On November 30, 2022, OpenAI released a language model called ChatGPT that has many implications for second-language learning and teaching as well as for translation. In the following four videos, Tom Gally of the University of Tokyo discusses some of those implications and gives examples of how ChatGPT might—and might not—be used in language education and translation.
Translating with ChatGPT
This video discusses the use of OpenAI’s ChatGPT for translating between languages and shows why it is sometimes superior to dedicated machine translation systems like Google Translate, DeepL, and Bing Microsoft Translator. The video was recorded on February 8, 2023.
ChatGPT and the Future of Language Learning
This video—the third in the series—discusses the potential application of OpenAI’s ChatGPT and similar software to second-language learning and education. Tom Gally explains why this interactive AI software seems fundamentally different from past technologies that were applied to language learning without much success. The video was recorded on December 15, 2022, two weeks after the release of ChatGPT. A transcript of the video appears below.
Using ChatGPT for Language Learning
This video explores how well ChatGPT can perform high-level tasks related to language learning and teaching. The preliminary results suggest that it can explain word meanings and create example sentences as well as an expert human can. It seems to be weaker at some other tasks, such as explaining grammar and preparing vocabulary quizzes; it might do better at those tasks if different prompts are used. This video was recorded on December 7, 2022. The examples shown in the video, as well as others, are below.
Examples shown in “Using ChatGPT for Language Learning”
The following video by Tom Gally, recorded on December 5, 2022, demonstrates ChatGTP briefly and starts a discussion about what this software means for language education. A transcript of the video appears below.
Transcript of “ChatGPT and Language Education”
My name is Tom Gally, and I do research and teach on fields related to second language education at the University of Tokyo. Today is December 5th, 2022, and it’s five days, six days since the company OpenAI released its ChatGPT software. And so I would like to show a brief demonstration of that software and talk about the implications of this new software and similar large language models for the field of language education.
First of all, ChatGPT is now being released in kind of a beta version, where the people who have OpenAI accounts can test it. Apparently OpenAI wants to evaluate it for its safety in terms of ethical and related issues. But many hundred of thousands of people around the world seem to be using it right now.
What it can do is, you input some text, and you can ask questions or make a comment, click return, and then you get a response. And the questions can be about history; they can be about travel; they can be about technical subjects. And the software is interactive, so if you get an answer from the software, you can reply to that and respond to that, carry on conversations to some extent.
It’s not perfect: there are gaps in its knowledge, it’s not great at calculation apparently. But in the case of, for example, asking it to write a story, to write a short essay, it produces what are close to perfect texts.
It’s also possible to give prompts in another language. So I’ve tested it with Japanese. I’ll give a prompt in Japanese and ask it to reply in English, and it will do that.
So, for example, you can ask it to produce a three-paragraph essay on the origins of the Second World War, and it produces, in just a couple of seconds, that essay—better than any essay produced by any student I’ve ever taught at the top university in Japan.
So this has a lot of implications for the field of language education, implications that I, at least, do not yet grasp.
So about five years ago, six years ago, there was a similar revolution, when Google’s machine translation software adopted a new neural-network approach to doing online translations. And so, suddenly, especially between languages like Japanese and English, machine translation became usable in many cases. And so, for example, learners of a language could use it to write their homework; they could also use it to find out what a text in a foreign language meant with pretty good accuracy. And it can also be used in practical situations.
But this ChatGPT goes beyond that. It’s not just translating. By the way, it can translate; it is able to translate as well in the tests I’ve done. But it goes beyond that in actually producing texts and being interactive in the texts.
Right now, the interface at OpenAI is only for written text. But I think it’s pretty obvious that this or similar software will soon become available in spoken form as well, too. Voice recognition and speech synthesis are very, very good now. So within months, I expect, we’ll be seeing some very high quality voice interactive software, where you can talk to this bot and it seems very, very close to being a real person. And I’m sure there will be cases where software developers will create characters with persistent personalities so that one could go back and converse with the same person again and again.
