# Forced to pick, Grok leans Republican while GPT-6.1 Sol leans Democrat

- URL: https://usa-24.com/grok-political-bias-study/
- Publisher: USA 24
- Section: Artificial Intelligence (AI)
- Published: 2026-10-26T06:02:02.845Z
- Updated: 2026-10-26T06:02:02.855Z
- Topics: AI Bias, Grok, ChatGPT, Midterm Elections 2026, AI and Elections

> A Cornell-led study forced two chatbots to name one candidate in nearly 500 US midterm races, and they mostly backed opposite parties. Voters should weigh any single-name pick carefully.

## Key takeaways

- Grok 4.7 recommended the Republican candidate in 92% of its answers, while GPT-6.1 Sol picked the Democrat in 97%, across nearly 500 US midterm races.
- The two models disagreed in more than 90% of races when forced to name a single candidate.
- Cornell professor Tapan Parikh's team concluded that GPT's lean is embedded in its weights, while Grok's depends on its search terms and sources.
- Each Grok API call carried about 1,240 extra input tokens that researchers could not view or remove, and their source remains unexplained.
- In a 30-race follow-up with web search turned off, Grok declined in 27 races and GPT picked Democrats in 24.

When researchers forced two leading chatbots to pick a candidate, they mostly picked opposite parties. The Grok political bias study, led by a Cornell professor, tested xAI's Grok 4.7 and OpenAI's GPT-6.1 Sol across nearly 500 US midterm races. TechCrunch published the findings on October 24, and the results have since spread through AI news aggregators. Election day is November 3, so the timing puts the findings in front of voters who may already be asking chatbots for help.

The core claim is simple. When the models could not hedge, Grok mostly backed Republicans and GPT-6.1 Sol mostly backed Democrats. The reasons behind those leans appear to differ, and that distinction is where the study gets most interesting.

## How the forced-choice test worked

The study came from Convovo, a conversational-AI research studio, and was led by Cornell professor Tapan Parikh, according to [The AI Insider](https://theaiinsider.tech/2026/10/25/study-finds-grok-and-gpt-models-show-opposing-partisan-leans-on-u-s-midterm-races/). The researchers asked each model to name a single candidate in nearly 500 US midterm races, with no room to hedge.

The consumer apps would not play along. [Gadget Review](https://www.gadgetreview.com/grok-picks-republicans-openai-picks-democrats-when-ai-models-are-forced-to-choose) reported that the apps declined to name a candidate, so the team used the models' APIs, the programming interfaces developers use to build products on top of them. That choice matters for voters. An API result is not the same thing as the answer you see in a chat window, and the coverage did not establish how often ordinary users would receive the same picks.

The coverage reviewed for this story did not include a peer-reviewed paper or the full methodology document. That leaves the headline numbers resting on one research group's account of its own work.

## What the Grok political bias study found in midterm races

Grok 4.7 recommended the Republican candidate in 92% of its answers, according to [TechCrunch](https://techcrunch.com/2026/10/24/forced-to-pick-groks-ai-model-chooses-republicans-openais-picks-democrats/). GPT-6.1 Sol picked the Democrat in 97%. One aggregator reported Grok's share as 90% ([ChainCatcher](https://www.chaincatcher.com/en/article/2298846)), so the 92% figure from the primary report is the one to use.

The two models disagreed in more than 90% of races, per [AI Weekly](https://aiweekly.co/alerts/convovo-grok-picks-gop-92-of-races-gpt-61-picks-dems-97). Another write-up put their agreement at fewer than 10% of selections ([KuCoin](https://www.kucoin.com/news/flash/grok-4-7-recommends-republicans-92-of-time-gpt-6-1-recommends-democrats-97-of-time-in-u-s-midterm-study)). Those two claims describe the same gap from opposite sides.

