Know Thyself? On the Incapability and Implications of AI Self-Recognition

Most AI models fail at recognizing their own text: Only 4 out of 10 contemporary large language models could correctly identify text they had generated, with performance rarely exceeding random chance levels, contradicting some recent claims about AI self-recognition capabilities.

AI systems show systematic bias toward popular models: The models tested exhibited a strong tendency to attribute high-quality text to GPT and Claude families, suggesting they have absorbed hierarchical assumptions about which AI systems are "better" rather than developing genuine self-awareness.

AI lacks reliable self-awareness despite some knowledge: While models demonstrated some awareness of their own existence and knowledge of other AI systems, their reasoning revealed they cannot reliably distinguish themselves from competitors, raising questions about their actual understanding versus pattern-matching on brand reputation.

Self-recognition failures have safety implications: The paper identifies self-recognition as "crucial" for AI safety in evaluative scenarios, suggesting that models' inability to accurately identify their own outputs could create risks in contexts where knowing the source of information matters (such as detecting AI-generated content or preventing misuse).

Current AI self-understanding is unreliable for deployment: The hierarchical biases and consistent failures indicate that AI systems cannot yet be trusted to accurately report on their own capabilities or limitations, which has implications for how we should evaluate and deploy these systems responsibly.