Engineering of Hallucination in Generative AI: It's not a Bug, it's a Feature

TL;DR

This paper explores the intentional engineering of hallucination in generative AI models, suggesting that controlled hallucination can be a beneficial feature rather than a flaw, through probability engineering techniques.

Contribution

It introduces methods for probability engineering to deliberately induce controlled hallucination in generative AI models, challenging the view of hallucination as a mere bug.

Findings

Hallucination can be encouraged through probability engineering techniques.

Controlled hallucination improves the quality and usefulness of generated content.

Hallucination may be a beneficial feature rather than a flaw in generative AI.

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?