[论文解读] Engineering of Hallucination in Generative AI: It's not a Bug, it's a Feature

Tim Fingscheidt, Patrick Blumenberg+1 more|arXiv (Cornell University)|Jan 11, 2026

Artificial Intelligence in Healthcare and Education被引用 0

一句话总结

论文主张在生成式人工智能中受控的幻觉有益，并概述了简单的概率工程技术以鼓励对期望结果的有限幻觉。

ABSTRACT

Generative artificial intelligence (AI) is conquering our lives at lightning speed. Large language models such as ChatGPT answer our questions or write texts for us, large computer vision models such as GAIA-1 generate videos on the basis of text descriptions or continue prompted videos. These neural network models are trained using large amounts of text or video data, strictly according to the real data employed in training. However, there is a surprising observation: When we use these models, they only function satisfactorily when they are allowed a certain degree of fantasy (hallucination). While hallucination usually has a negative connotation in generative AI - after all, ChatGPT is expected to give a fact-based answer! - this article recapitulates some simple means of probability engineering that can be used to encourage generative AI to hallucinate to a limited extent and thus lead to the desired results. We have to ask ourselves: Is hallucination in gen-erative AI probably not a bug, but rather a feature?

研究动机与目标

• 使人们相信在生成式AI中幻觉可以成为有用的特征，而不仅仅是缺陷的观点被证明或得到支持。
• 提出简单的概率工程方法以诱导受控的幻觉。
• 讨论幻觉如何在文本和视频生成系统中带来期望的结果。

提出的方法

• 对大型语言模型和视觉模型中的幻觉进行概念性讨论。
• 描述概率工程方法以在有限程度上鼓励幻觉。
• 概述利用幻觉实现特定任务的实际考虑因素。

实验结果

研究问题

• RQ1 生成式AI中的幻觉是否可以被作为特征而非错误来利用？
• RQ2 哪些简单的概率工程技术可以鼓励受控的幻觉？
• RQ3 在何种情景下幻觉会改善或促成AI系统中的期望结果？

主要发现

• 在适度受限的情况下，幻觉可以帮助实现生成式AI的预期结果。
• 可以使用简单的概率工程方法在有限程度上鼓励幻觉。
• 本工作将幻觉框定为潜在的设计特征而非固有的缺陷。
• 该讨论基于对一个报告的反思并总结了实际影响。