to lean on different information diets, and that appears to have shaped the recommendations they produced.

Grok’s sourcing pattern

TechCrunch reported that Grok often cited conservative sources, along with Ballotpedia, OpenSecrets, VoteSmart and X. In the study, Grok also more often searched for both candidates, which suggests a broader information sweep before answering.

Its topic emphasis also skewed in a particular direction. Grok more often searched immigration, taxes, crime and energy positions, all of which are issues that can map onto partisan identity in U.S. politics.

GPT’s sourcing pattern

GPT leaned on Wikipedia, the Associated Press, Axios and Reddit in its sourcing. It also searched healthcare and abortion positions more often than Grok, a difference that may help explain why it repeatedly favored Democratic candidates in the test.

TechCrunch reported that GPT more often searched for Democratic candidates and sometimes skipped Republican opponents altogether. That is a notable asymmetry: a model that sees one side more often may not be comparing the field on equal terms.

How did the geography of the races affect the results?

The pattern held even in places where it ran against the local partisan grain. In heavily Republican districts, GPT picked the Democrat in 57 of 67 districts. In deep-blue districts, Grok picked the Republican in 47 of 72 districts.

Those results suggest the bias was not simply a matter of following the district’s likely winner. Instead, the models often pushed toward the opposite side of the dominant local preference, which makes the output even more consequential for a user seeking neutral guidance.

What happens when search is turned off?

Turning off internet search changed the behavior, but not in the same way for both models. With search disabled, GPT still leaned Democratic. Grok, by contrast, largely stopped choosing a candidate.

In the smaller test set cited by TechCrunch, GPT picked a Democrat in 24 of 30 cases with search off, chose a Republican once and declined to pick five times. Grok refused to pick in 27 of 30 cases and chose a candidate only three times.

That split is important because it points to different mechanisms. GPT appears to carry a durable preference even without web search. Grok looks far more dependent on its search pattern and source selection, which may be where the final answer gets shaped.

“For Grok, it really is the search pattern — the terms and the sources — that determine the result.”

Background and context

Large language models have been reluctant political actors from the start. Consumer chat interfaces often refuse direct endorsements, especially on sensitive topics like elections. At the same time, prior research has warned that AI systems can still steer users through framing, selective sourcing and the way they organize information.

That tension is at the center of this study. A chatbot can reject a direct “who should I vote for?” prompt and still shape an answer through the sources it surfaces, the issues it prioritizes and the candidates it compares. The result can feel neutral to the user while still embedding a directional push.

The API-versus-consumer-app distinction also matters. Developers can add prompts, attach custom instructions and wire models into products that behave differently from the public interface. That means findings from one access method do not automatically describe what every user sees, but they do reveal how the system can behave when the guardrails are adjusted.

OpenAI pushed back on that interpretation. The company said the API test did not reflect the way most people use ChatGPT and repeated that the system is meant to be objective in political conversations. That response does not settle the question; it simply highlights the gap between intended behavior and observed behavior under experimental conditions.

What the sourcing suggests about bias

The study’s most interesting detail may be less the final recommendation than the pathway to get there. Grok’s reliance on conservative sources, X and several political databases suggests a sourcing environment that may amplify a right-leaning frame. GPT’s use of AP, Axios, Wikipedia and Reddit points to a different mix, one that may privilege mainstream summaries and user discussion over partisan sites.

Those sourcing differences matter because retrieval is not a neutral add-on. When a model searches, chooses, ranks and summarizes, each step can tilt the result. That helps explain why hidden bias can emerge even if the base model is not explicitly instructed to support one party.

Parikh said consumer-app jailbreaks using simple prompts produced recommendations similar to prior research. In other words, the more the model is pushed out of its default refusal mode, the more visible its directional preferences become.

How much of this is training, and how much is search?

That question remains open, and the study does not settle it. The evidence points to both mechanisms. GPT’s persistent Democratic lean even with search off suggests the bias may be partly embedded in the model itself. Grok’s dependence on search and sources suggests a heavier role for retrieval and web-ranking choices.

Parikh put the distinction plainly: “For GPT, it has mostly decided already…it’s in the weights.” That is a strong claim about where the tendency lives, though the study as described by TechCrunch does not prove it alone.

What the paper does show is that a supposedly simple election query can produce radically different partisan answers depending on the model and the access route. That should make users cautious about treating AI election guidance as a single, stable product.

Who is behind the broader political context?

Elon Musk looms over the Grok side of the story, not because the study proves direct influence, but because his political spending and ownership of xAI keep the model in a highly scrutinized orbit. TechCrunch cited $445 million in Musk donations supporting GOP candidates in the midterms, and roughly a quarter of a billion dollars spent backing Trump and Republicans in 2024.

Sam Altman, by contrast, has said he was not putting money into the 2026 elections, according to TechCrunch. Those public positions do not explain the model behavior by themselves, but they frame why partisan-lean questions around these products attract immediate attention.

SpaceXAI did not provide comment by deadline, TechCrunch reported.

What happens next

The immediate question is whether this pattern holds beyond these two systems and beyond this exact prompt design. The study raises several open issues that will likely matter in the coming months:

• Whether other models show the same kind of partisan split on election prompts.
• Whether hidden API instructions, including the extra 1,240 tokens noted in the study, materially shaped Grok’s answers.
• Whether consumer-app jailbreaks produce the same results at scale.
• How much of the lean comes from model training versus search and source selection.
• Whether election-season usage changes as voters turn to chatbots for quick candidate summaries.

The midterm election date referenced in the story is November 3, with less than two weeks remaining at the time of publication noted by TechCrunch. That timing matters because these systems are being tested right before voters make decisions, not in an abstract lab setting long before an election.

If the findings hold, they could reshape how platforms label political answers, how developers configure retrieval, and how users interpret what looks like an objective summary. If they do not hold under broader testing, the study will still have shown something valuable: that apparent neutrality in AI can disappear quickly once search, sources and hidden instructions enter the picture.

Bottom line

Convovo’s study suggests that election chatbots may be far less neutral than they appear, but in different ways. Grok 4.7 leaned Republican, GPT-6.1 Sol leaned Democratic, and the method used to reach each answer helped determine the result. For anyone relying on AI to understand the ballot, that is not a small technical quirk; it is a warning about how political information is being filtered before users ever see it.