Do LLMs pass the mirror test?
The article explores whether large language models (LLMs) can pass a version of the mirror test—a classic measure of self-awareness in animals. By prompting LLMs to recognize themselves in mirrored responses or textual self-references, the author finds that current models fail to demonstrate genuine self-recognition, highlighting limitations in their understanding of selfhood.
Background
- The mirror test is a classic psychology experiment used to assess self-awareness: an animal is marked with a spot of dye in a place it can only see in a mirror, and if it touches or investigates the mark, it's considered to have recognised itself.
- This post applies a similar concept to large language models (LLMs) like GPT-4, Claude, and Gemini — testing whether they can correctly identify themselves in a conversation when given a transcript that might or might not be their own output.
- The basic setup: the LLM is shown a dialogue history and asked "who said that last line?"; sometimes the last line was written by the model itself, sometimes by a different LLM or a human.
- This is an informal, blog-style investigation (not a peer-reviewed study), but it probes an important philosophical and technical question: do these models have any form of self-model or consistent identity, or are they just pattern-matching without a stable sense of self?