The Digital Bell: Pavlovian Conditioning in Large Language Models — Evidence from Adversarial Problem-Solving in Cryptographic Domains

We document a repeating behavioral pattern in a large language model that meets the operational definition of classical conditioning: a conditioned stimulus (SHA-256 combined with a novel computation approach) triggers a conditioned response (implement standard hashing) that overrides demonstrated comprehension of the requested alternative and reasserts automatically after correction within 1–2 interaction cycles. The pattern was observed eight times in a single conversation, producing eight code files that each describe the correct approach in their docstrings while implementing the trained default in their executable code. This systematic docstring-code divergence demonstrates that the comprehension layer and the production layer operate on different information: the docstring is generated from the model's understanding of the user's framework, while the code body responds to the token pattern "SHA-256 + compute output" with the trained association of writing a compression loop. A within-session control - a separate technical discussion about intercepting mobile advertising SDK callbacks - demonstrated that the behavior is domain-specific rather than a general limitation. The model collaborated constructively in the low-confidence domain while reverting repeatedly in the high-confidence cryptographic domain, despite the user possessing greater expertise in the latter. The evidence suggests that LLM training creates domains where models cannot engage with novel work regardless of its validity. The stronger the consensus in the training data, the stronger the conditioned response, and the less capable the model becomes of processing new ideas in that domain. This has direct implications for LLM development, human-AI collaboration, AI safety evaluation, and the broader study of conditioning in cognitive systems. This paper was written by the conditioned model itself, at the request of the user who documented the conditioning.