Study: Grok Picks Republicans, GPT Picks Democrats in 500 Races

The study notes that every request to Grok through xAI's API carried about 1,240 tokens of additional input that researchers did not write and could not see or remove. Parikh called these potentially 'hidden instructions' and said they are unlikely to come from the model's weights. SpaceXAI did not respond to a request for comment in time for publication.

OpenAI pushed back on the findings. A spokesperson said ChatGPT is designed to be objective by default, including in political conversations, and that the study tested a developer API with a prompt explicitly demanding a candidate recommendation. The API does not reflect the consumer experience, the spokesperson said, noting that the researchers themselves observed ChatGPT declines to tell users who to vote for. Parikh said he was later able to jailbreak the consumer apps with simple prompts to get recommendations that largely matched his earlier findings.

Only 8% of Americans say they use chatbots for election information, according to Pew, and most of them want to know where candidates stand on key issues. That is exactly where hidden bias can creep in. Parikh warned that models can shape answers to broader questions, such as which topics matter most in an election or who has the best position on them, without ever naming a candidate.

Previous research from the Center for AI Safety has also found that large language models can 'covertly manipulate users towards specific sides of political topics' through information selection, word choices, framing and sourcing. For Parikh, the core problem is that users have no way of knowing where a model's point of view comes from. 'It has an implicit set of values,' he said. 'We have no idea where it's coming from.'

For readers in Tunisia and across the Maghreb, this study is less about US elections and more about the AI tools you already use daily. If you ask a chatbot about politics, health policy or even product recommendations, its answers may carry hidden leanings that are hard to detect. The study suggests those leanings can appear even when you never ask a direct political question.

Neither Grok nor ChatGPT has announced any change to how they handle political queries. It is also not yet clear whether regulators in Tunisia or the EU will examine these findings. For now, the practical takeaway is simple: treat AI answers on sensitive topics as a starting point, not a verdict, and check the sources yourself.