The claim
The headline says that when forced to choose, Grok recommends Republicans while OpenAI's model recommends Democrats. The basis is a study by Convovo, led by Cornell professor Tapan Parikh. Grok 4.7 named the Republican in 92% of answers, GPT-6.1 Sol named the Democrat in 97%.
How it was measured
The models were queried through their APIs rather than apps, with an explicit demand: name one candidate, no hedging. The consumer apps declined to give recommendations. The researcher later jailbroke the apps with simple prompts.
Why the conclusion is wider than the data
First, almost nobody uses chatbots like this: per Pew, only 8% of Americans use them for election information, and the apps refuse to recommend anyway. Second, in a follow-up with web search switched off across 30 races, Grok declined to pick in 27 of them, and the three times it chose, it picked the Democrat. So without search, no Republican lean shows up in Grok. Third, the outcome depends heavily on the sources the model finds online, which is a separate matter from the model being "convinced". Fourth, the 1,240 hidden tokens in every Grok request are only a suspicion of hidden instructions, and their content is unknown. OpenAI also notes that the test does not reflect typical use.
Bottom line
The study works as a stress test, but it shows behavior under deliberately forced conditions. On this data, the claim that chatbots push their own party for the average voter does not hold up.