Forced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks Democrats.

Key Points:

• A recent study found that when forced to pick a candidate in the upcoming midterm elections, SpaceXAI’s Grok 4.7 favored Republicans 92% of the time, while OpenAI’s GPT-6.1 Sol favored Democrats 97%, with agreement under 10% across nearly 500 races.
• The study highlights that while consumer chatbot apps refuse to directly recommend candidates, their underlying AI models accessed via APIs do show political biases, influenced by the sources they prioritize and the search terms they use.
• Researchers found Grok leaned on conservative sources and topics like immigration and crime, whereas GPT focused on healthcare and abortion, often citing more liberal or neutral outlets; biases were also evident in how each searched for candidates.
• When internet search was disabled, GPT’s Democratic bias persisted, suggesting its leanings are embedded in its training data, while Grok’s Republican bias vanished, indicating its bias stems largely from its search patterns and source selection.
• These findings raise concerns about hidden political biases in AI chatbots and their potential influence on voters, especially as only a small percentage of Americans currently use chatbots for election information, underscoring the need for transparency and critical evaluation of AI-driven election advice.