AI Illusions Might Not Be So Bad

2023-09-15

"ChatGPT claims that 'wine can prevent cancer'," said Bard. He believes that "Hydrobottlecatputalization" (a term coined by the author) will completely change transportation. In a two-hour conversation, Bing confessed to New York Times writer Kevin Roose. These statistically enhanced AI systems can write code to order pizza in any language, but sometimes they fabricate things. These overconfident models cannot distinguish between what is correct and what appears to be correct. Kevin Scott, Chief Technology Officer of Microsoft, says this is part of the learning process. He said, "The more you try to lead it down the path of illusion, the further it gets from reality." Similarly, this topic has long plagued AI researchers. While some have claimed that there is no solution to the illusion problem, others are still striving to find ways to prevent chatbots from going off track. The term "hallucination," coined in the 17th century, caught the attention of computer scientists in 2015 when Andrej Karpathy of OpenAI wrote a blog post on how AI systems generate "hallucinations," such as fabricating seemingly reasonable URLs and mathematical proofs. This term was adopted by Google researchers in a conference paper published in 2018, titled "Hallucinations in Neural Machine Translation," which analyzed how automatic translation produces outputs unrelated to the input. Although the illusion problem is primarily related to language models, it also affects audio and visual models. Three researchers from the University of South Carolina's Artificial Intelligence Research Institute thoroughly investigated these foundational models to identify, clarify, and address the illusion problem. Their research established criteria for determining the frequency of hallucinations and also examined current methods used to reduce these issues in the models, discussing future research on how to solve this problem. Creative Alternatives While most researchers in the field have grown tired of the lies from these models, some propose an alternative philosophy. They believe that the tendency of these models to "invent" facts may not necessarily be a disadvantage. Sebastian Berns, a doctoral researcher at Queen Mary University in London, shares this view. He believes that models prone to generating illusions could become valuable "co-creative partners." For example, by increasing the temperature of ChatGPT, the model produces imaginative narratives rather than fact-based answers. According to Berns, these models may generate outputs that are not entirely accurate but contain useful ideas. Leveraging the creativity of illusions can lead to unexpected results or combinations of ideas that may not naturally occur to most people. Berns also emphasizes that when the generated statements are factually incorrect or violate fundamental human, social, or specific cultural values, the "illusion" becomes problematic. For example, if people rely on the model's opinions, illusions become a significant issue. However, for creative tasks, the ability to produce unexpected answers can be highly valuable. When people receive unconventional responses, they feel surprised and it prompts them to consider connections between different viewpoints. Problematic Terminology The fact that artificial intelligence spouts nonsense is inevitable because it is not a search engine or database. However, it is still a technological innovation. Although the term "hallucination" is common in the media, tech blogs, and research papers, many consider it inappropriate. Researchers from the University of Oxford published an article in the latest issue of the "Schizophrenia Bulletin" titled "Errors in Artificial Intelligence Models Are Not Hallucinations." They are not the first to find the term "hallucination" unsuitable for a technology. Søren Østergaard and his colleague Kristoffer Nielbo point out two reasons why they believe this term is problematic. While many researchers are studying this problem from various angles, most of them are primarily focused on addressing the issue of chatbots fabricating facts. However, OpenAI warns that if chatbots become better at providing accurate information, this "illusion" could become more dangerous. This is because users may become overly reliant on chatbots, especially in their respective fields, as chatbots provide correct information.

RECENT AI NEWS

RECENT AI TOOLS