(点击右边三个点，可调整速度，电脑上可下载)

听力参考原文 ↓↓↓

[00:00.04]Newly-developed artificial intelligence (AI) systems

[00:05.76]have demonstrated the ability to perform

[00:09.32]at human-like levels in some areas.

[00:13.52]But one serious problem with the tools remains

[00:18.28]– they can produce false or harmful information repeatedly.

[00:24.16]The development of such systems, known as "chatbots,"

[00:28.72]has progressed greatly in recent months.

[00:32.80]Chatbots have shown the ability to interact

[00:36.48]smoothly with humans and produce complex writing

[00:41.56]based on short, written commands.

[00:44.72]Such tools are also known as "generative AI"

[00:49.60]or "large language models."

[00:52.84]Chatbots are one of many different AI systems

[00:57.12]currently under development.

[00:59.64]Others include tools that can produce new images,

[01:04.32]video and music or can write computer programs.

[01:09.84]As the technology continues to progress,

[01:13.76]some experts worry that AI tools

[01:17.52]may never be able to learn how to avoid false,

[01:21.72]outdated or damaging results.

[01:25.92]The term hallucination has been used to describe

[01:30.88]when chatbots produce inaccurate or false information.

[01:36.40]Generally, hallucination describes something that

[01:40.40]is created in a person's mind,

[01:43.48]but is not happening in real life.

[01:47.12]Daniela Amodei is co-creator

[01:50.28]and president of Anthropic,

[01:53.36]a company that produced a chatbot called Claude 2.

[01:58.92]She told the Associated Press,

[02:01.84]"I don't think that there's any model today

[02:05.40]that doesn't suffer from some hallucination."

[02:09.92]Amodei added that such tools

[02:12.84]are largely built "to predict the next word."

[02:17.48]With this kind of design, she said,

[02:20.36]there will always be times when the model

[02:23.52]gets information or context wrong.

[02:27.92]Anthropic, ChatGPT-maker OpenAI

[02:32.28]and other major developers of such AI systems

[02:36.52]say they are working to make AI tools

[02:40.32]that make fewer mistakes.

[02:42.92]Some experts question how long

[02:45.96]that process will take

[02:48.12]or if success is even possible.

[02:52.00]"This isn't fixable," says Professor Emily Bender.

[02:57.16]She is a language expert and

[02:59.72]director of the University of Washington's

[03:02.80]Computational Linguistics Laboratory.

[03:07.12]Bender told the AP she considers

[03:10.52]the general relationship between AI tools

[03:14.60]and proposed uses of the technology a "mismatch."

[03:20.08]Indian computer scientist Ganesh Bagler

[03:24.08]has been working for years to get AI systems

[03:28.28]to create recipes for South Asian foods.

[03:32.92]He said a chatbot can generate misinformation

[03:37.28]in the food industry that could hurt a food business.

[03:42.16]A single "hallucinated" recipe element

[03:45.72]could be the difference between a tasty meal

[03:49.00]or a terrible one.

[03:51.76]Bagler questioned OpenAI chief Sam Altman

[03:56.36]during an event on AI technology

[03:59.48]held in India in June.

[04:02.64]"I guess hallucinations in ChatGPT are still acceptable,

[04:08.36]but when a recipe comes out hallucinating,

[04:11.80]it becomes a serious problem," Bagler said.

[04:16.68]Altman answered by saying

[04:19.24]he was sure developers of AI chatbots

[04:23.20]would be able to get "the hallucination problem

[04:27.16]to a much, much better place" in the future.

[04:31.52]But he noted such progress could take years.

[04:35.96]"At that point we won't still talk about these," Altman said.

[04:41.56]"There's a balance between creativity

[04:44.64]and perfect accuracy, and the model will need

[04:49.12]to learn when you want one or the other."

[04:53.52]Other experts who have long studied the technology

[04:57.68]say they do not expect

[04:59.84]such improvements to happen anytime soon.

[05:04.60]The University of Washington's Bender

[05:07.52]describes a language model as a system

[05:11.28]that has been trained on written data

[05:14.28]to "model the likelihood of different strings of word forms."

[05:20.24]Many people depend on a version of this technology

[05:24.36]whenever they use the "autocomplete" tool

[05:28.24]when writing text messages or emails.

[05:32.48]The latest chatbot tools

[05:35.20]try to take that method to the next level,

[05:38.60]by generating whole new passages of text.

[05:42.88]But Bender says the systems are still

[05:46.28]just repeatedly choosing

[05:48.44]the most predictable next word in a series.

[05:52.84]Such language models "are designed to make things up.

[05:57.52]That's all they do," she noted.

[06:00.64]Some businesses, however, are not so worried

[06:04.76]about the ways current chatbot tools

[06:07.92]generate their results.

[06:10.44]Shane Orlick is head of marketing technology company Jasper AI.

[06:17.04]He told the AP, "Hallucinations are actually an added bonus."

[06:23.44]He explained many chatbot users were pleased

[06:27.84]that the company's AI tool

[06:30.28]had "created takes on stories or angles

[06:34.52]that they would have never thought of themselves."

[06:38.56]I'm Bryan Lynn.

____________________

Words in This Story

generate – v. to produce something

inaccurate – adj. not correct or exact

context – n. all the facts, opinions, situations, etc. relating to a particular thing or event

mismatch – n. a situation when people or things are put together but are not suitable for each other

recipe – n. a set of instructions and ingredients for preparing a particular food dish

bonus – n. a pleasant extra thing

angle – n. a position from which something is looked at