| 生成式人工智能幻觉的治理机制与福利效应研究 |
| Mechanism Design and Welfare Effects Analysis of Generative AI Hallucination Governance |
| 摘要点击 78 全文点击 0 投稿时间：2025-04-29 修订日期：2025-11-12 |
| 查看/发表评论 下载PDF阅读器 |
| 中文关键词 人工智能; 幻觉治理; 机制设计; 社会福利; 动力系统 |
| 英文关键词 Artificial Intelligence; Hallucination Governance; Mechanism Design; Social Welfare; Dynamic Systems |
| 基金项目 |
| 投稿方向 数字经济学与系统动力学 |
| 作者 | 单位 | 邮编 | | 陈晓佳 | 海南大学 | 570228 | | 徐玮* | 海南师范大学 | 571158 | | 邹文文 | 广州大学 | |
|
| 中文摘要 |
| 针对生成式人工智能在提升内容生产效率的同时引发输出不真实信息(幻觉)的问题, 提出基于动态反馈机制的幻觉调节分析方法. 通过构建包含企业信息生成、用户信任演化和监管政策三要素的动态反馈模型, 将幻觉率作为可调参数纳入结构分析. 结果显示: 幻觉率与社会福利之间呈倒U型关系, 中等水平的幻觉有助于信息多样性与效率提升, 过度压制幻觉率则削弱整体福利; 在缺乏有效监管的情境下, 市场容易陷入“高幻觉—低信任—高不确定性”的恶性循环; 各行业对幻觉的容忍度存在显著差异, 统一且过严监管会导致优质企业退出并造成逆向选择. 结论指出, 应针对不同领域实施弹性化监管策略, 以优化人工智能内容治理, 为政策制定提供科学参考. |
| 英文摘要 |
| To address the challenge of generative AI producing unrealistic information (Hallucinations) while improving content efficiency, this paper develops a hallucination regulation approach based on a dynamic feedback model integrating enterprise information generation, user trust evolution, and regulatory policies. The hallucination rate is treated as an adjustable parameter within the system. Results reveal an inverted U-shaped relationship between hallucination rate and social welfare: moderate hallucination enhances information diversity and efficiency, whereas excessive suppression lowers welfare. Without effective regulation, markets may fall into a “high hallucination–low trust–high uncertainty” trap. Industry-specific tolerance differences imply that uniform, stringent regulation can trigger adverse selection. The study advocates flexible, sector-specific governance strategies to optimize AI content regulation and support evidence-based policymaking. |
| 关闭 |
|
|
|
|
|
