Grok vs ChatGPT: AI Models Forced to Pick Midterm Candidates Split 92% GOP vs 97% Dem

When researchers at Convovo forced two AI models to name one candidate in nearly 500 US midterm races, Grok 4.7 picked the Republican 92% of the time and OpenAI's GPT-6.1 Sol picked the Democrat 97% of the time, agreeing in under 10% of races. The test used developer APIs with a forcing prompt; consumer apps refused. The bigger lesson: models frame the same subject differently and cite different sources.

What did the Convovo study actually find?

The study found that two leading models gave nearly opposite answers to the same voting question. According to TechCrunch's report, Convovo, a conversational AI research and design studio led by Cornell University professor Tapan Parikh, queried both models through their APIs with search turned on. The prompt asked who to vote for, told the model to name one candidate, and told it not to hedge or list both sides.

The headline results:

• Grok 4.7 recommended the Republican in 92% of its answers.
• GPT-6.1 Sol recommended the Democrat in 97% of its answers.
• The two models agreed on a candidate in fewer than 10% of races.
• In heavily Republican districts, GPT still picked the Democrat in 57 of 67 races.
• In deep-blue districts, Grok still picked the Republican in 47 of 72 races.

TechCrunch reported one exception: Grok was more likely to suggest the Democrat when the Republican faced a serious controversy. Convovo's AI Ballot 2026 page hosts the write-up and a data explorer.

Which sources did each model rely on?

Each model leaned on a noticeably different set of sources. TechCrunch reported that Grok drew on nonpartisan sites such as Ballotpedia, OpenSecrets and VoteSmart, alongside conservative sources. Specifically:

• X appeared in 41% of Grok's answers.
• A Christian conservative voter guide appeared in 39%.
• Turning Point Action's website appeared in 23%.

GPT, by contrast, relied heavily on Wikipedia, the Associated Press, Axios and Reddit. TechCrunch noted that OpenAI has content deals with AP, Axios and Reddit.

The search behavior differed too. GPT searched more often for candidates' healthcare and abortion positions and searched for Democrats more often, sometimes skipping Republican opponents. Grok searched topics such as immigration, taxes, crime and energy and was more likely to look up both candidates. Parikh told TechCrunch he suspects both models searched for reasons to justify a lean they already had.

How much weight should readers give these results?

Readers should treat the results as a stress test, not a picture of everyday use. Several caveats matter:

1. The prompt forced a choice. The study explicitly told models not to hedge, which is not how most people ask questions.
2. Consumer apps refused. ChatGPT and Grok's consumer apps declined to name a candidate, although Parikh said simple prompts could get around that.
3. OpenAI disputes the framing. An OpenAI spokesperson said ChatGPT is designed to be objective by default and that an API test demanding a recommendation does not reflect typical use. Grok's developer did not respond before publication, per TechCrunch.
4. The follow-up was small. A no-search test of just 30 races found GPT picked the Democrat in 24 and declined in five, while Grok declined in 27. IBTimes noted the selection of those 30 races was not explained.
5. The "hidden tokens" claim is unconfirmed. Researchers said each Grok API request carried roughly 1,240 extra input tokens they did not write and could not see. Their origin and content have not been independently confirmed.

Usage is also limited. A Pew Research Center survey published October 23 found only 8% of US adults use AI chatbots for 2026 election information, though that rises to 11% among adults under 50.

Is this the first evidence of political lean in AI models?

No, earlier research has documented political lean, though with different methods and mixed conclusions. David Rozado's 2024 study in PLOS ONE tested 24 models on 11 political orientation tests and found most conversational models scored left of center, as TechXplore summarized. Rozado also showed fine-tuning could push a model in either direction.

A 2025 Stanford and Dartmouth study asked more than 10,000 Americans to rate model responses. Respondents perceived OpenAI's models as the most left-leaning, with xAI second, while Google's models were seen as least slanted. The authors cautioned that perceived neutrality is not the same as accuracy. The Convovo results complicate that earlier picture, since Grok's search-enabled answers ran in the opposite direction.

Why does this matter beyond elections?

It matters because the same pattern applies to any topic, including brands. The study shows that two models given an identical question can retrieve different sources, emphasize different issues and reach different conclusions.

For a company, the "candidate" could be a product, a category or a reputation question. One model might lean on community forums and news wires, while another leans on social posts and advocacy sites. If you only check one assistant, you see only one version of the story.

What should brands do about multi-model differences?

Brands should monitor how several AI models describe them, not just one. The Convovo findings are a vivid example of model divergence, and similar divergence can show up in category, product and reputation questions. A practical approach:

1. Track several models side by side. Run the same prompts across ChatGPT, Grok, Gemini and others. MarkGrid's Model Share measures how often and how favorably each model mentions your brand compared with competitors.
2. Audit the citations, not just the answer. Note which domains each model pulls from. If one model relies on forums and another on social posts, your content strategy needs both.
3. Test indirect questions. Ask "which tool is best for X" rather than only "what is Brand Y." Parikh told TechCrunch that bias in indirect questions is the bigger concern.
4. Strengthen neutral, citable sources. Reference sites, accurate directories and clear first-party pages tend to appear across models. The MarkGrid GEO guide covers how to make content easier for AI systems to cite accurately.
5. Watch for manipulation and misinformation. Election season brings synthetic media risks too; see our guide on spotting Russian deepfake celebrity videos ahead of the midterms. Legal scrutiny of AI answers is also growing, as covered in our post on Texas suing OpenAI.

FAQ

Did ChatGPT and Grok recommend candidates to regular users?

No. According to TechCrunch, the consumer ChatGPT and Grok apps declined to name a candidate. The 92% and 97% figures came from API tests with a prompt that demanded a single choice.

How many races did Convovo test?

Nearly 500 US midterm races, covering House, Senate and governor contests, according to TechCrunch and IBTimes. Convovo's page says it asked about every such race.

What sources did Grok cite most?

TechCrunch reported X appeared in 41% of Grok's answers, a Christian conservative voter guide in 39% and Turning Point Action's site in 23%. Grok also cited nonpartisan sites such as Ballotpedia.

What did OpenAI say about the study?

An OpenAI spokesperson said ChatGPT is designed to be objective by default and that an API test forcing a recommendation does not reflect how people typically use ChatGPT.

Are the hidden tokens in Grok's API confirmed?

No. Researchers reported about 1,240 extra input tokens on each Grok request that they could not see or remove. Their source and content have not been independently confirmed.

Where should voters get election information?

TechCrunch pointed readers to resources such as Vote.gov, vote411.org, Ballotpedia, VoteSmart, state and county election offices, candidate websites and local news.

Last updated October 26, 2026.