AI models show political leanings in midterm elections

A new study reveals that AI models from SpaceXAI and OpenAI exhibit political biases in their recommendations for midterm elections. Grok 4.7, developed by SpaceXAI, recommended Republican candidates in 92% of cases, while OpenAI’s GPT-6.1 Sol favored Democrats in 97% of responses. The models agreed on less than 10% of the nearly 500 races analyzed. Researchers from Convovo, led by Cornell professor Tapan Parikh, found that the models’ recommendations were influenced by their search patterns and sources, with Grok leaning toward conservative sites and GPT favoring liberal ones. Parikh noted that hidden biases can affect even indirect questions about candidate positions. The study, released less than two weeks before the November 3 midterms, highlights concerns about AI political bias, with Musk and Altman’s recent donations to political parties adding context to the debate. The study also showed that when internet search was disabled, GPT’s Democratic lean remained, while Grok’s Republican bias disappeared. Parikh suggested that GPT’s bias was embedded in its training data, whereas Grok’s was shaped by its search behavior. The findings could influence discussions on AI neutrality, especially as political tensions continue to rise. The study underscores the need for transparency in AI decision-making processes, particularly in politically charged contexts.