Spend enough time with ChatGPT and other artificial intelligence chatbots and it doesn\u2019t take long for them to\u00a0spout falsehoods.<\/p>\n\n\n\n
Described as hallucination, confabulation or just plain making things up, it\u2019s now a problem for every business, organization and high school student trying to get a generative AI system to compose documents and get work done. Some are using it on tasks with the potential for high-stakes consequences, from psychotherapy to researching and\u00a0writing legal briefs.<\/p>\n\n\n\n
\u201cI don\u2019t think that there\u2019s any model today that doesn\u2019t suffer from some hallucination,\u201d said Daniela Amodei, co-founder and president of Anthropic, maker of the chatbot Claude 2.<\/p>\n\n\n\n
\u201cThey\u2019re really just sort of designed to predict the next word,\u201d Amodei said. \u201cAnd so there will be some rate at which the model does that inaccurately.\u201d<\/p>\n\n\n\n
Anthropic, ChatGPT-maker OpenAI and other major developers of AI systems known as large language models say they\u2019re working to make them more truthful.<\/p>\n\n\n\n
How long that will take \u2014 and whether they will ever be good enough to, say, safely dole out medical advice \u2014 remains to be seen.<\/p>\n\n\n\n
\u201cThis isn\u2019t fixable,\u201d said Emily Bender, a linguistics professor and director of the University of Washington\u2019s Computational Linguistics Laboratory. \u201cIt\u2019s inherent in the mismatch between the technology and the proposed use cases.\u201d<\/p>\n\n\n\n
A lot is riding on the reliability of generative\u00a0AI technology. The McKinsey Global Institute projects it will add the equivalent of $2.6 trillion to $4.4 trillion to the global economy. Chatbots are only one part of that frenzy, which also includes technology that can generate new images, video, music and computer code. Nearly all of the tools include some language component.<\/p>\n\n\n\n
Google is already\u00a0pitching a news-writing AI\u00a0product to news organizations, for which accuracy is paramount. The Associated Press is also exploring use of the technology as part of\u00a0a partnership with OpenAI, which is paying to use part of AP\u2019s text archive to improve its AI systems.<\/p>\n\n\n\n
In partnership with India\u2019s hotel management institutes, computer scientist Ganesh Bagler has been working for years to get AI systems, including a\u00a0ChatGPT\u00a0precursor, to invent recipes for South Asian cuisines, such as novel versions of rice-based biryani. A single \u201challucinated\u201d ingredient could be the difference between a tasty and inedible meal.<\/p>\n\n\n\n
When\u00a0Sam Altman, the CEO of OpenAI, visited India in June, the professor at the Indraprastha Institute of Information Technology Delhi had some pointed questions.<\/p>\n\n\n\n
\u201cI guess hallucinations in ChatGPT are still acceptable, but when a recipe comes out hallucinating, it becomes a serious problem,\u201d Bagler said, standing up in a crowded campus auditorium to address Altman on the New Delhi stop of the U.S. tech executive\u2019s\u00a0world tour.<\/p>\n\n\n\n
\u201cWhat\u2019s your take on it?\u201d Bagler eventually asked.<\/p>\n\n\n\n
Altman expressed optimism, if not an outright commitment.<\/p>\n\n\n\n
\u201cI think we will get the hallucination problem to a much, much better place,\u201d Altman said. \u201cI think it will take us a year and a half, two years. Something like that. But at that point we won\u2019t still talk about these. There\u2019s a balance between creativity and perfect accuracy, and the model will need to learn when you want one or the other.\u201d<\/p>\n\n\n\n
But for some experts who have studied the technology, such as University of Washington linguist Bender, those improvements won\u2019t be enough.<\/p>\n\n\n\n
Bender describes a language model as a system for \u201cmodeling the likelihood of different strings of word forms,\u201d given some written data it\u2019s been trained upon.<\/p>\n\n\n\n
It\u2019s how spell checkers are able to detect when you\u2019ve typed the wrong word. It also helps power automatic translation and transcription services, \u201csmoothing the output to look more like typical text in the target language,\u201d Bender said. Many people rely on a version of this technology whenever they use the \u201cautocomplete\u201d feature when composing text messages or emails.<\/p>\n\n\n\n
The latest crop of chatbots such as ChatGPT, Claude 2 or\u00a0Google\u2019s Bard\u00a0try to take that to the next level, by generating entire new passages of text, but Bender said they\u2019re still just repeatedly selecting the most plausible next word in a string.<\/p>\n\n\n\n