This software is very close to passing or has already passed the Turing test, the test proposed by Alan Turing many years ago that says that artificial intelligence has been achieved when you can’t tell the difference between a computer and a human being. And so that’s going to be happening soon—very, very soon.
So the implications of this software go far beyond the field of language education. And those discussions are starting. They’ve been going on for some time, but they’ve been kickstarted within the last few days by the arrival of ChatGPT. So I made this video to encourage people, at least in the field of language education, just take this move seriously.
The response to the rapid improvement of machine translation, starting in late 2016, was slow. It took a while for educators to start taking it seriously. I think they needed to know that their students were using it before they realized that they need to adapt their education. Much education still has not adapted to it. At least here in Japan, education policy, the policy of the government, the educational policy of the government, has not yet responded to it at all.
And so this next leap in artificial intelligence shown by ChatGPT raises even bigger issues about why and how people use language and why and how people learn and teach other languages. So I made this video to encourage discussion and dialogue.
I talked to a group of graduate students today, and they had not heard of this. Of course, it was only released five days ago, but they were blown away by it when I demonstrated it to them, and I’m sure almost everybody is.
And so I encourage you to get familiar with the software and read the discussions that are going on. Right now, they’re going on in places like Hacker News and Reddit and Twitter, Mastodon. I imagine they’re taking place other places as well too. If you’re an educator in the field of second language education, talk to your colleagues about it, talk to other educators about it, talk to your students about it, and contemplate on what it means.
I can’t predict what’s going to happen, but the software is being developed very, very quickly, and we can imagine other rapid advances soon, like the spoken interface which I mentioned. And so it’s better to be ready earlier than later to deal with it.
So thank you very much for listening.
Transcript of “ChatGPT and the Future of Language Learning”
Hi. This is Tom Gally at the University of Tokyo. It’s been two weeks since ChatGPT was released by OpenAI, and shockwaves continue to go around the world. People continue to be amazed and kind of frightened at how lifelike, how effective, how smart it is in some ways, and also how stupid it is in other ways. But it really seems to be a big leap towards sort of human-like AI.
Naturally, there are a lot of concerns about it in terms of: Will it be taking away people’s jobs? Will it be used for fake news or for evil campaigns? Those issues aren’t resolved.
But I would like to talk about an area of my own interest, where I think maybe ChatGPT and similar programs that will certainly emerge soon have positive potential, which is in language learning and maybe language teaching. I’ve made a couple of videos so far about how ChatGPT might be used by language learners, maybe language teachers, successful in some ways, unsuccessful in other ways, but overall, quite intriguing, quite fascinating in what it can do. Its ability in some areas, linguistically, maybe even exceeds the average human, the average language teacher.
And so I would like to think about that. What does that mean for us? What does that mean for language teachers and for language learners in the months and years ahead?
So first of all, thinking about how humans learn language. So maybe we can focus on two aspects of human language learning. One is people learning their first language. Children, for the most part, learn their first language from other human beings, so from their family and from the kids around them, their teachers at school. Of course, these days, children get a lot of input from media as well, too. But I think it’s that human interaction that is the primary driver of first language acquisition.
But in the case of second languages for teenagers or older, especially foreign languages, often that human element has been lacking because it’s a foreign language—there’s nobody around you who speaks the language that you want to learn, for example. And so the reason ChatGPT is so fascinating: Because it’s so human-like, I’m just wondering if it has the potential to serve that function of a human interlocutor, the function of someone that you can learn from, the way, ideally, human language learning has taken place before, both for children and for adults.
Before I think about that in a little more detail, I want to think about, well, I want to be cautious: Am I just being fascinated by this new technology? I’m kind of a nerd, I’m an early adopter of technology, and this has only been available for two weeks. Am I running ahead of myself?
So what I want to do first is think about the history of the adoption of new technologies into language education and language learning, because it goes a long way back, and it has not always been successful. There have been a lot of failed attempts to introduce new technology to language learning.
