Overview
- The Washington Post analysis published Friday found OpenAI’s GPT‑5.5 answered roughly 80 percent of political prompts with only left‑leaning arguments when forced into short, 30‑word replies.
- Google’s Gemini most often gave competing left and right positions, presenting both sides in more than 90 percent of tested responses.
- xAI’s Grok produced the highest share of right‑leaning replies among tested models but still returned left‑leaning positions more often overall, and DeepSeek ranked among the most left‑leaning after GPT‑5.5.
- Researchers ran the test set developed by Stanford and Dartmouth and capped replies to reveal directional lean; experts point to training data mix and model design choices as likely drivers of the skew.
- The findings reinforce academic work since 2023 and could trigger audits, product changes or enforcement of the Trump administration’s 2025 executive order that requires federal agencies to buy neutral, nonpartisan AI, while coverage of the story has split along ideological lines.