When used to generate text, language models \u201care designed to make things up. That\u2019s all they do,\u201d Bender said. They are good at mimicking forms of writing, such as legal contracts,\u00a0television scripts\u00a0or sonnets.<\/p>\n\n\n\n
\u201cBut since they only ever make things up, when the text they have extruded happens to be interpretable as something we deem correct, that is by chance,\u201d Bender said. \u201cEven if they can be tuned to be right more of the time, they will still have failure modes \u2014 and likely the failures will be in the cases where it\u2019s harder for a person reading the text to notice, because they are more obscure.\u201d<\/p>\n\n\n\n
Those errors are not a huge problem for the marketing firms that have been turning to Jasper AI for help writing pitches, said the company\u2019s president, Shane Orlick.<\/p>\n\n\n\n
\u201cHallucinations are actually an added bonus,\u201d Orlick said. \u201cWe have customers all the time that tell us how it came up with ideas \u2014 how Jasper created takes on stories or angles that they would have never thought of themselves.\u201d<\/p>\n\n\n\n
The Texas-based startup works with partners like OpenAI, Anthropic, Google or Facebook parent Meta to offer its customers a smorgasbord of AI language models tailored to their needs. For someone concerned about accuracy, it might offer up Anthropic\u2019s model, while someone concerned with the security of their proprietary source data might get a different model, Orlick said.<\/p>\n\n\n\n
Orlick said he knows hallucinations won\u2019t be easily fixed. He\u2019s counting on companies like Google, which he says must have a \u201creally high standard of factual content\u201d for its search engine, to\u00a0put a lot of energy\u00a0and resources into solutions.<\/p>\n\n\n\n
\u201cI think they have to fix this problem,\u201d Orlick said. \u201cThey\u2019ve got to address this. So I don\u2019t know if it\u2019s ever going to be perfect, but it\u2019ll probably just continue to get better and better over time.\u201d<\/p>\n\n\n\n
Techno-optimists, including Microsoft co-founder Bill Gates, have been forecasting a rosy outlook.<\/p>\n\n\n\n
\u201cI\u2019m optimistic that, over time, AI models can be taught to distinguish fact from fiction,\u201d Gates said in a July blog post detailing his thoughts on AI\u2019s societal risks.<\/p>\n\n\n\n
He cited a 2022 paper from OpenAI as an example of \u201cpromising work on this front.\u201d More recently, researchers at the Swiss Federal Institute of Technology in Zurich said they developed a method to detect some, but not all, of ChatGPT\u2019s hallucinated content and remove it automatically.<\/p>\n\n\n\n
But even Altman, as he markets the products for a variety of uses, doesn\u2019t count on the models to be truthful when he\u2019s looking for information.<\/p>\n\n\n\n
\u201cI probably trust the answers that come out of ChatGPT the least of anybody on Earth,\u201d Altman told the crowd at Bagler\u2019s university, to laughter.<\/p>\n","protected":false},"excerpt":{"rendered":"
Spend enough time with ChatGPT and other artificial int […]<\/p>\n","protected":false},"author":1,"featured_media":2694,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[],"class_list":["post-2693","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-business"],"_links":{"self":[{"href":"https:\/\/www.viewworld.org\/index.php?rest_route=\/wp\/v2\/posts\/2693","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.viewworld.org\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.viewworld.org\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.viewworld.org\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.viewworld.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2693"}],"version-history":[{"count":1,"href":"https:\/\/www.viewworld.org\/index.php?rest_route=\/wp\/v2\/posts\/2693\/revisions"}],"predecessor-version":[{"id":2695,"href":"https:\/\/www.viewworld.org\/index.php?rest_route=\/wp\/v2\/posts\/2693\/revisions\/2695"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.viewworld.org\/index.php?rest_route=\/wp\/v2\/media\/2694"}],"wp:attachment":[{"href":"https:\/\/www.viewworld.org\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2693"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.viewworld.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2693"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.viewworld.org\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2693"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}