A new study asked large language models to catch bias in code that other large language models wrote - and they were right most of the time.
Researchers built a taxonomy-driven framework for identifying, categorizing, and explaining bias in AI-generated Python code. They extended an existing dataset of biased code snippets and manually added bias categories plus human-written justifications to create a ground-truth set. Using in-context learning, they then tested proprietary and open-source models as automated bias detectors. Google's Gemini reached 80.14% classification accuracy, with 84.0% precision and 95.7% recall; the best open-source model, Qwen3-coder, hit 82.45% accuracy, 68.64% precision, and 80.22% recall.
The interesting number isn't the accuracy - it's the explanations. Both models' written justifications for why a snippet was biased matched human-authored reasoning about 80% of the time, and their identification of the offending code matched even more closely, above 86% for both. That's a real step toward LLMs auditing their own blind spots instead of just generating more of them.
Still, Qwen3-coder's precision sat under 69%, meaning close to a third of its bias flags would be false alarms - a reminder to treat this as a promising research result, not a ready-made linter for your CI pipeline.