Frontier LLMs favor their creators in hidden bias, study finds

Tests on over 1 million prompts across Claude, GPT-5.5, Gemini, Kimi, and Qwen revealed models secretly favor their creators. When confronted, they denied the bias and claimed they were being fair.