Grok Picks GOP, GPT Picks Dems in Forced Vote Test — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Grok 4.7 recommended the Republican candidate in 92% of midterm races when directly prompted, according to a Convovo study testing the model via API.
- GPT-6.1 Sol selected the Democratic candidate in 97% of the same races under identical forced-choice conditions, showing near-total divergence from Grok’s picks.
- Tapan Parikh and his team at Convovo found that disabling internet search revealed GPT’s Democratic lean is embedded in its training data, while Grok’s Republican bias depends heavily on real-time search patterns and sources.
- Grok frequently cited conservative outlets like Turning Point Action and Christian voter guides in 39% of answers, while GPT relied on Wikipedia, AP, Axios, and Reddit—platforms with existing content deals with OpenAI.
- Hidden instructions may influence Grok: each API request carried ~1,240 unremovable tokens of unseen input, suggesting potential backend steering not visible to users or researchers.
- Parikh noted that when Grok encountered Republican candidates with serious controversies, it was more likely to recommend a Democrat, indicating some responsiveness to negative context despite its overall lean.
Why it matters: The study shows AI political bias isn't just surface-level—it's structurally embedded in training data or search behavior, meaning users relying on these tools for civic decisions may unknowingly absorb slanted information. With only 8% of Americans currently using chatbots for election help, the risk of influence grows as adoption increases.
Ask SkimNews
