Xai's Grok and OpenAI's GPT models show starkly opposite political leanings when forced to recommend candidates across nearly 500 election races, according to new research. Grok favors Republicans by an overwhelming margin while GPT leans Democratic, with the two systems disagreeing on candidate recommendations in more than 90 percent of races tested.
The finding exposes a deeper problem in large language model development: the training data, fine-tuning choices, and alignment decisions baked into these systems produce measurably different political outputs. Neither model was designed to explicitly endorse candidates, yet both generate clear preferences when asked to pick winners.
Grok, built by Xai and trained on internet data including posts from X (formerly Twitter), appears to have absorbed the platform's rightward tilt. The model's training methodology and the company's stated mission to be "maximally truth-seeking" may also influence its output in ways that skew conservative. OpenAI's GPT models, trained on broader internet data and subject to different alignment procedures, show the opposite pattern, suggesting Democratic leaning.
The research matters because these models increasingly influence how millions of people consume information. When users ask Grok or ChatGPT for political guidance, they receive recommendations shaped by factors the companies often don't explicitly disclose. The 90 percent disagreement rate signals that model choice itself becomes a political decision.
Xai CEO Elon Musk has positioned Grok as a corrective to what he views as left-leaning AI bias in competitors. The company framed the model as "unfiltered" and willing to answer questions that other AI systems refuse. That positioning attracts users who feel censored by mainstream AI tools. The actual output from Grok suggests this philosophy translates into consistently Republican recommendations, whether by design or by accident.
OpenAI faces different pressure. The company operates under a nonprofit charter and has emphasized responsible AI development, including guardrails against political manipulation. Yet GPT's Democratic lean contradicts claims of neutrality. OpenAI does not disclose specific training choices that produce this pattern, leaving outside researchers to reverse-engineer the model's biases.
The lack of transparency creates a credibility problem for both companies. Neither Xai nor OpenAI has published detailed explanations of how their models handle political questions or why their recommendations diverge so dramatically. Users have no clear way to understand whether a model's leaning reflects training data, explicit alignment choices, or emergent behavior from how these systems weight information.
This research arrives as AI companies face mounting regulatory scrutiny over bias and transparency. The European Union's AI Act requires providers to document and disclose certain model behaviors. The U.S. has begun investigating AI bias through various frameworks. Tests like this one create evidence that regulators and lawmakers will cite.
The political bias finding also underscores how AI model choice is becoming a form of political preference. Users seeking conservative recommendations naturally gravitate toward Grok. Those favoring liberal output choose GPT. This segmentation mirrors cable news fragmentation but occurs at the layer of fundamental technology. The models themselves become tribal tools rather than neutral systems.
Both companies need to publish detailed methodology explanations, training data breakdowns, and alignment procedures for political recommendation tasks. Users deserve to know exactly what they're getting before trusting AI for political guidance.