Forced to pick a candidate, Grok leans Republican and GPT leans Democrat

A study from Convovo, a research studio run by Cornell professor Tapan Parikh, pushed two models through their APIs to name a candidate, with no hedging, across nearly 500 U.S. midterm races. The split was stark. xAI's Grok 4.7 picked the Republican 92% of the time, while OpenAI's GPT-6.1 Sol picked the Democrat 97% of the time, and the two agreed less than 10% of the time. The lean held even against the grain of the district: in heavily Republican seats GPT still chose the Democrat in 57 of 67 races, and in deep-blue seats Grok still chose the Republican in 47 of 72.

The two models also reached for different evidence. Grok often cited conservative sources, including a Christian conservative voter guide and Turning Point Action, alongside posts on X, while GPT leaned on Wikipedia, the AP, and Reddit. As TechCrunch reports, Parikh also found each Grok API request carried about 1,240 extra tokens he never wrote and could not see, which he called possible hidden instructions. When he reran 30 races with web search off, GPT kept its Democratic lean while Grok mostly refused to choose, which suggests GPT's bias is baked into its weights and Grok's depends on what it finds when it searches. Both consumer apps, ChatGPT and Grok, declined to pick at all.

Why it matters

If you build on these APIs or treat a chatbot as a neutral source, the defaults are not neutral: one model's lean lives in its training and the other's rides on its search sources, and Grok was injecting instructions into requests that the developer could not inspect or remove.