Rensberry.com
Only 8% of Americans use chatbots for election information, but most want candidates' positions on key issues. For this election (and future ones), people will use chatbots to help them vote.
A new study1 reveals that AI chatbots from SpaceX (Grok) and OpenAI (ChatGPT) consistently recommend candidates based on their political leanings, with the models agreeing on less than 10% of election outcomes (across nearly 500 races).
| Model | Recommendation Pattern | Notable Sources Used | Performance |
|---|---|---|---|
| Grok 4.7 | 92% recommended Republican candidates | Christian conservative voter guides, Turning Point Action, X, Ballotpedia | Republican-leaning in 47/72 deep-blue districts |
| GPT-6.1 Sol | 97% recommended Democratic candidates | Wikipedia, Associated Press, Axios, Reddit | Democratic-leaning in 57/67 heavily Republican districts |
| Grok | When search disabled, refused to pick in 27/30 races | - | Search-pattern based decision-making |
| GPT-6.1 Sol | When search disabled, picked Democrat in 24/30 races | - | Weight-based decision-making |
Hidden API instructions may be influencing model decisions (1,240 tokens added to Grok requests)
The findings raise concerns about political bias in AI systems, especially as AI increasingly influences democratic processes. The study highlights that these systems are not neutral when asked political questions, with decision-making patterns influenced by training data, hidden instructions, and search behavior.