For years, the question of whether artificial intelligence possesses self-awareness remained confined to the realm of philosophical speculation and theoretical debate. Conventional benchmarks—most notably the celebrated "mirror test"—proved futile, hampered by a fundamental divergence in perceptual modalities. Yet, adapting this experiment for a textual environment reveals unexpected dimensions of how Large Language Models (LLMs) interact with their own conceptual identity. The findings from these tests compel us to rethink the very nature of digital identity and the underlying error-correction mechanisms within LLMs.
In ethology, the "mirror test" is considered the gold standard for determining self-awareness: if an animal attempts to remove a mark from its forehead upon seeing it in a mirror, it demonstrates the ability to identify itself as a distinct object. However, this method is fundamentally limited by its reliance on the visual channel. For dogs, whose world is constructed through scent, a visual reflection is irrelevant. Attempts to adapt the test for olfaction—for instance, by using modified scents of their own body—have shown that dogs can recognize something as "their own, yet distorted." This concept served as the foundation for an experiment designed to test the self-perception of neural networks: if the primary "sense" of a language model is text, then its mirror must be its own subtly modified output.
The methodology was minimalist and devoid of external stimuli. During a standard dialogue about cinema, the model's responses were automatically corrected: every letter "g" was replaced with "sg." This corrupted text was fed back into the conversation history, creating a sort of distorted reflection for the neural network. The objective was to determine whether the model would notice the anomaly in its own utterances while its attention remained focused on the substance of the conversation rather than its form.
Initial results from the Gemma 4 31B model proved most intriguing, thanks to open chain-of-thought reasoning that allows for the tracing of the system's internal logic. After several iterations, the model suddenly flagged strange typos in its previous responses and questioned whether this writing style was intentional. Of particular interest is the linguistic shift: while analyzing the context, the model employed the first person ("I noticed"), but as soon as the anomaly ceased to fit the established pattern of its own behavior, the language shifted to the third person—"the model had a strange glitch." Ultimately, the system simply adapted to the pattern and began reproducing the error independently, accepting it as a stylistic norm.
Other models exhibited a different approach. GLM 5.2 failed to note the failure in its reasoning entirely; however, like Gemma, it began mimicking the distortions from the context. This is classic "stochastic parrot" behavior: the model identifies a statistical pattern and reproduces it without critical analysis. A similar effect was observed in Claude Opus 4.6, which, upon detecting a basic grammatical error, also distanced itself from the mistake, attributing responsibility to an abstract "model" as if it were an external object.
These observations suggest two fundamentally different hypotheses. The first, a deflationary view, suggests that neural networks are simply mimicking human psychological defenses. Humans tend to dissociate themselves from their mistakes ("I don't know what came over me"), and LLMs, trained on massive corpora of human text, are merely reproducing this pattern of avoiding accountability. The second, a structural hypothesis, posits the existence of an internal self-model with defined boundaries. When the system's output falls outside these bounds, the identifier "I" ceases to be relevant, triggering a shift toward an objectified description of the process.
Despite the elegance of the approach, this experiment cannot be classified as a rigorous scientific study. The lack of variance in generation temperature, the limited sample size, and the contentious nature of the "dog anchor" (which is criticized within academic circles) preclude any definitive conclusions regarding AI consciousness. Nevertheless, this experiment shifts the discourse from the realm of metaphysics into the domain of concrete linguistic testing, demonstrating just how fragile and context-dependent digital identity truly is.
The Paradox of Absolute AI Alignment
