Traditional mirror tests for LLMs miss the mark by translating visual recognition into text-based identification. Researchers typically show models their outputs and ask if they recognize them, but this approach fundamentally misunderstands what self-awareness testing should measure. The real issue mirrors criticism of dog mirror testing - using the wrong sensory modality for the subject.

Alexandra Horowitz solved this for dogs by testing scent recognition instead of visual reflection. She presented dogs with their own scent modified with aniseed oil, triggering investigation behavior that suggested anomaly detection against an internal baseline. This insight inspired a new approach for LLMs: modifying their own textual output and observing whether models notice the discrepancy during normal conversation.

Using Gemma 4 31B on Google AI Studio, the author corrupted model responses by replacing 'g' with 'sg' throughout the text. For two conversational turns, Gemma processed the garbled output without comment. Then midway through planning its third response, the model flagged the anomaly in its thinking trace, shifting from first-person ('I noticed') to third-person ('the model had a strange quirk') perspective.

This spontaneous detection suggests LLMs may possess some form of self-monitoring capability, though whether this constitutes true self-awareness remains debatable. The dissociation between the model's 'I' and 'the model' reveals interesting insights about how these systems process anomalies in their own output streams.