AI Economy
New comparisons of Grok and ChatGPT reveal that major AI models produce systematically different political outputs — and neither company has a satisfying explanation for why.
NewsOnScale Staff
October 26, 2026
There is a question that AI companies have been dodging since the first chatbot learned to discuss abortion policy: whose values are baked into the model? A new round of testing appears to have forced the issue into the open, with side-by-side comparisons of xAI's Grok and OpenAI's ChatGPT producing starkly different political outputs when both systems were asked to make partisan choices.
The results, which have circulated widely in technology and political media circles, suggest something more structural than a one-off glitch. Grok, the model built by Elon Musk's xAI, skewed toward Republican-aligned answers when pressed. OpenAI's models leaned Democratic. Neither company has offered a detailed technical account of how this happens or what they intend to do about it.
## Why This Is Not a Fringe Concern
The AI industry has long tried to frame political bias in its models as a calibration problem — something to be tuned away in future updates, like a noisy sensor. That framing lets companies avoid the harder question: whose calibration, toward whose definition of neutrality?
The scale at which these systems operate makes the stakes concrete. ChatGPT alone reports hundreds of millions of active users. Grok is deeply integrated into X, a platform that serves as a primary news source for a significant portion of the American electorate. When these systems subtly frame political questions, suggest which concerns are legitimate, or — as this testing shows — explicitly favor one party's worldview when forced to choose, they are not passive information pipes. They are political actors, whether their operators intend that or not.
This is precisely the kind of infrastructure story that traditional tech coverage underweights. The bias is not dramatic. It does not look like a campaign ad. It accumulates in thousands of small moments: a framing here, a recommended source there, an answer that treats one side's assumptions as baseline common sense.
## The Transparency Gap
What makes this story particularly significant for anyone tracking the AI agent economy is not that the models have political tendencies — that was always likely, given that they are trained on human-generated text produced within politically structured societies. The problem is that neither xAI nor OpenAI maintains anything approaching genuine public transparency about how political content is handled in model training, fine-tuning, or the reinforcement learning from human feedback that shapes final outputs.
OpenAI publishes a system card with some safety disclosures. xAI publishes considerably less. Neither document would allow an independent researcher to audit why Grok lands where it lands on a contested political question, or why ChatGPT lands somewhere different.
This is a governance gap with civic consequences. Political campaigns, advocacy organizations, and government agencies are already deploying these tools. Voters increasingly encounter AI-mediated information without any label indicating that the system answering their question was built with particular values, by particular people, under particular commercial pressures.
## What Accountability Would Actually Look Like
A minimally accountable standard for AI systems operating at civic scale would include, at minimum, published documentation of how politically sensitive training data is handled, independent third-party auditing of outputs across political categories, and clear disclosure to users when a system is generating content in domains where its training creates documented tendencies.
None of the major frontier model companies currently meet that standard. The testing that produced this week's headlines was done from the outside, by people with no access to training pipelines or internal evaluations.
The AI industry's preferred narrative is that it is building neutral tools that reflect the world as it is. The more honest version is that every model reflects choices — about data, about what counts as a good answer, about which human feedback to reward. Those choices have political consequences. Until companies are required to account for them in verifiable ways, the public is left inferring the architecture of its new information infrastructure from the outputs it occasionally notices.
That is not a technology problem. It is a power problem dressed in technical language.