Scholarlink

<p>Large Language Models (LLMs) suffer from hallucinations—probabilistic fabrications that erode epistemic trust. These are not technical bugs but emergent illusions inherent to their non-agentive ontology. This paper argues that statistical models are not knowers but instruments; treating them as autonomous epistemic agents reflects a misalignment in role assignment, rather than a rejection of their instrumental epistemic value. We propose a novel governance framework from East Asian philosophy to address the crisis of <i>synthetic plausibility</i>. This framework synthesizes Buddhist ontology—specifically the Madhyamaka understanding of dependent origination (<i>pratītyasamutpāda</i>) and the Two Truths doctrine—with Confucian ethics to deconstruct LLM outputs as <i>śūnya</i>: empty of intrinsic essence (<i>svabhāva</i>) yet conventionally functional. Devoid of self or intent, their outputs are relational fictions, analyzable via the Two Truths doctrine to distinguish conventional coherence (<i>saṃvṛti</i>) from ultimate ungroundedness (<i>paramārtha</i>). This ontological shift is balanced by Confucian ethics, invoking <i>zhèngmíng</i> (rectification of names) to correct misassignments—such as labeling LLMs “experts”—that harm user-tool relations and social harmony (<i>lǐ</i>). Resolving the paradox of expecting deterministic truth from probabilistic systems, our framework moves governance from internal accuracy to interface labeling, using metadata signals and standards like provenance watermarking, complementing Model Cards and C2PA. This operationalizes trust via UI disclosures, reducing over-reliance by aligning expectations with probabilistic realities. By accepting illusion as the machine’s natural state while governing its naming, we build resilient epistemic ecosystems and a meta-epistemology where trust arises from calibrated expectations. This cross-cultural approach critiques anthropomorphic overreach and provides actionable protocols for global AI policy, ensuring synthetic tools support rather than undermine human judgment.</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

Rectifying illusion: a Buddhist–Confucian framework for LLM hallucinations

  • 期刊论文

  • 发表日期：

    2026年3月8日

查看全文
  • Sridharan Sankaran

摘要

Large Language Models (LLMs) suffer from hallucinations—probabilistic fabrications that erode epistemic trust. These are not technical bugs but emergent illusions inherent to their non-agentive ontology. This paper argues that statistical models are not knowers but instruments; treating them as autonomous epistemic agents reflects a misalignment in role assignment, rather than a rejection of their instrumental epistemic value. We propose a novel governance framework from East Asian philosophy to address the crisis of synthetic plausibility. This framework synthesizes Buddhist ontology—specifically the Madhyamaka understanding of dependent origination (pratītyasamutpāda) and the Two Truths doctrine—with Confucian ethics to deconstruct LLM outputs as śūnya: empty of intrinsic essence (svabhāva) yet conventionally functional. Devoid of self or intent, their outputs are relational fictions, analyzable via the Two Truths doctrine to distinguish conventional coherence (saṃvṛti) from ultimate ungroundedness (paramārtha). This ontological shift is balanced by Confucian ethics, invoking zhèngmíng (rectification of names) to correct misassignments—such as labeling LLMs “experts”—that harm user-tool relations and social harmony (lǐ). Resolving the paradox of expecting deterministic truth from probabilistic systems, our framework moves governance from internal accuracy to interface labeling, using metadata signals and standards like provenance watermarking, complementing Model Cards and C2PA. This operationalizes trust via UI disclosures, reducing over-reliance by aligning expectations with probabilistic realities. By accepting illusion as the machine’s natural state while governing its naming, we build resilient epistemic ecosystems and a meta-epistemology where trust arises from calibrated expectations. This cross-cultural approach critiques anthropomorphic overreach and provides actionable protocols for global AI policy, ensuring synthetic tools support rather than undermine human judgment.