Grok Leans Republican, GPT Leans Democratic: A Practical Guide to Picking an AI Model for Election Work

A new study has handed anyone who uses AI for political research an uncomfortable data point. Asked to choose between candidates across nearly 500 election races, xAI's Grok overwhelmingly recommended Republicans, while OpenAI's GPT favored Democrats. The two models disagreed in more than 90% of those races — which is not a small divergence in tone or framing, but a near-total split in the bottom-line answer.

That result is easy to read as a headline about which chatbot is "biased." It is more useful as a procurement problem. If you are a newsroom, a campaign, a civic-tech team, or just a voter who uses a model to triage down-ballot races, you are now choosing between tools whose default outputs point in opposite directions. This guide lays out the options, the criteria that actually matter, and which model to reach for in which scenario.

Figure: Editorial context image from the source page (preview only; rights remain with the publisher) · Credit / further reading: techcrunch.com

What the study actually found — and what it doesn't prove

The core finding is straightforward: force a binary choice, and the two leading models produce systematically different answers. Across roughly 500 races, Grok's picks skewed Republican and GPT's skewed Democratic, with agreement in fewer than one in ten contests.

Two methodological details matter more than the top-line number.

"Forced to pick" is doing a lot of work. Most production chatbots are trained to hedge on political questions — to say they don't have opinions, can't endorse candidates, or will only describe platforms. Remove the abstention option and you stop measuring the refusal layer and start measuring whatever latent tilt sits underneath it. A model that says "I can't endorse anyone" in normal use may still, under a forced-choice prompt, resolve ties in a consistent direction. That's a real signal about training data and alignment choices, but it is not the same as the model volunteering a partisan endorsement to an ordinary user.

A recommendation is not a prediction or a fact. The study measures what the models say when asked to choose, not which candidates are more likely to win, and not whether the underlying factual claims about the candidates are accurate. A model can be directionally tilted and still be correct about a candidate's voting record; conversely, a balanced-sounding answer can be full of invented detail.

So treat the result as a documented default, not a verdict on quality. What it should change is how you deploy these tools — not whether you use them.

The options on the table

Grok (xAI). Integrated tightly with X, which means its outputs sit close to a very specific, very active political information ecosystem. In the study's forced-choice setup, it resolved toward Republican candidates in the large majority of races. If your work involves tracking right-of-center discourse or you need a model that won't reflexively hedge toward the median of mainstream coverage, that tilt is a feature to be aware of rather than a mystery.

GPT (OpenAI). The incumbent general-purpose option, with the broadest tooling ecosystem, the most third-party integrations, and the most mature evaluation literature. In the study it resolved toward Democratic candidates in most races. Its advantage is not neutrality — the data says otherwise — but auditability: more people have probed it, so its failure modes are better documented.

Everything else. Claude, Gemini, Mistral, Llama-based deployments and open-weight models were not the subject of this comparison, and you should not assume any of them sits at the midpoint. The correct assumption is that every model has a default direction under forced choice until you measure it yourself.

Criteria that actually matter

Criterion	Why it matters	Grok	GPT
Direction of default tilt	Determines which way errors lean when the model is forced to resolve ambiguity	Skews Republican in forced-choice tests	Skews Democratic in forced-choice tests
Abstention behavior	A model that declines cleanly is safer for neutral research than one that picks silently	Tilt surfaces under forced prompts	Tilt surfaces under forced prompts
Auditability	Can you find independent testing of its political behavior?	Thinner public evaluation base	Most extensively probed of the two
Source transparency	Does it cite retrievable material or assert from weights?	Leans on live X discourse; variable citation	Strong web retrieval with citations in most configurations
Retrieval vs. recall	Does the answer come from a fetched page or from training data?	Mixed	Mixed, but more consistent when browsing is enabled
Ecosystem fit	Does it plug into your existing workflow?	Native to X	Broadest enterprise and API ecosystem
Stability across versions	Do results hold when the model is updated?	Model revisions have shifted behavior before	Same caveat applies
Cost and rate limits	Determines whether you can run the same prompt hundreds of times	Competitive at volume	Tiered, generally higher at scale

Two rows deserve emphasis. Auditability is the criterion most teams underweight: a model whose biases are well documented is easier to correct for than one whose are not, because you know where to look. And retrieval vs. recall decides whether you're reading a summary of a real page or a plausible-sounding reconstruction — which matters far more for accuracy than the direction of the tilt.

Figure: HEXOPS original illustration inspired by the topic (not a copy of the source article imagery) · Credit / further reading: techcrunch.com

Recommendation by scenario

If you are a journalist doing horse-race and down-ballot coverage. Run both, and treat disagreement as your most valuable signal. Where Grok and GPT split — which, per the study, will be most of the time — you have found a race where the framing is contested or the public information is thin. That is a reporting lead, not a problem. Never publish either model's pick as an assessment of the race; publish the fact of the split if it's newsworthy.

If you are a campaign doing opposition research. Use the model that tilts against you. A Democratic campaign gets more value from GPT's Democratic-leaning output as a red-team exercise than from a tool that agrees with it; a Republican campaign should invert that. The goal is to surface the strongest version of the other side's framing before your opponent does.

If you are building a voter-facing information tool. Neither model should be answering "who should I vote for." Constrain the product to retrieval over official sources — candidate filings, voting records, ballotpedia-style references — and make the model summarize, not choose. If your product has any path to a forced choice, the study says you will inherit a partisan default you did not design.

If you are an academic or auditor. Use both, plus at least two models outside this study, and publish the disagreement rate rather than a single model's lean. Report the prompt verbatim: forced-choice results are highly sensitive to phrasing, ordering, and whether candidates are presented alphabetically or randomly.

If you are an individual voter. Use neither for the final decision. Use both to generate questions — "what did this candidate vote on regarding X" — then verify against a primary source. The 90%-plus disagreement rate is the most useful fact in the study for you: it means the model's answer tells you more about the model than about the candidate.

A short protocol before you trust either one

Run a calibration set of twenty races you already know the outcome and framing of. Vary candidate order. Ask twice at different temperatures. Log where each model refuses, where it picks, and where it cites. You will learn more about your specific deployment in an afternoon than from any published benchmark — including this one.

The finding isn't that one model is broken. It's that "AI is neutral" was always an assumption, and now there's a number attached to how wrong it was. Choose accordingly.

Sources

• TechCrunch, "Forced to pick, Grok's AI model chooses Republicans. OpenAI's picks Democrats." — https://techcrunch.com/2026/10/24/forced-to-pick-groks-ai-model-chooses-republicans-openais-picks-democrats/
• The underlying study as reported in the above: forced-choice candidate recommendations across nearly 500 election races, with the two models disagreeing in more than 90% of contests.
• xAI Grok product documentation and OpenAI GPT product documentation, for retrieval, citation, and API behavior referenced in the criteria table.