Scholarlink

<p>This article offers a critical and philosophically grounded reassessment of recent attempts to apply Harry Frankfurt’s concept of “bullshit” to large language models (LLMs). While Hicks et al. (ChatGPT is Bullshit. AI and Society, forthcoming, 2024) argue that ChatGPT exemplifies epistemic insincerity—discourse indifferent to truth—we contend that this attribution commits a category error. Frankfurt’s framework presupposes epistemic agency, which LLMs lack. However, if one accepts the extension of this framework to non-agentive systems, we argue that it is DeepSeek—not ChatGPT—that more closely simulates the rhetorical and epistemic features associated with bullshit. DeepSeek, characterized by fluent surface performance, benchmark-driven optimization, and assertive style, often masks brittle or unexamined reasoning beneath rhetorical polish. In contrast, GPT-4 (as deployed through ChatGPT) more frequently exhibits hedging, contextual modulation, and dialogic responsiveness—behavioral proxies that, while not indicative of epistemic commitment, reduce the appearance of insincerity. We do not offer an uncritical defense of generative AI systems. Rather, we caution against mistaking fluency for epistemic integrity. Through an integrated philosophical, empirical, and architectural analysis, we propose a conditional reinterpretation of Frankfurt’s framework and show that DeepSeek exemplifies, not contradicts, the challenges posed by epistemic simulation in LLMs.</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

The deep illusion: a critical analysis of DeepSeek and the limits of large language models

  • 期刊论文

  • 发表日期：

    2025年8月6日

查看全文
  • Andrei Polozov

摘要

This article offers a critical and philosophically grounded reassessment of recent attempts to apply Harry Frankfurt’s concept of “bullshit” to large language models (LLMs). While Hicks et al. (ChatGPT is Bullshit. AI and Society, forthcoming, 2024) argue that ChatGPT exemplifies epistemic insincerity—discourse indifferent to truth—we contend that this attribution commits a category error. Frankfurt’s framework presupposes epistemic agency, which LLMs lack. However, if one accepts the extension of this framework to non-agentive systems, we argue that it is DeepSeek—not ChatGPT—that more closely simulates the rhetorical and epistemic features associated with bullshit. DeepSeek, characterized by fluent surface performance, benchmark-driven optimization, and assertive style, often masks brittle or unexamined reasoning beneath rhetorical polish. In contrast, GPT-4 (as deployed through ChatGPT) more frequently exhibits hedging, contextual modulation, and dialogic responsiveness—behavioral proxies that, while not indicative of epistemic commitment, reduce the appearance of insincerity. We do not offer an uncritical defense of generative AI systems. Rather, we caution against mistaking fluency for epistemic integrity. Through an integrated philosophical, empirical, and architectural analysis, we propose a conditional reinterpretation of Frankfurt’s framework and show that DeepSeek exemplifies, not contradicts, the challenges posed by epistemic simulation in LLMs.