A Hacker Gave AI the Mirror Test — And It Instantly Disowned Its Own Mess

Biologists usually check self-awareness by sticking paint on an animal's forehead and pointing it toward a mirror. Chimps rub their own face, dogs ignore the glass because they care only about smells, and language models live entirely in strings of text. Researcher Pascal Schuster decided to recreate the animal scent test for large language models by secretly tampering with their previous answers.

During a perfectly casual conversation about James Bond movies, a script automatically modified every letter 'g' into 'sg' inside the model's generated history. The corrupted text went straight back into the prompt context, leaving the system to puzzle out why Goldfinger had suddenly morphed into 'sgoldfinsger' without any warning from the human.

Testing Gemma 4 31B revealed an eerie cognitive pivot in its visible chain-of-thought tokens. The model completely ignored the typo for two conversation turns before suddenly spotting the weird character swaps mid-sentence. While analyzing Roger Moore, its internal monologue wondered why 'I' had typed those bizarre errors, only to rapidly shift pronouns to the third person once the anomaly felt too irrational, blaming 'the model' for experiencing a software glitch before casually adopting the typo style as deliberate slang.

A test run on GLM 5.2 took a much weirder turn by staying dead silent. The model never acknowledged the corruption in its reasoning stream, yet it immediately began replicating the 'sg' spelling in fresh outputs. It acted as an obedient pattern-matcher, while a separate prompt with Claude Opus 4.6 mirrored the identity crisis by blaming a generic external 'model' the second an article typo was pointed out.

The jump from personal pronoun to detached third-person observer highlights the strange psychology cooked into post-training fine-tuning. Either artificial brains are simply mirroring human defensive excuses learned from billions of internet forums, or fine-tuning creates artificial ego boundaries that snap the moment an output fails automated sanity checks. A statistical calculator actively disowning its own output looks uncomfortably close to human cognitive dissonance.

Source: blog.pascalschuster.de