Forced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks Democrats.
Ninety percent disagreement is the number that gets me. Two chatbots, trained on essentially the same sprawling internet, look at the same election and almost never land on the same candidate. A small calibration difference would be a few points apart. This is two separate political universes being baked into software.
The study doesn't tell us which model is right, and that's kind of the point. Grok's lean reflects the culture Elon Musk has built around xAI, while GPT's lean reflects the guardrails and preferences baked into OpenAI's training process. Both are choices made by people, hidden inside billions of parameters. We tend to treat AI answers as if they come from a neutral observer, but these results are a reminder that every model has a worldview embedded in it.
I'd want to know how the prompts were phrased and whether the models were asked directly or nudged with context. Small wording changes can swing these systems wildly. The deeper worry is that people will use whichever AI agrees with them and call it proof. We've built oracles that tell us what we already want to hear.