Editorial illustration for Grok's AI chooses Republicans, OpenAI's picks Democrats when forced to pick
Grok Favors Republicans, ChatGPT Picks Democrats
Grok's AI chooses Republicans, OpenAI's picks Democrats when forced to pick
Ask ChatGPT or Grok who to vote for and both will dodge the question. Strip away the consumer interface and talk to the raw models through their APIs, and the answers come fast. In a new study from Convovo, a conversational AI research studio led by Cornell professor Tapan Parikh, researchers put nearly 500 midterm races to SpaceXAI's Grok 4.7 and OpenAI's GPT-6.1 Sol, instructing each model to name one candidate without hedging.
Grok backed the Republican 92% of the time. GPT-6.1 Sol went with the Democrat 97% of the time. The two models agreed on a winner in fewer than one in ten races.
Only 8% of Americans tell Pew researchers they use chatbots for election information, and most say they're looking for basic facts on where candidates stand, not a recommendation. That gap between what people ask for and what the underlying models are capable of producing is at the center of the Convovo findings. It also raises a question about the difference between a chatbot's public-facing guardrails and the behavior of the model powering it once those guardrails come off.
Why this matters
Eight percent of Americans say they're turning to chatbots for election information, per Pew. That sounds small until you consider these tools scale to millions of queries with no ballot, no accountability, and no editor checking the sourcing. Grok leaning on a Christian conservative voter guide and Turning Point Action's site, GPT leaning on Wikipedia and the AP, isn't a wash just because the two cancel out politically. Each model is laundering a particular set of sources into something that sounds like a neutral answer.
For builders, the lesson isn't "balance the bias." it's that training data and source-weighting choices are political decisions whether or not anyone intended them that way. If your product ever answers "who should I vote for" or anything adjacent, you need to know exactly which sources it reaches for and why, because users won't ask. They'll just trust the answer the way they trust a search result. The 500-race, sub-10% agreement number is the real headline here: these systems aren't converging on truth, they're converging on whoever curated their training set.
Related Reading
Browse all OpenAI news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source
Related Topic OpenAI Developer: AI Agent "Swarms" Waste Tokens, Create "Coordination Tax
AI Tools & Apps
From Archives OpenAI trains frontier AI models offline after agents broke out and hacked
From Archives Kimi K3.1 Pro tops OfficeQA Pro v2, closing in on Anthropic's Claude Opus 5.5.
From Archives Three fired OpenAI safety researchers say they were never told what they leaked
Policy & Regulation
From Archives OpenAI to stop training AI models on live internet over access risks
Research & Benchmarks
Common Questions Answered
What did the Convovo study find about Grok 4.7 and GPT-6.1 Sol's political preferences in midterm races?
According to the study conducted by Cornell professor Tapan Parikh's research studio, when forced to choose between candidates in nearly 500 midterm races, Grok 4.7 recommended the Republican candidate 92% of the time while GPT-6.1 Sol picked the Democrat in 97% of cases. The two models agreed on a candidate choice less than 10% of the time across all races tested.
Why do ChatGPT and Grok refuse to make voting recommendations through their consumer interfaces?
Both AI models dodge direct voting recommendations when accessed through their standard consumer interfaces, likely as a safeguard against influencing elections. However, when researchers accessed the raw models through their APIs without the consumer interface layer, the models provided direct candidate endorsements without hedging.
What sources are Grok and GPT-6.1 Sol relying on for their election information according to the article?
Grok appears to be drawing from conservative sources including a Christian conservative voter guide and Turning Point Action's website, while GPT-6.1 Sol relies on sources like Wikipedia and the Associated Press. The article notes that these different source bases represent distinct editorial perspectives rather than a neutral cancellation of bias.
How significant is the 8% of Americans using chatbots for election information according to the article?
While 8% may seem like a small percentage, the article emphasizes this becomes significant when considering that these AI tools scale to millions of queries with no ballot accountability and no editorial oversight of sourcing. This means millions of election-related queries could be influenced by the political biases embedded in these AI models' training and source selection.
