https://www.giskard.ai/knowledge/llms-recognise-bias-but-also-reproduce-harmful-stereotypes

We released new findings from our Phare LLM Benchmark on bias in leading language models. Instead of traditional "fill-in-the-blank" tests, we had 17 leading LLMs generate thousands of stories, then asked them to judge their own patterns.

In short: Leading LLMs can recognise bias but also reproduce harmful stereotypes