Do LLMs pass the mirror test?

Current LLM mirror tests ask models to identify their own outputs in a lineup, but this argument holds that tests should instead probe anomaly detection against an internal baseline — analogous to Horowitz's olfactory dog test. The proposed analog is presenting a model with subtly altered versions of its own prior responses and measuring whether it flags the discrepancy as "mine, but wrong." Whether that constitutes self-awareness is left open; the practical claim is that existing text-based mirror tests are poorly designed instruments, not that models lack self-recognition.