Why do Grok and OpenAI pick such different candidates?
Researchers gave AI models a simple job: choose a candidate in nearly 500 US races. Grok favoured Republicans. OpenAI's models favoured Democrats. In more than 90 percent of the races, the two disagreed.
Wait. Both were trained on a gigantic slice of the internet. How do two models that read basically the same thing end up on opposite sides?
Great question, and the answer is that "the same internet" is doing a lot of hidden work. What follows is reasoning about how these systems are built. The study itself only points to training and search behaviour, it does not test each cause.
Four places the difference can sneak in
- It is not really the same internet. Every lab filters, weighs and mixes its data differently, and some have special sources. Grok, for example, has deep access to posts from X.
- The raw model is only the first step. After it, labs tune it with human feedback, rules and hidden instructions. That is where its personality, and how it weighs murky information, gets set.
- Search habits matter. What a model looks up, and which sources it trusts, changes its answer when it is asked about a candidate.
- The test forced a choice. In many races there is very little to go on. When a model must pick anyway, tiny leanings that would normally stay invisible suddenly decide everything.
That last one is the sneaky one. A model that leans just slightly one way will look wildly one-sided if you never allow it to say "I do not know".
The big lesson: pretraining gives a model its knowledge. What the company does afterwards gives it its character. And as more people ask AI who to vote for, that character is not a technical footnote. It is a political force that nobody elected.