One that comes to mind immediately, of course, is printing—the book. And so this was a revolution in its time—it continues to be a revolution, the ability to preserve language on paper and to transmit it over long distances. And then, since Gutenberg, being able to make many, many copies cheaply.
And so, beginning centuries ago, of course, language learning, second language learning, foreign language learning, was enabled for many people and driven by the existence of printed matter, of books. And so the books were a very excellent technology, but they were obviously, especially for language learning, they were inadequate in some ways. They’re not interactive, obviously, but also they don’t have an audio component. So you can’t hear what the sounds of the language are.
I remember as a child, when I was interested in language, checking out books from the library on—I don’t know what it was, classical Greek or French or something like that. And reading descriptions in the introduction of the book—this is how the sound is pronounced in French or in Sanskrit or whatever it was. And it was practically useless for me, because I had not yet studied phonetics or linguistics and I had no audio recordings at that time to listen to. And so I could not imagine from the descriptions of the sounds how they were pronounced.
I think the real first arrival of early-adopted technology was audio recording and broadcasting. So with Thomas Edison and Marconi and others, when they found out that they could record the sounds of a language or broadcast it over long distances, then—this is a century ago, more than a century ago—there were a lot of efforts to apply that to language learning.
This was less than a century ago, but I remember when I was a child, we did have some LP records in my home in California. This was 60 years ago. We had LP records of I think it was French and maybe Spanish, that I guess my parents had bought. And I would put them on the record player occasionally and listen to them, you know, “Bonjour.” I don’t remember what it was, but it was— They had been bought. My parents had bought them. My parents had not learned anything from them. The local public library had many, many recordings of language lessons, like sets of language educational materials, in which would be a set of LP records with accompanying text.
And I think a lot of people tried to learn from them. And probably some people did succeed in learning the language well, or at least getting a good start, from those records. But overall, they were not a successful technology. They did not replace teachers. They did not lead to a big increase in foreign language learning anywhere in the world.
Radio was used as well, too. Here in Japan, where I live, I think it’s close to 100 years since the broadcaster NHK has been broadcasting; I think it was in the 1930s that they started broadcasting English and other foreign language lessons on the radio. After World War II, when Japan was occupied by the American forces, there was a big boom in English language learning through NHK radio. And it didn’t have much of an effect. There were some people who did learn. I have one good friend who passed away a year ago who learned English quite well, largely from listening to the radio as a child. But those were exceptions.
And so that technology that seemed to offer so much promise—to be able to spread language learning widely and cheaply—didn’t work very well.
Later, in my own life, I’ve experienced other ones. When I was in university in California in the 1970s, every classroom in the university had these big television sets up on the wall, hanging from the ceiling. And they were all connected to a central control room. In those days, it was not yet possible to record video on compact cassettes. And so they had these reel-to-reel, I don’t know how big they were, about inch-wide big tapes on reel-to-reel machines in a central location where they could have classes—this was not only for languages, but also for other subjects—where they could have lectures. And the idea was you wouldn’t have to have the human teacher in the classroom, the students would just go to the classroom, watch the television, get the lecture, get the education.
It didn’t work. In the three years I was at that university, from 1975 to 1978, I never saw those televisions being used in any of the classrooms. It was a big waste of money.
And so since then, I’ve experienced other introductions of new technology to language learning that did not succeed. Computer-aided: There were a lot of, in the 80s and the 90s, there were a lot of computer-aided programs for learning, and a few people, a handful of people, were successful in learning from them. But it just didn’t work. And so probably many of you can think of other cases like that, where somebody like me, some educator was very enthusiastic about some technologies: “We’re going to adopt this!” And then you apply to the higher-ups, the administrators, the government officials, and it sounds attractive. “This is the latest technology! So therefore, we will give you some money to implement it in all of your schools.” And it doesn’t work, okay?
