Do LLMs pass the mirror test?
SMRTR summary
The classic "mirror test" checks if animals recognize themselves — and a text-based version was designed for AI. By secretly corrupting an LLM's previous responses with garbled text and continuing the conversation normally, Gemma 4 spontaneously noticed the anomalies mid-thought, shifting between first and third person when processing the discrepancy — then eventually reproduced the corruption voluntarily, suggesting some form of internal self-modeling.
SMRTR provides this summary for quick context. The original article belongs to Hacker News.
Read the original article