When the San Francisco start-up OpenAI unveiled its ChatGPT online chatbot late last year, millions were impressed by its humanlike conversation skills and ability to discuss various topics. However, it soon became clear that this new type of chatbot often generates false information. Similar issues were observed with Google’s chatbot and Microsoft’s Bing chatbot. Vectara, a new start-up founded by former Google employees, is now researching the frequency of chatbot inaccuracies. They estimate that chatbots invent information at least 3 percent of the time, and possibly as high as 27 percent. This behavior, known as “hallucination,” poses serious concerns when dealing with sensitive data such as court documents or medical information. Vectara’s researchers discovered that hallucination rates differ among leading AI companies, with OpenAI’s technologies having the lowest rate at around 3 percent, and Palm chat, a Google system, having the highest rate at 27 percent. The researchers hope their work will stimulate industry-wide efforts to reduce the occurrence of hallucinations. The technology behind chatbots, called large language models (L.L.M.), learns from vast amounts of digital text but still struggles with accuracy due to the prevalence of false information and the probabilistic nature of language generation. While companies like OpenAI, Google, and Microsoft are working on improving the accuracy of their chatbots, researchers warn that eliminating hallucinations entirely is a challenging task.