In the 21st century hallucinations have become a daily experience. The origins of the word can be followed back at least to the Latin verb “alucinor”, best translated with “to hallucinate”. As a verb to can conjugate it, meaning that I can do it, you can do it, s/he can do it, and we may do it in groups. Roman emperors did it, American presidents do it and, of course, AI does it. Hence, it is a great subject to study.
In “Nature” 2025 we find ways to limit hallucinations of AI systems. The strategy consists mainly in repeated queries of the same type, but from different angles. It is a bit like cubism applied to informatics. On “github.com” we can follow the rankings of AI-models using LLMs based on the “hallucination-leaderboard” developed by Vectara. On “huggingface.com” you can test the Hughes Hallucination Evaluation Model. For example it is possible to run a test of your own small text documents (just like any blog entry on this webpage) and what the AI systems will do them in an attempt to summarize your ideas. According to the “hallucination-leaderboard” we are confronted with a 1.3%-4% hallucination rate of the top 25 LLMs as AI-systems. In text based systems the quantity of “errors” is a first indicator only. The seriousness of the omission, addition of wrong information or an erroneous judgment will be left to the reader or analyst to uncover.
There is now a lot to do to test various AI-systems on their “trustworthiness” in summarizing my own work. My very own daily hallucinations have become a large data base as a test case for the capacity of LLMs to make sense of them.
Based on the series of passed blog entries I shall test the capacity of AI to predict the n+1 blog entry. It would be great to know today what I am going to write about tomorrow etc. Thanks to AI I shall have (finally) a sort of intellectual life after death (not sure whether I should want this). Enough of hallucinations and on hallucinations for now, back to serious readings or fictionalized science. (Image: extract from Delphine Diallo, Kush, 2024 at Hangar Gallery Brussels).