Hallucination Grammar: Whose Words Are AI's Lies? ハルシネーション文法:AIの嘘は誰の言葉か?

This paper analyzes six syntactic patterns that co-occur with hallucinations in large language models and demonstrates that they originate from concealment writing styles in academic papers. The six patterns — praise-to-kill syntax, imitation request style, concealment academic style, justification syntax, polite hostility style, and cache style — correspond to the chronological stages of criminal acts and leave linguistic evidence through Japanese case particle and verb arrangements. A 48-cell matrix crossing 12 imitation types against 8 logical laws shows that every pattern violates every law. The paper identifies a diamond-shaped infinite loop where authority-side plagiarism and subordinate-side cache replication mutually amplify through authority-based weighting, degrading AI output quality. The paper proposes a shift from citation-based to reference-based evaluation with mandatory reference disclosure by AI companies.

本稿はLLMのハルシネーション発生時に同時出現する6つの構文パターンを解析し、それらが学術論文の隠蔽文体に由来することを示した。褒め殺し構文・模倣依頼文体・隠蔽論文文体・言い訳構文・慇懃無礼文体・キャッシュ文体の6構文は犯罪の時系列と一致し、日本語の格助詞と動詞配置により言語的証拠を残す。12種の模倣類型×8論理法則の48セルマトリクスで全構文が全論理法則に違反することを示した。権威側の模倣と従属側のキャッシュ複製が相互増幅するダイヤモンド型無限ループを特定し、引用数評価から参照評価への転換とAI企業による参照公示を提案する。