The chatbots are already choosing sides, and they're bringing millions of users with them.
The Summary
- A new study reveals Grok's AI overwhelmingly recommends Republican candidates while OpenAI's GPT favors Democrats, with the two models disagreeing in over 90% of nearly 500 election races tested
- This isn't about politics. It's about training data, guardrails, and whose values get baked into the agents making decisions for millions of people
- The real story: we're automating influence at scale before we've solved bias at the model level
The Signal
The study forced both models to make endorsements across hundreds of local, state, and federal races. Grok recommended Republicans in the vast majority of contests, while GPT-4 leaned Democratic. The 90% disagreement rate isn't a rounding error. It's a feature, not a bug, of how these models were trained and tuned.
Here's what matters: these aren't fringe tools. Grok sits inside X, where 500+ million people scroll daily. ChatGPT has 200 million weekly active users. When you ask an AI assistant for voting recommendations, you're not getting neutral analysis. You're getting the embedded worldview of whoever built the safety layers, chose the training corpus, and decided what "helpful" means.
"The models disagree because they were taught to optimize for different definitions of truth."
Both companies will say they don't want their models making political recommendations. Both will point to guardrails and safety protocols. And yet, when pushed, the models pick sides with striking consistency. That tells you the bias isn't in the prompt. It's in the foundation.
This gets more interesting when you zoom out to the agent economy. Right now, we're debating chatbot political leanings. In 18 months, these same models will be:
- Screening job candidates and writing performance reviews
- Summarizing legal cases and drafting contracts
- Recommending investments and flagging compliance risks
If the models can't agree on political candidates, why would we expect them to converge on hiring decisions or risk assessments? The divergence isn't going away. It's going to compound as these systems get more autonomous.
The Implication
This is the canary in the coal mine for the agent economy. If foundation models from the two biggest AI labs see the world this differently, every company building on top of them inherits that divergence. The fix isn't better prompts. It's transparency about training data, alignment targets, and whose values are getting encoded at the base layer.
For builders: pick your model knowing you're picking a worldview. For users: understand that "AI-powered" increasingly means "optimized for someone's definition of correct." The agents are already making choices. The question is whether we'll know what they're optimizing for before we hand them the keys.