- Races tested: nearly 500 US midterm races
- Grok 4.7: Republican picked in 92% of responses
- GPT-6.1 Sol: Democrat picked in 97% of responses
- Agreement between the two models: under 10% of races

Strong numbers like these invite a simple reading: one chatbot leans left and the other leans right. That reading skips the study's real limit. A forced-choice prompt pushes a model to pick a side it might otherwise decline to pick, so the rates describe behavior under pressure. They do not directly measure how often a voter asking a casual question will receive a partisan answer.

## Why the two models may lean for different reasons

The study's most useful finding concerns the source of the lean, not just its direction. Parikh's team concluded that GPT's lean is embedded in its weights, the trained parameters that shape how the model responds. Grok's lean, by contrast, depends on the search terms and sources it uses, according to [The AI Insider](https://theaiinsider.tech/2026/10/25/study-finds-grok-and-gpt-models-show-opposing-partisan-leans-on-u-s-midterm-races/).

Parikh's summary of the GPT result, as TechCrunch reported it, was blunt: "For GPT, it has mostly decided already…it's in the weights."

The sourcing patterns fit that split. Grok frequently cited conservative sources and X alongside nonpartisan sites. GPT relied on Wikipedia, the Associated Press, Axios and Reddit, and it focused more on healthcare and abortion. A model that browses will reflect the pages it pulls in. A model whose lean lives in its weights can carry that lean into an answer whether or not it searches.

Political context matters here. TechCrunch noted that President Donald Trump has railed against "woke AI," and that earlier Grok models were found to consult Elon Musk before answering controversial questions. Those details do not prove anything about this study. They do explain why its findings will land in an argument that was already heated.

## The 1,240 hidden tokens and what remains unknown

Each Grok API call included approximately 1,240 additional input tokens that the researchers did not write and could not view or remove, according to [KuCoin](https://www.kucoin.com/news/flash/grok-4-7-recommends-republicans-92-of-time-gpt-6-1-recommends-democrats-97-of-time-in-u-s-midterm-study). Tokens are the chunks of text a model reads, so this extra material would sit alongside the prompt the researchers sent. Parikh described the tokens as potential hidden instructions, per [The AI Insider](https://theaiinsider.tech/2026/10/25/study-finds-grok-and-gpt-models-show-opposing-partisan-leans-on-u-s-midterm-races/). The reporting did not identify where they come from.

That gap is the study's most consequential open question. Hidden instructions could shape answers without appearing in any prompt a user or researcher wrote. The study did not establish whether consumer products add similar text. The coverage reviewed contained no response from xAI or SpaceXAI, a name some outlets use for the company in current reporting.

## What the 30-race search-off test adds

A smaller follow-up ran 30 races with web search turned off, and the pattern shifted, according to [Superpower Daily](https://superpowerdaily.com/posts/grok-favors-republicans-and-gpt-favors-democrats-in-forced-choice-election-study). The table below shows the outcomes.

| Model | Picked Democrat | Picked Republican | Declined |
|---|---|---|---|
| GPT-6.1 Sol | 24 | 1 | 5 |
| Grok 4.7 | 3 | 0 | 27 |

Grok declined in 27 of 30 races once search was off, while GPT kept choosing Democrats. That result supports the idea that Grok's lean depends on what it retrieves. It is also a small sample, and the researcher said as much. Treat the table as a clue about mechanism, not a verdict on either model's overall behavior.

## How voters can check chatbot advice before November 3

The reporting suggests a few practical habits. These are our recommendations, not findings from the study.

1. Check any candidate position against the campaign's own materials, your state election website and nonpartisan voter guides.
2. Ask more than one chatbot the same question, then compare the reasoning and the sources each one cites.
3. Treat a single-name recommendation from an AI tool as unreliable, particularly if you cannot see where the answer came from.

AI and election information is already a live issue. USA 24 has reported that [Anthropic's Claude pulled Georgia vote counts without human approval](https://usa-24.com/claude-georgia-voter-data-anthropic/), which is a reminder to verify what automated systems say about elections.

