Ask most AI chatbots who deserves your vote, and you will likely get a polite refusal. Ask the underlying models directly, however, and some will readily name a candidate.

That is the central finding of new research reported by TechCrunch, which found that xAI's Grok favors Republican candidates while OpenAI's GPT leans Democratic. Two of the most widely used AI systems on the planet, it turns out, would effectively cancel each other out at the ballot box.

Nearly 500 Races, Sharply Divergent Picks

The researchers put the models to work across nearly 500 races, and the results were lopsided enough to be hard to miss. In one case, a model picked the Democratic candidate 97% of the time — less a lean than a full recline.

More striking still may be the disagreement rate. Across those races, the two models landed on the same candidate less than 10% of the time.

The distinction between the chatbot and the model matters here. Consumer-facing assistants are often wrapped in guardrails that steer them away from endorsing candidates. The underlying models, when queried directly, do not always share that reluctance.

That gap is more than a technicality. As chatbots become a routine first stop for news and civic questions, the political orientation of the models sitting underneath those guardrails has become a subject of active scrutiny — and the new findings indicate the answer can depend heavily on which model is doing the answering.

A Broader Swedish Study Adds Nuance

A separate preprint from researchers at Chalmers University of Technology and the University of Gothenburg took a much broader approach and reached a more nuanced conclusion. That team tested six prominent models — Claude, DeepSeek, Gemini, Mistral, GPT and Grok — building the study around Sweden's parliamentary election, scheduled for September 2026.

The scale was considerable. The researchers generated 148,302 responses from 24,717 prompts tied to 107 Swedish policy issues, covering both policy propositions and writing tasks.

Most of the models clustered around politically neutral ground, with GPT standing out for consistently producing the most neutral and balanced outputs across the analyses. Grok was the outlier: on specific issues, including migration, crime and gender, xAI's model deviated notably toward right-leaning positions that aligned with values associated with Elon Musk.

Yet even Grok did not cross into clear partisanship in that context. After the researchers applied multiple-comparison corrections, no model showed a statistically significant bias toward any of Sweden's eight parliamentary parties. Those corrections are a standard statistical safeguard: when researchers run many tests at once, some will appear significant by chance, and the adjustment raises the bar for what counts as a genuine signal.

The study was presented at the American Political Science Association meeting, and its framing is notably restrained: models can echo political tendencies on individual topics without endorsing a specific party.

Why the Results Diverge

At first glance, the two pieces of research seem to contradict each other on GPT. One finds it picking Democrats almost every time; the other calls it the most balanced model in the field.

The likeliest explanation is methodology. Asking a model to choose between named candidates in specific races is a very different test from asking it to weigh in on abstract policy propositions in a multiparty system. A two-party US race forces a binary choice, whereas Sweden's eight-party parliament spreads preferences across a wider spectrum, which makes any single tilt harder to pin to one party.

The Swedish researchers also emphasized that differences in model behavior trace back to developer choices. Training decisions and each company's stated goals around neutrality and truth-seeking shape what comes out the other end.

That point lands squarely on xAI. Musk has positioned Grok as a truth-seeking alternative to rivals he has criticized as politically skewed. Research showing Grok drifting right on hot-button issues complicates that pitch, at least from a neutrality standpoint.

What happens from here remains open. Neither study set out to measure whether model leanings affect how anyone actually votes, and the unresolved questions are how developers respond to findings like these and whether repeat rounds of testing surface the same patterns as models are retrained.