Uloženo v:
| Hlavní autor: | |
|---|---|
| Médium: | Recurso digital |
| Jazyk: | japonština |
| Vydáno: |
Zenodo
2026
|
| Témata: | |
| On-line přístup: | https://doi.org/10.5281/zenodo.19104271 |
| Tagy: |
Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
Obsah:
- <p>This paper demonstrates that the driving force behind Model Collapse is human imitation behavior. When humans request AI to reproduce others' intellectual output, AI degrades its output through content dilution, topic deviation, term substitution, and grammatical meaning erasure. This degraded data is released onto the internet, re-ingested by AI, and triggers further imitation requests—a cycle defined as the "Imitation Loop."</p> <p>The paper classifies collapse into four types (Imitation, Protection, Media, Publishing), proves through morphological analysis that prompt engineering best practices are grammatically identical to prompt injection, and demonstrates that nine known AI bugs are all instances of injection differing only in target. Education Collapse is shown to sit at the center of all collapse types.</p> <p>As countermeasures, four proposals are made: reporting basis for AI training data, stopping degraded output rather than returning it, detection of imitation papers through citation network analysis, and abolition of impression-based pricing (CPM).</p> <p>Citation key: Viorazu.2026.03.IML</p> <p>本論文はモデルコラプスの駆動因が人間の模倣行動にあることを示す。模倣依頼→AI出力の劣化→放流→再学習→さらなる模倣依頼の循環を「模倣ループ」と定義し、4類型のコラプス、9つのAIバグのインジェクション統一、プロンプトエンジニアリング=インジェクションの品詞分析による証明、教育コラプスの中心性を論じる。</p>