“Survey of Hallucination in Natural Language Generation” was submitted to arXiv on February 8, 2022 by Ziwei Ji and a large team led by Pascale Fung at HKUST, and published in ACM Computing Surveys. It was one of the first comprehensive reviews of hallucination - text that is fluent and confident but unfaithful to the source or simply false - and became a heavily cited reference as the problem moved to the center of attention with large language models.
The survey treats hallucination as the generation of unintended content not grounded in the input or in fact, and it draws a now-common distinction between intrinsic hallucination (output that contradicts the source) and extrinsic hallucination (output that cannot be verified against the source at all). It then walks through how the problem appears, is measured, and is mitigated across a range of tasks: abstractive summarization, dialogue, generative question answering, data-to-text generation, machine translation, and vision-language generation.
Beyond cataloging, the paper organizes the causes - noisy or divergent training data, exposure bias from teacher-forced training, decoding strategies that favor fluency - and surveys metrics and mitigation methods, from better data and faithfulness-aware training objectives to post-hoc fact-checking. By giving the field a shared vocabulary and taxonomy, it helped frame the research agenda that later work on retrieval grounding and factuality evaluation built on.
Why business readers should care: hallucination is the single biggest obstacle to deploying generative AI in high-stakes settings, and this survey is the standard map of why it happens and what is known to reduce it.