Systematically Reducing Hallucination in Generative AI from Users to Data to Models

Abstract

AI hallucination is not just about a one-off technical glitch at the model level but is a systemic issue with the interplay between humans, the data, and the model itself. Most existing reviews focus on isolated parts of model development, like model architecture or data governance, and do not consider the whole model development chain from users to data to the model as an integrated whole and analyze it accordingly. To systematically review advances in research addressing hallucinatory outputs in generative Artificial Intelligence, this paper takes a tripartite approach covering users, data, and models. From the user perspective, from low prompting capabilities, cognitive biases to information discernment, which result in hallucinations, and then collectively devise prompt engineering, interactive clarification, and information literacy enhancement as three countermeasures to reduce the effects of hallucinations. From the data perspective, it explores how data quality, retrieval-augmented generation, and knowledge graphs can help strengthen the factual consistency of data and outline their intrinsic weaknesses. In terms of the model, it outline factuality-oriented decoding strategies, alignment training paradigms, and self-inspection with corrective refinement techniques. This paper clarifies the above three aspects and how they work together to form a closed-loop hallucination cycle, starting with user feedback as a departure point, moving to the next step of turning error-correction signals into training instances at the data tier, and lastly to internalising the ability of veracity discrimination in the model parameters via preference alignment and knowledge rectification at the model tier.