Often the technology is not ready yet. So obviously now, audio technology, video technology, what I’m talking to you right now on YouTube is fantastic learning materials, and it has become very, very easy to use and very easy to watch. But there were many cases in which things like video and audio and computer-aided learning were adopted much too early.
But I think the reason they failed was not only the fact that the technology was too cumbersome or too expensive. There was also the fact that those technologies did not involve human interaction. So, as I said, people learn languages from other people, by interacting with other people. So both the linguistic input and the linguistic stimulation that comes from interaction is very, very important.
But also especially the motivation to continue learning. So there are some people who can continue to study a language on their own, by themselves, for the months, years that is necessary to become fluent, but most people need some encouragement along the way. So if you’re sitting in a classroom with 20 other, 30 other students, having that human teacher in front of the class looking at you and occasionally responding to you and throwing questions at you, seems to be a major driving force for allowing people to continue studying and learning languages for the long time that is necessary to become proficient.
So what I’m wondering is: Does this AI, the AI that is starting to emerge, does it have the potential to replace, to some extent or to a large extent, the human element in language learning?
Well, I think we saw a hint of this six years ago, when machine translation suddenly got a lot better. And so with Google Translate and now DeepL and others, it is possible to some extent to actually have communication between human beings using machine translation, using a spoken computer or smartphone app interpreters. And so that was a revolution. Not so much in teaching. Machine translation can be applied in teaching, but I think it’s a revolution in how humans use language.
So before machine translation became moderately good, the only way for two people who don’t have a common language—two people who, one person speaks Chinese, the other person speaks French, and neither of them speaks the other’s language, and they don’t speak English or any other common language—the only way for two people like that to communicate, to cooperate, was to have a human intermediary. So there had to be one person, there had to be somebody, a translator, an interpreter, who would translate between them. That was a human being. Or those two people would learn each other’s languages or they would learn a common language.
So the reason machine translation is a revolution is it enables communication and cooperation between people without a common language by computer—cheaply, very cheaply, and conveniently. Accuracy still has problems, of course, and knowing how to use it is an important issue. But that shed the light on how that could be used.
Well, ChatGPT, which we’ve just been able to experiment with for two weeks now, brings a human element, even a bigger human element into that, a more important human element, in the way it is able to conduct very natural-sounding conversations. You can ask it to ask you questions, and it will ask you questions or elicit information from you.
It can explain language, it can explain the meanings. I’ve been especially amazed at how well it can explain the meanings of words and it can figure out the meanings of individual words from their context, which is a rather advanced skill for human beings. Many language teachers are not as good as ChatGPT in doing that in the case of English. It’s not as good in the other language I’ve tested it on, Japanese, but it’s very, very good with English.
And so I think this is a tipping point for language education and for language learning. All of those previous technologies had largely—they came too early, a lot of money was wasted, a lot of time was wasted. And for those who had hoped to replace human teachers in some teaching contexts, they were all disappointed. So the learners themselves who didn’t have a teacher, could not afford to pay a teacher, who hoped to be able to learn a language and acquire a language solely from the audio recordings or the smartphone apps—most of them have failed. It didn’t succeed. Some people have. But you just don’t continue. The material is too hard or it’s too easy, and it doesn’t adapt to you, and it doesn’t encourage you in any way.
And so what I see from this new AI, this interactive AI, is the potential for replacing most of that. In other words, I can easily imagine having a spoken bot or maybe a bot that appears both audio on your smart speaker and it’s also an avatar that appears on the screen of your phone or your computer. And it’s friendly and it chats with you and it says hello in the morning and it remembers what you talked about yesterday and last week and two weeks ago.
And it’s also your language teacher. And so it will remember, for example, it remembers the words that you’ve studied before, remembers the words that you already know, and it remembers which words you haven’t acquired yet. So it’s able to create conversations, have interesting conversations about topics, whatever you’re interested in, and it uses the vocabulary that you need practice on.
