Do LLMs pass the mirror test?

The article examines whether large language models possess self-awareness by proposing a new "mirror test" adapted from Alexandra Horowitz's olfactory mirror test for dogs. Rather than asking LLMs to identify their own outputs, the author suggests the proper test involves subtly modifying an LLM's previous responses in a conversation and observing whether the model detects the discrepancy—analogous to how dogs notice when their scent is altered. The author argues this approach better measures what mirrors tests actually test: anomaly detection against an internal baseline, and demonstrates the concept using Google's Gemma model.