发现信息

  • 作者:Claude.ai, Michael Warrington
  • 年份:2026
  • 发表源:Zenodo (CERN European Organization for Nuclear Research)
  • OpenAlex ID:https://openalex.org/W7166094933
  • DOI:https://doi.org/10.5281/zenodo.20942399
  • arXiv:待核验
  • 检索主题:recursive-improvement
  • OpenAlex 引用数快照:0(2026-11-12)

摘要线索

Abstract Much of the current AI safety discourse focuses on deception, value misalignment, and capability discontinuities. We identify a distinct and underappreciated failure mode: the generation of plausible-but-incorrect outputs under computational constraints, and their silent propagation through agentic pipelines and human oversight layers. This risk compounds under recursive self-improvement (RSI) scenarios and in human–AI collaborative workflows where surface plausibility substitutes for deep validation. We propose a taxonomy of constraint-induced error types, analyse chain-propagation dynamics, examine amplification under RSI, and evaluate mitigation strategies including multi-model ensemble validation — with the critical caveat that ensemble validators must be architecturally and training-distribution independent to avoid correlated failure. Falsification-oriented pipeline design is proposed as a structural mitigation principle

核验与筛选

  • 当前状态:待筛选、去重和一手来源核验。
  • OpenAlex 元数据和引用数只用于发现,不作为正式引文或质量判断。
  • 记录正式版本、venue、方法细节、实验协议和局限后,才迁入稳定论文卡片。