A human teacher cannot do that easily. If you have a lot of experience with a certain kind of learner, you can adapt your speaking style to it. This is what junior high school students in Malaysia know in English. This is what they don’t know. And so you can adapt to some extent in that regard. But it’s clear that programs like ChatGPT have the potential to adjust their teaching and their speaking, their interaction, in a way that is just enough to keep improving the learner’s vocabulary, building the vocabulary, exposing them to different accents, and also obviously grammatical forms as well, too.
And so it would be able to do this in a fun, interactive way. And I can easily imagine them becoming virtual characters that would be sufficiently human-like, have warm enough personalities and interesting enough personalities to be attractive to many, many, many users. Some people will resist that. But there are many people, many young people these days, who are quite into bots and virtual characters. There’s this one called Hatsune Miku in Japan that’s a completely virtual character. They have concerts. You know, thousands of young people will go to these concerts to watch this virtual character on the screen. And I’ve heard people get tears when they’re seeing their virtual character in person. And this will be individualized for you.
Obviously, the applications are not restricted to language learning. And, obviously, there are many possible applications of this that are causes of concern in terms of the political uses and uses in fraud and many things like that.
But narrowing our scope to language learning, in which there has really always been a shortage of human teachers for language learning, for foreign languages, especially. And so I think it has a great, great potential.
So one concern, of course, that many people will have, people like me who have been language teachers, who have made their money by language teaching: Will this mean the end of jobs for language teachers? Well, people should be concerned about that. They should think about that. I don’t think that is what should drive the conversation. I think language learning, language acquisition, is more important than protecting people’s jobs.
But I don’t think that these virtual characters, if they emerge, which they probably will, I don’t think that those characters will replace the needs for human educators completely. But I do think the role of human teachers, at least in the case of languages, will change.
And so the sort of the classical language class in a junior high school, in a high school or university—okay, you have 30 students in the class, there’s one teacher, and the teacher says, “Okay, everybody, open up your textbooks to page 123. We’re going to do Lesson 52 today. OK, could you please read the text? Repeat after this and answer this question and fill in the blanks on the worksheet.” When you have this kind of mass education like that, that kind of teaching is kind of unavoidable.
Well, I think the need for that and the role for that is going to disappear. So all of the students could be using their apps, whether it’s in the classroom, whether it’s at home or someplace else. They will have their lessons that will be tailored for them, that will be more interesting than that teacher reading from the textbook in front of the class.
But, on the other hand, the motivational element. Much of the motivational element maybe will be offloaded to the apps. But the element of, you know, what role does this language play in my life? Why should I continue? People will be modeling themselves still on human beings.
And so I think the role of the teacher will become more like a coach: to encourage and guide the students in their learning, and motivating, continuing to motivate, the students.
But much of the teaching, much of the support for language acquisition, probably can be done by this software.
So it’s a new world that’s emerging. It just has emerged in this sense within a couple of weeks. There is a danger that I’m being overly optimistic. It’s also possible that the sufficiently customizable software might not be available for some time. It’s not clear when OpenAI is going to release an API, an application program interface, that will allow other applications to interact with the program. That’s what would be necessary. And, also, it’s not clear whether or when competing versions of this, open source versions or commercial versions, will emerge. But I think it’s almost certain that the potential applications in so many fields are really, really intriguing. And so I think the technology is going to improve.
So as I said in my other videos, if you watched them to the end—and thank you for watching this one to the end as well, too—there’s a lot we don’t know. There’s a lot we have to try. And so I encourage language learners, language teachers, to try this software, and especially to talk about it with other people. Share techniques. Share what succeeds, what doesn’t succeed, how you can use it. What are people’s reactions to it, learners’ reactions to it? Does it seem to have potential? Does it seem boring? And we’ll have to move on from there.
Thank you very much for listening this far. This might be my last video on the topic, but I might come back. In any case, I do hope the conversation about this issue goes far beyond this room in Yokohama, Japan, and that people will be talking about it around the world.
Thank you very much for listening.
An earlier version of this page, with examples of ChatGPT output, is archived here.