cs.CY提交于 2026-11-02已译

侵蚀真相默认值：基础模型幻觉与虚假信息在真实环境中对人类易感性的因果分析

Eroding the Truth-Default: A Causal Analysis of Human Susceptibility to Foundation Model Hallucinations and Disinformation in the Wild

Alexander Loth · Martin Kappes · Marc-Oliver Pahl

中文摘要

随着基础模型（foundation models, FMs）接近人类水平的流畅度，区分合成内容与真实内容已成为可信网络智能的关键挑战。本文提出 JudgeGPT 与 RogueGPT 双轴框架，将“真实性”与“归因”解耦，以研究人类易感性的机制。通过分析五种基础模型（包括 GPT-4 和 Llama-2）上的 918 次评估，采用结构因果模型（Structural Causal Models, SCMs）作为主要框架，构建关于检测准确率的可检验因果假设。与党派叙事相反，政治倾向与检测表现关联可忽略（r=-0.10）。相反，“假新闻熟悉度”成为候选中介变量（r=0.35），表明接触可能对人类辨别者起到对抗训练的作用。研究识别出“流畅性陷阱”：GPT-4 输出（HumanMachineScore: 0.20）绕过来源监控机制，使其与人类文本难以区分。这些发现表明，“预揭穿”干预应针对认知来源监控而非人口统计细分，以确保可信的信息生态系统。

关键要点

01提出 JudgeGPT 与 RogueGPT 双轴框架，解耦真实性与归因以研究人类易感性机制。
02分析五种基础模型（含 GPT-4、Llama-2）的 918 次评估，用结构因果模型检验检测准确率假设。
03政治倾向与检测表现关联可忽略（r=-0.10），假新闻熟悉度成为候选中介变量（r=0.35）。
04识别“流畅性陷阱”：GPT-4 输出（HumanMachineScore: 0.20）绕过来源监控，与人类文本难以区分。
05建议“预揭穿”干预针对认知来源监控而非人口统计细分，以保障可信信息生态。

解读

尚无解读。

原始英文摘要

arXiv:2601.22871v3 Announce Type: replace-cross Abstract: As foundation models (FMs) approach human-level fluency, distinguishing synthetic from organic content has become a key challenge for Trustworthy Web Intelligence. This paper presents JudgeGPT and RogueGPT, a dual-axis framework that decouples "authenticity" from "attribution" to investigate the mechanisms of human susceptibility. Analyzing 918 evaluations across five FMs (including GPT-4 and Llama-2), we employ Structural Causal Models (SCMs) as a principal framework for formulating testable causal hypotheses about detection accuracy. Contrary to partisan narratives, we find that political orientation shows a negligible association with detection performance ($r=-0.10$). Instead, "fake news familiarity" emerges as a candidate mediator ($r=0.35$), suggesting that exposure may function as adversarial training for human discriminators. We identify a "fluency trap" where GPT-4 outputs (HumanMachineScore: 0.20) bypass Source Monitoring mechanisms, rendering them indistinguishable from human text. These findings suggest that "pre-bunking" interventions should target cognitive source monitoring rather than demographic segmentation to ensure trustworthy information ecosystems.