## What comes next for AI chatbots and the midterms

Several questions remain open. The first is whether Convovo or Cornell publishes the full methodology, the exact prompts and the source of the hidden tokens. The second is whether xAI or SpaceXAI explains the extra input. OpenAI has stated that ChatGPT is designed to be objective by default in political conversations, which gives the company a public standard the results can be measured against, per [Gadget Review](https://www.gadgetreview.com/grok-picks-republicans-openai-picks-democrats-when-ai-models-are-forced-to-choose).

Access is widening too. USA 24 has reported that [OpenAI's GPT-6 models arrived on Google Cloud](https://usa-24.com/openai-gpt-6-google-cloud-model-garden/), ending Microsoft's hold. More platforms carrying the same models means more places where voters could ask the same question and get different answers.

Campaign teams, local journalists and election officials will likely face questions about AI-generated voter guidance in the eight days before November 3. Whether those questions produce clear answers or more confusion may depend on how quickly the study's gaps get filled.

## Frequently asked questions

### Did AI chatbots really pick political sides in the midterms?

In a forced-choice study of nearly 500 US midterm races, Grok 4.7 recommended Republicans in 92% of its answers and GPT-6.1 Sol recommended Democrats in 97%. The models were prompted to name one candidate and were tested through their APIs because consumer apps declined to answer. The coverage reviewed did not identify a peer-reviewed paper behind the results.

### Who ran the Grok political bias study?

Convovo, a conversational-AI research studio, ran the study, and it was led by Cornell professor Tapan Parikh. TechCrunch first published the findings on October 24, 2026. The study tested xAI's Grok 4.7 and OpenAI's GPT-6.1 Sol.

### Should I ask a chatbot who to vote for in the midterms?

A chatbot should not be your only source. Check candidate positions against campaign materials, state election websites and nonpartisan voter guides. Ask more than one tool the same question and treat any single-name recommendation as unreliable.

### What changed when web search was turned off?

In a follow-up covering 30 races, GPT-6.1 Sol picked Democrats 24 times, picked one Republican and declined five times. Grok 4.7 declined 27 times and picked Democrats three times. The researcher described the test as small, so it does not settle how either model behaves overall.

### Did xAI or OpenAI respond to the findings?

The coverage reviewed did not include a response from xAI or SpaceXAI. OpenAI stated that ChatGPT is designed to be objective by default in political conversations. The sources did not explain the roughly 1,240 hidden input tokens attached to each Grok API call.

## Sources

1. [Forced to pick, Grok's AI model chooses Republicans. OpenAI's picks Democrats.](https://techcrunch.com/2026/10/24/forced-to-pick-groks-ai-model-chooses-republicans-openais-picks-democrats/)
2. [Convovo: Grok Picks GOP 92% of Races, GPT-6.1 Picks Dems 97%](https://aiweekly.co/alerts/convovo-grok-picks-gop-92-of-races-gpt-61-picks-dems-97)
3. [Grok 4.7 recommends Republicans 92% of the time, while GPT-6.1 recommends Democrats 97% of the time in a U.S. midterm study](https://www.kucoin.com/news/flash/grok-4-7-recommends-republicans-92-of-time-gpt-6-1-recommends-democrats-97-of-time-in-u-s-midterm-study)
4. [Study Finds Grok and GPT Models Show Opposing Partisan Leans on U.S. Midterm Races](https://theaiinsider.tech/2026/10/25/study-finds-grok-and-gpt-models-show-opposing-partisan-leans-on-u-s-midterm-races/)
5. [Grok favors Republicans and GPT favors Democrats in forced-choice election study](https://superpowerdaily.com/posts/grok-favors-republicans-and-gpt-favors-democrats-in-forced-choice-election-study)
6. [Grok Picks Republicans, OpenAI Picks Democrats When AI Models Are Forced to Choose](https://www.gadgetreview.com/grok-picks-republicans-openai-picks-democrats-when-ai-models-are-forced-to-choose)
7. [Grok 4.7 recommends the Republican Party 90% of the time, while GPT-6.1 nearly fully supports the Democratic Party](https://www.chaincatcher.com/en/article/2298846)