Developers have tricks to stop artificial intelligence from making things up, but large language models are still struggling to tell the truth, the whole truth and nothing but the truth.
(21 Jan 2025) In one 2024 study, various chatbots made mistakes between about 30% and 90% of the time on references, getting at least two of the paper’s title, first author or year of publication wrong1. Chatbots come with warning labels telling users to double-check anything important. But if chatbot responses are taken at face value, their hallucinations can lead to serious problems, as in the 2023 case of a US lawyer, Steven Schwartz, who cited non-existent legal cases in a court filing after using ChatGPT.
Chatbots err for many reasons, but computer scientists tend to refer to all such blips as hallucinations. It’s a term not universally accepted, with some suggesting ‘confabulations’ or, more simply, ‘bullshit’2. The phenomenon has captured so much attention that the website Dictionary.com picked ‘hallucinate’ as its word of the year for 2023.
Because AI hallucinations are fundamental to how LLMs work, researchers say that eliminating them completely is impossible3. But scientists such as Zou are working on ways to make hallucinations less frequent and less problematic, developing a toolbox of tricks including external fact-checking, internal self-reflection or even, in Zou’s case, conducting “brain scans” of an LLM’s artificial neurons to reveal patterns of deception.
Nature has more details here.