Spend enough time with ChatGPT and other artificial intelligence chatbots and it doesn鈥檛 take long for them to .
Described as hallucination, confabulation or just plain making things up, it鈥檚 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 .
鈥淚 don鈥檛 think that there鈥檚 any model today that doesn鈥檛 suffer from some hallucination,鈥� said Daniela Amodei, co-founder and president of Anthropic, maker of the chatbot Claude 2.
鈥淭hey鈥檙e really just sort of designed to predict the next word,鈥� Amodei said. 鈥淎nd so there will be some rate at which the model does that inaccurately.鈥�
Anthropic, ChatGPT-maker OpenAI and other major developers of AI systems known as large language models say they鈥檙e working to make them more truthful.
How long that will take 鈥� and whether they will ever be good enough to, say, safely dole out medical advice 鈥� remains to be seen.
鈥淭his isn鈥檛 fixable,鈥� said Emily Bender, a linguistics professor and director of the University of Washington鈥檚 Computational Linguistics Laboratory. 鈥淚t鈥檚 inherent in the mismatch between the technology and the proposed use cases.鈥�
A lot is riding on the reliability of generative . 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.
Google is already product to news organizations, for which accuracy is paramount. The Associated Press is also exploring use of the technology as part of , which is paying to use part of AP鈥檚 text archive to improve its AI systems.
In partnership with India鈥檚 hotel management institutes, computer scientist Ganesh Bagler has been working for years to get AI systems, including a precursor, to invent recipes for South Asian cuisines, such as novel versions of rice-based biryani. A single 鈥渉allucinated鈥� ingredient could be the difference between a tasty and inedible meal.
When , visited India in June, the professor at the Indraprastha Institute of Information Technology Delhi had some pointed questions.
鈥淚 guess hallucinations in ChatGPT are still acceptable, but when a recipe comes out hallucinating, it becomes a serious problem,鈥� Bagler said, standing up in a crowded campus auditorium to address Altman on the New Delhi stop of the U.S. tech executive鈥檚 .
鈥淲hat鈥檚 your take on it?鈥� Bagler eventually asked.
Altman expressed optimism, if not an outright commitment.
鈥淚 think we will get the hallucination problem to a much, much better place,鈥� Altman said. 鈥淚 think it will take us a year and a half, two years. Something like that. But at that point we won鈥檛 still talk about these. There鈥檚 a balance between creativity and perfect accuracy, and the model will need to learn when you want one or the other.鈥�
But for some experts who have studied the technology, such as University of Washington linguist Bender, those improvements won鈥檛 be enough.
Bender describes a language model as a system for 鈥渕odeling the likelihood of different strings of word forms,鈥� given some written data it鈥檚 been trained upon.
It鈥檚 how spell checkers are able to detect when you鈥檝e typed the wrong word. It also helps power automatic translation and transcription services, 鈥渟moothing the output to look more like typical text in the target language,鈥� Bender said. Many people rely on a version of this technology whenever they use the 鈥渁utocomplete鈥� feature when composing text messages or emails.
The latest crop of chatbots such as ChatGPT, Claude 2 or try to take that to the next level, by generating entire new passages of text, but Bender said they鈥檙e still just repeatedly selecting the most plausible next word in a string.
When used to generate text, language models 鈥渁re designed to make things up. That鈥檚 all they do,鈥� Bender said. They are good at mimicking forms of writing, such as legal contracts, or sonnets.
鈥淏ut 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,鈥� Bender said. 鈥淓ven if they can be tuned to be right more of the time, they will still have failure modes 鈥� and likely the failures will be in the cases where it鈥檚 harder for a person reading the text to notice, because they are more obscure.鈥�
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鈥檚 president, Shane Orlick.
鈥淗allucinations are actually an added bonus,鈥� Orlick said. 鈥淲e have customers all the time that tell us how it came up with ideas 鈥� how Jasper created takes on stories or angles that they would have never thought of themselves.鈥�
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鈥檚 model, while someone concerned with the security of their proprietary source data might get a different model, Orlick said.
Orlick said he knows hallucinations won鈥檛 be easily fixed. He鈥檚 counting on companies like Google, which he says must have a 鈥渞eally high standard of factual content鈥� for its search engine, to and resources into solutions.
鈥淚 think they have to fix this problem,鈥� Orlick said. 鈥淭hey鈥檝e got to address this. So I don鈥檛 know if it鈥檚 ever going to be perfect, but it鈥檒l probably just continue to get better and better over time.鈥�
Techno-optimists, including Microsoft co-founder Bill Gates, have been forecasting a rosy outlook.
鈥淚鈥檓 optimistic that, over time, AI models can be taught to distinguish fact from fiction,鈥� Gates said in a July blog post detailing his thoughts on AI鈥檚 societal risks.
He cited a 2022 paper from OpenAI as an example of 鈥減romising work on this front.鈥�
But even Altman, as he markets the products for a variety of uses, doesn鈥檛 count on the models to be truthful when he鈥檚 looking for information for himself.
鈥淚 probably trust the answers that come out of ChatGPT the least of anybody on Earth,鈥� Altman told the crowd at Bagler鈥檚 university, to laughter.
READ ALSO:
READ ALSO:
