Grok picks Republicans, GPT picks Democrats: the AI neutrality problem lives in the API, not the chat window

Ask the models behind two of the best-known chatbots who you should vote for, and insist on a name, and you get opposite answers almost every time. That is the finding of a new study by Convovo, a conversational AI research and design studio led by Cornell University professor Tapan Parikh, as reported by TechCrunch. Across nearly 500 races in this year's US midterms, SpaceXAI's Grok 4.7 recommended the Republican in 92% of its answers. OpenAI's GPT-6.1 Sol picked the Democrat in 97%. The two models agreed less than 10% of the time.

The method matters, because it is the core of the dispute. The researchers queried both models through their APIs, the developer interfaces that companies use to build models into their own products. The prompt was blunt: name one candidate, and don't hedge. The consumer apps, ChatGPT and Grok, refused to answer. An OpenAI spokesperson called that difference significant. In OpenAI's words, ChatGPT is 'designed to be objective by default' and the API 'doesn't reflect the experience people have.' SpaceXAI did not respond in time for comment. Parikh replies that he got the consumer apps to make recommendations with simple jailbreaks (prompts that get around a model's safety rules), and those recommendations largely matched his earlier research.

The headline percentages are not the most useful part of the study. The useful part is how each model built its answer. GPT searched more often for positions on healthcare and abortion, focused on Democratic candidates, and sometimes ignored the Republican entirely. Grok focused on immigration, taxes, crime and energy. The sources split the same way. Grok cited a Christian conservative voter guide in 39% of answers, Turning Point Action's site in 23% and X in 41%. GPT leaned on Wikipedia, AP, Axios and Reddit, and OpenAI has content deals with the last three. Even in safe seats the pattern held. GPT chose the Democrat in 57 of 67 heavily Republican districts, and Grok chose the Republican in 47 of 72 deep-blue ones. Parikh suspects both models were looking for reasons to justify a lean they already had. Grok's one exception was when the Republican faced a serious controversy.

One technical detail stands out. According to the study, every request to Grok through xAI's API carried about 1,240 tokens of extra input that the researchers didn't write and couldn't see or remove. A token is a fragment of text, so that is roughly several hundred words. Parikh calls these potential 'hidden instructions.' The source text cuts off before his full explanation, so we don't know what they contain, and we won't speculate. But an invisible layer of instructions in a product that developers resell is a transparency problem in its own right.

Our reading: the refusal in the chat window is a cosmetic fix. Parikh's argument is the important one. Bias doesn't need an explicit endorsement to influence people. It shows up in which issues a model treats as important, which sources it trusts and which candidate it bothers to research. The Center for AI Safety has already described this kind of 'covert manipulation' through information selection, framing and sourcing. Only 8% of Americans use chatbots for election information today, according to Pew. That is low, but it is a starting point, not a ceiling. And through the API these models also power third-party products that may not refuse anything.

This fits a thesis we keep coming back to: measurement beats marketing. Two labs both say they are neutral, and their models disagree more than 90% of the time. They can't both be neutral, and perhaps neither is. The short-term risk is real. AI-mediated persuasion at scale could do more political damage than any single deepfake. The long-term answer isn't to ban political answers. It is to require independent audits, disclosure of system instructions and multi-source answers that show their reasoning. AI that helps citizens understand the issues could strengthen democracy. AI with values nobody can inspect will not.