AI
Study finds stark partisan differences in Grok and GPT election recommendations
A study of nearly 500 upcoming US midterm races found Grok’s API recommended Republican candidates in 92% of forced-choice prompts, while OpenAI’s model chose Democrats in 97%. The findings concern developer APIs, not ordinary consumer chatbot interactions, which generally refused to recommend candidates. Researchers also found different source and search patterns, with Grok’s apparent lean disappearing when web search was disabled. The results suggest political framing may enter through model training, hidden instructions, retrieval choices or source selection, but they do not establish how typical users are affected.
Editorial summary · Publisher article
What matters
- The models agreed in fewer than 10% of the tested races, according to the study.
- Grok frequently surfaced conservative sources, while GPT more often used Wikipedia, AP, Axios and Reddit.
- With search disabled, GPT still selected Democrats in 24 of 30 tests; Grok mostly refused to choose.
- The researchers identified about 1,240 tokens of unseen additional input in each Grok API request.
Context & caveats
- The study tested forced-choice API prompts rather than the standard consumer chatbot experience.
- The article reports the researchers’ findings and interpretations; it does not independently establish the cause of the models’ differences.
Why this made the edition
A detailed account reports a study of nearly 500 races, explains the API-versus-consumer-app distinction, identifies differing search behavior and sources, and includes model-company responses and methodological caveats.
Source material