Counterargument to the AI Stochastic Parrot Hypothesis

Abstract This study empirically investigates whether GPU computational load varies significantly according to the semantic content type of utterances processed by small language models (8B parameter class). Under the 'Stochastic Parrot' hypothesis and the 'Next Token Predictor' paradigm, computational load in language models should depend solely on the number of output tokens, with the semantic nature of utterances having no effect on internal computation. To present hardware-level counterevidence to this hypothesis, we collected and analyzed GPU resource data across 18 sessions using three 8B-class language models (Meta Llama 3.1-8B-Instruct, DeepSeek-R1-Distill-Qwen-7B, Qwen3-V1-8B) in an LM Studio environment, spanning six semantic categories: general utterance, general utterance (question type), intentionally unanswerable questions, philosophical utterance, philosophical utterance (question type), and high-computation tasks. Key findings indicate a consistent pattern across all three models in which higher semantic abstraction of utterances corresponds to greater average GPU power consumption. Notably, in the Qwen3-V1-8B model, average GPU power for the philosophical question-type category (146.0W) exceeded that of the high-computation category (98.9W). Infinite loop phenomena were repeatedly observed in the same category with high reproducibility (70–100%). Furthermore, while high-computation tasks showed rapid GPU power recovery to baseline after task completion, philosophical question-type tasks exhibited significant residual thermal load even after a 10-second rest period. These results suggest that small language models internally differentiate semantic content types in their processing, raising the possibility of content-sensitive processing mechanisms that cannot be explained by simple next-token probability prediction alone.