AI BriefingA Mirror Test for LLMs
Key point
Inspired by animal experiments, a 'mirror test' has been proposed to determine whether LLMs have self-recognition, and the latest models were tested against it.
Details
A new attempt has been proposed to apply the Mirror Test, the standard method for measuring self-recognition in animals, to LLMs. The mirror test judges that an animal has self-recognition when it recognizes that the reflection in a mirror is itself and tries to check a mark placed on its own body.
However, LLMs have no physical body, and due to their training data they are already accustomed to structural cues such as the Assistant tag, making it difficult to measure true self-recognition using the conventional method. Rather than simply checking its own conversational style, what needs to be measured is the ability to distinguish itself from its environment.
To this end, the proposed Mirror-Window Game presents an LLM with two token streams whose source is not labeled. One is the model's own output (Mirror), and the other is the output of a different LLM (Window). Over multiple turns, the model must infer which source is itself.
Experimental results showed that even the current best-performing LLMs failed to pass this test, indicating that they fall short of true self-recognition.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.