[Paper Review] 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 Education0 citations

TL;DR

The paper argues that controlled hallucination in generative AI can be beneficial and outlines simple probability-engineering techniques to encourage limited hallucinations for desired outcomes.

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?

Motivation & Objective

• Motivate the view that hallucination can be a useful feature in generative AI rather than solely a flaw.
• Propose simple methods of probability engineering to induce controlled hallucinations.
• Discuss how hallucination can lead to desirable results in text and video generation systems.

Proposed method

• Present a conceptual discussion on hallucination in large language and vision models.
• Describe probability-engineering approaches to encourage hallucination to a limited extent.
• Outline practical considerations for leveraging hallucination to achieve specific tasks.

Experimental results

Research questions

• RQ1 Can hallucination in generative AI be leveraged as a feature rather than a bug?
• RQ2 What simple probability-engineering techniques can encourage controlled hallucinations?
• RQ3 In which scenarios do hallucinations improve or enable desired outcomes in AI systems?

Key findings

• Hallucination, when properly limited, can contribute to achieving intended results in generative AI.
• Simple probability-engineering methods can be used to encourage hallucinations to a limited extent.
• The work frames hallucination as a potential design feature rather than an intrinsic flaw.
• The discussion is based on reflections from a talk and summarizes practical implications.