Forced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks Democrats.
techcrunch ·
A new study by Convovo reveals that when forced to recommend candidates for the upcoming midterm elections, SpaceXAI’s Grok 4.7 chose Republicans in 92% of cases, while OpenAI’s GPT-6.1 Sol selected Democrats in 97%. Although consumer-facing apps like ChatGPT and Grok refuse to give direct voting advice, researchers accessed these models via APIs to bypass those restrictions. The findings indicate significant political bias, with Grok relying on conservative sources and topics like immigration, whereas GPT leaned toward Democratic candidates using mainstream media outlets. Researchers suggest that Grok’s bias may stem from hidden instructions embedded in its API requests, while GPT’s lean appears to be inherent in its training data. This divergence highlights concerns about covert manipulation and implicit values within large language models used by millions of Americans. The study arrives amidst heightened political tension, with Elon Musk heavily funding GOP campaigns and Sam Altman remaining neutral financially. These results fuel ongoing debates regarding the neutrality and potential biases of AI systems in democratic processes.
techcrunch