Scholarlink

<p>The proliferation of autonomous AI systems has created a critical “responsibility gap”, where the opacity of decision-making processes makes accountability elusive. Prevailing alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF) and Constitutional AI (CAI), address ethical behavior through training and principle-adherence but fail to produce a verifiable, contemporaneous record of moral reasoning. This paper introduces Ternary Moral Logic (TML), a novel system architecture that converts AI ethical deliberation from an abstract process into a cryptographically secured, evidentiary record. TML implements a third logical state—the Sacred Pause (0)—alongside conventional permit ( + 1) and prohibit (−1) states, which is triggered when a model encounters moral complexity. This pause initiates a non-blocking, parallel process that generates a Moral Trace Log via a mandatory Always Memory component. These logs are rendered immutable through a Hybrid Shield that uses multi-chain blockchain anchoring. We present qualitative findings indicating TML's potential to reduce harmful outputs while maintaining performance, and we argue that its architecture provides the technical substrate necessary to meet the traceability and record-keeping mandates of emerging regulations like the EU AI Act. By transforming moral hesitation into verifiable forensic data, TML establishes a new discipline: Ethical Forensics for AI Systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Auditable AI: tracing the ethical history of a model

  • 期刊论文

  • 发表日期：

    2025年12月28日

查看全文
  • Lev Goukassian

摘要

The proliferation of autonomous AI systems has created a critical “responsibility gap”, where the opacity of decision-making processes makes accountability elusive. Prevailing alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF) and Constitutional AI (CAI), address ethical behavior through training and principle-adherence but fail to produce a verifiable, contemporaneous record of moral reasoning. This paper introduces Ternary Moral Logic (TML), a novel system architecture that converts AI ethical deliberation from an abstract process into a cryptographically secured, evidentiary record. TML implements a third logical state—the Sacred Pause (0)—alongside conventional permit ( + 1) and prohibit (−1) states, which is triggered when a model encounters moral complexity. This pause initiates a non-blocking, parallel process that generates a Moral Trace Log via a mandatory Always Memory component. These logs are rendered immutable through a Hybrid Shield that uses multi-chain blockchain anchoring. We present qualitative findings indicating TML's potential to reduce harmful outputs while maintaining performance, and we argue that its architecture provides the technical substrate necessary to meet the traceability and record-keeping mandates of emerging regulations like the EU AI Act. By transforming moral hesitation into verifiable forensic data, TML establishes a new discipline: Ethical Forensics for AI Systems.