tool
Artificial intelligence model actually produces non-existent information.
Halusination is the condition that a model of artificial intelligence - especially a large language model - an information that is not actually present, incorrect or source, completely self-certainty and fluently, as if it is corrected. This behavior is caused by the basic working principle, not "worn" of the model: a language model is actually not a reality database, but it is a system that estimates the next most likely token based on statistical words and concept images in the educational data. Since the model is under the pressure of producing a fluent response in grammar and üslup even in a matter that it does not know, it can produce a possible but fitting answer instead of actual information; for example, it can skip to an unexistent academic article or give a wrong date.
Halusination is one of the most serious obstacles in front of the use of reliable in high-risk areas such as medicine, law and finance. Several approaches are used to reduce the problem: based on the actual documents of the model with RAG, verifying the output of the model in real-time with external sources, learning to consolidate with human feedback (RLHF) to encourage the model to be honest in the case of uncertainty and use more clean, accurate training data. Since no technical hallucination is completely eliminated, users are often advised to verify model outputs with independent sources in critical decisions.
Halusination is one of the most serious obstacles in front of the use of reliable in high-risk areas such as medicine, law and finance. Several approaches are used to reduce the problem: based on the actual documents of the model with RAG, verifying the output of the model in real-time with external sources, learning to consolidate with human feedback (RLHF) to encourage the model to be honest in the case of uncertainty and use more clean, accurate training data. Since no technical hallucination is completely eliminated, users are often advised to verify model outputs with independent sources in critical decisions.