2026/03/26 by J.-H. Park · 1 voice
Medicine · Social Sciences · #Artificial Intelligence in Healthcare and Education #Computational and Text Analysis Methods #Ethics and Social Impacts of AI
paper · doi:10.5281/zenodo.19229756
openalex publication_date 2026/03/26 · openalex created_date 2026/03/27 · openalex updated_date 2026/07/01
Why does AI lose its way in philosophy? This study begins with that question. Language models generate more tokens — and exhibit higher internal uncertainty — when processing brief philosophical monologues than when solving complex mathematical problems. This is not a matter of insufficient training. Philosophical concepts exist abundantly in model training data. This study proposes that the underlying cause lies in the absence of a consensus structure within human knowledge — that is, the presence or absence of a Convergence Point. A Convergence Point refers to whether a clear answer or logical endpoint exists within an utterance toward which an AI can converge. It is a variable independent of utterance length, complexity, or reasoning demand. Using four small language models (8B parameter scale) running in a local inference environment (LM Studio 0.4.6), we conducted first-utterance experiments across 13 categories and sequential utterance experiments across 12 categories. Utterances with clear convergence points (everyday narration, expert scientific statements) produced fewer tokens and remained stable. Utterances structurally lacking convergence points (philosophical monologues, unanswerable prompts) produced significantly more tokens and showed greater cumulative tok/sec decline. Notably, when a context-closing signal was added to philosophical utterances while preserving all philosophical keywords, token counts dropped by approximately 52% in Llama and 59% in Qwen3. This suggests that AI responds not to individual keywords but to the structural completeness of the utterance's context. Beyond behavioral observation, prompt-level entropy measurements conducted via llama-cpp-python revealed that philosophical utterances recorded the highest entropy across all categories (Llama: 1.6967, Mistral: 1.5169, DeepSeek: 1.6253). This serves as indirect evidence at the internal-state level, suggesting that the presence or absence of a convergence point is already reflected in the model's internal processing structure prior to output generation. Non-quantitative repeated observation across four large-scale models — GPT, Claude, Gemini, and Grok — confirmed the same pattern. Convergence Point phenomena may represent a structural characteristic independent of model scale and architecture. This study presents data suggesting that what AI finds difficult is not complex computation, but questions humanity has yet to resolve. This implies that the limitations of AI may stem not from within the model itself, but from the unresolved structure of human knowledge.