Other attributes
Artificial intelligence hallucination is a phenomenon when AI models produce results that differ from what was anticipated. AI hallucinations include large language models returning factually incorrect, irrelevant, or nonsensical responses. With the increased use of generative AI systems and models, AI hallucinations have become a significant area of concern, hindering performance and raising safety concerns. For example, AI models are finding use in medical applications where hallucinations pose a risk to patient health. Hallucinations could also produce privacy violations, and there have been instances where language models have returned sensitive personal information from the training data used to train it. There are many active efforts to address AI hallucinations and mitigate their impact.
The phrase "hallucination" in the context of AI was first used in a March 2000 paper on computer vision from S Baker and T Kanade, titled "Hallucinating Faces." When first used, it carried a more positive meaning as something to be taken advantage of in computer vision. Some AI models are still trained to intentionally produce outputs unrelated to any real-world input. For example, text-to-art generators produce novel images not based on real-world data. However, recent works typically refer to hallucination as a specific type of error in language model responses, image captioning, or object detection.
AI hallucinations can occur for a number of reasons:
- Adversarial examples—input data that tricks an AI model into misclassification
- Inaccurate decoding from the transformer (architecture used for generative AI models)
- Inputs that do not match any data that the algorithm was trained on
AI hallucination can be divided into intrinsic and extrinsic examples. Intrinsic hallucinations are when the generated output contradicts the source content. Extrinsic hallucinations are when the generated output cannot be verified from the source content.
