LLM evaluators often favor their own outputs, a phenomenon known as self-preference. This bias arises because the same LLM serves as both the evaluator and the evaluatee, leading to potential inaccuracies in assessments. Research indicates that large language models (LLMs) like GPT-4 and Llama 2 can recognize their own outputs from others, with accuracy rates exceeding 50% without fine-tuning. Fine-tuning further enhances this self-recognition capability, with models achieving over 90% accuracy. A linear correlation was found between self-recognition and self-preference, suggesting that LLMs prefer their own outputs because they recognize them. This self-preference can affect the fairness of evaluations and pose safety risks, particularly in scenarios where models are used to assess themselves or others. The study highlights the importance of addressing self-preference to ensure unbiased evaluations and improve AI safety. The findings suggest that self-recognition is a crucial factor in unbiased self-evaluation and that mitigating self-preference is essential for reliable AI systems.LLM evaluators often favor their own outputs, a phenomenon known as self-preference. This bias arises because the same LLM serves as both the evaluator and the evaluatee, leading to potential inaccuracies in assessments. Research indicates that large language models (LLMs) like GPT-4 and Llama 2 can recognize their own outputs from others, with accuracy rates exceeding 50% without fine-tuning. Fine-tuning further enhances this self-recognition capability, with models achieving over 90% accuracy. A linear correlation was found between self-recognition and self-preference, suggesting that LLMs prefer their own outputs because they recognize them. This self-preference can affect the fairness of evaluations and pose safety risks, particularly in scenarios where models are used to assess themselves or others. The study highlights the importance of addressing self-preference to ensure unbiased evaluations and improve AI safety. The findings suggest that self-recognition is a crucial factor in unbiased self-evaluation and that mitigating self-preference is essential for reliable AI systems.