OpenAI Paper: Hallucinations apparently unavoidable
Large language models and AI chatbots that don't hallucinate? Even OpenAI thinks that's impossible. But there is a way out.
(Image: Novikov Aleksey/Shutterstock.com)
A new scientific paper published by OpenAI deals with hallucinations. This is false information and context that large language models (LLMs) and, therefore, also AI chatbots output. All AI companies are working to keep hallucinations to a minimum. Eliminating them completely, however, seems impossible. OpenAI itself now also writes this.
The paper is summarized by OpenAI in a blog post. It states that hallucinations cannot be avoided. “Spelling and parentheses follow consistent patterns, so errors in this area disappear with increasing size.” But arbitrary, rarely occurring.
Seldom occurring facts cannot be predicted based on patterns and thus lead to hallucinations. As an example, OpenAI cites birthdays, which cannot be predicted and are not recognizable by anything. They are random. AI cannot learn from coincidences.
AI models want to be rewarded
Not only that, AI models also strive to give an answer. The authors compare this to a multiple-choice test. If you select an answer even though you have no idea what the correct answer is, the chance of being right increases. This is also how LLMs act—they would rather give any answer than none.
Videos by heise
One way to reduce hallucinations is to evaluate wrong answers negatively, the authors suggest. Answers that are not given, on the other hand, could at least receive partial points, so to speak, to put an end to this blind guessing. Random hits should also no longer be rewarded. However, this is the basis of the models—they learn through reward. Reasoning processes, in which the models justify their answers, at least reduce hallucinations. However, OpenAI does not believe that they will disappear completely in the near future. They are working on reducing them further.
“Accuracy will never reach 100 percent because, regardless of model size search and reasoning capabilities, some real-world questions are inherently unanswerable.” But language models should be able to hold back: no answers instead of hallucinations are possible.
(emw)
