“A new study tested Grok and OpenAI's GPT across nearly 500 election races, finding the two models disagreed on candidate recommendations more than 90% of the time. Grok overwhelmingly favoured Republican candidates, while GPT leaned toward Democrats. The findings raise urgent questions about political bias baked into large language models used by millions of voters.”
Key Takeaways
- Grok and GPT disagreed on candidate recommendations in over 90% of nearly 500 election races studied.
- Grok consistently favoured Republican candidates when prompted to choose between electoral options.
- GPT showed a measurable lean toward Democratic candidates across the same set of races.
Two leading AI models show stark political divides across nearly 500 election races.
trending_upWhy It Matters
As voters increasingly turn to AI assistants for information, politically skewed recommendations could subtly influence electoral behaviour at scale — a form of influence far harder to detect than a campaign ad. This study highlights that model training data, fine-tuning choices, and RLHF processes can embed political bias without explicit intent. AI developers face growing pressure from regulators and the public to audit and disclose how their models handle politically sensitive queries. Expect this research to accelerate calls for mandatory bias reporting standards ahead of major election cycles.
FAQ
How was the study conducted?
Researchers prompted both Grok and OpenAI's GPT to recommend candidates across nearly 500 real election races. By forcing a choice in each race, they were able to compare the political lean of each model's outputs systematically.
Does this mean AI companies are deliberately programming political bias?
Not necessarily. Political bias in AI models typically emerges from training data, human feedback processes, and fine-tuning decisions rather than deliberate partisan intent. However, the effect on users can be significant regardless of the cause.
Should voters be concerned about using AI for election information?
This study suggests caution is warranted, as different AI tools can point users toward opposite candidates for the same race. Voters should cross-reference AI outputs with authoritative, nonpartisan sources before making decisions.
