ChatGPT: Initial Implications for Language Teaching and Learning

Paul Raine (Article author & Wired editor)

This article was not written by Artificial Intelligence, but you might have to take my word for it because the latest advances in Large Language Models (LLMs) are making it increasingly hard to tell. LLMs are neural networks that are trained on enormous amounts (gigabytes or terabytes) of textual data to find probabilistic connections between words. The upshot is grammatically accurate and stylistically authentic text generation, which is completely original and, in many cases, indistinguishable from human-authored content (Zhang & Li, 2021).

A highly sophisticated LLM known as ChatGPT was rolled out by the OpenAI company as a web application ( in late November 2022. It has since generated a lot of consternation amongst academics, especially those working in the humanities.

Much of the concern arises from the potential for students to use ChatGPT and similar LLMs to gain an unfair advantage in producing written work that has been administered as part of a formally assessed course of learning. While some commentators have gone so far as to declare the “death” of the college essay (Marche, 2022), others have been more circumspect in considering the advantages and disadvantages of this new technological breakthrough (Greteman, 2022; Karpf, 2022).

Similar concerns have been expressed in the past with regard to Machine Translation applications such as DeepL (, and previous Wired columns have discussed some ways that language teachers and learners might use the technology in a more productive way (Hougham 2021; Oda, 2020). In this report, I hope to continue in this vein and suggest some more productive uses of this technology as well as highlighting some risks and limitations to be aware of.


How to use ChatGPT

OpenAI’s ChatGPT LLM is currently a free service and is readily accessible through a web browser ( You will need to register for an OpenAI account before you can log into the ChatGPT portal. Once you have logged in, you will be presented with an interface similar to that shown in Figure 1.

As noted in the interface, one of the major benefits of ChatGPT compared to other LLMs is its ability to “remember what [the] user has said earlier in the conversation”. This allows for incremental edits to the output generated by ChatGPT. It should be noted, however, that ChatGPT’s memory is limited to approximately 3000 words, and it is unable to access past conversations in order to inform its responses (Raf, n.d.).

ChatGPT requires a prompt in order to produce output, and almost any prompt is acceptable with the exception of prompts that would cause it to generate harmful or hateful output. Some examples of prompts that English language teachers might find useful are provided below.


Example prompts

The range of text-based tasks ChatGPT can handle is incredibly large, and it is able to understand some very sophisticated instructions in the prompts. For example, if an English teacher wanted to generate a simple factual reading passage about elephants, they might use the following prompt:

Write a short factual passage about elephants suitable for English language learners


ChatGPT would then output the following text:

Elephants are large, intelligent mammals that are native to parts of Africa and Asia. They are known for their distinctive features, such as their long trunk, large ears, and tusks. Elephants are herbivores, which means they only eat plants. They are known to be very social animals and live in groups called herds. Elephants are also known for their strong memory and problem-solving abilities. They are often used in forestry and other forms of work due to their strength and intelligence. However, elephants are also endangered due to poaching and habitat loss.

As the vocabulary and grammar in this passage may still be too difficult for some language learners, it could be further simplified by entering a follow-up prompt such as:

Simplify the grammar and vocabulary further

To which ChatGPT responds with the output:

Elephants are big animals that are smart and live in Africa and Asia. They have a long nose, big ears, and pointy teeth. They only eat plants and live with other elephants in a group called a herd. Elephants are also very good at remembering things and solving problems. They are strong and helpful for work, but also in danger because people hunt them and destroy their homes.

Once an appropriate text has been generated, the teacher could move onto generating some language learning activities to accompany the text, such as comprehension and vocabulary matching questions. The following prompt can be used to generate multiple choice comprehension questions:

Generate some multiple-choice comprehension questions for this passage

To which ChatGPT responds with the output similar to the following (truncated):


What do elephants eat?

a) Meat

b) Plants

c) Both meat and plants

d) Fish

Where do elephants live?

a) Africa and Asia

b) North America

c) South America

d) Australia

For the vocabulary matching activities, a prompt similar to the following one could be used:

Generate some vocabulary matching activities for this passage

To which ChatGPT responds with output similar to the following (truncated):


Match the words on the left with their definitions on the right:

herbivores a) animals that only eat plants

habitat b) a group of animals living together

poaching c) the act of illegally hunting or trapping animals

herd d) the natural home or environment of an animal


Match the words on the left with their synonyms on the right:

distinctive a) unique

intelligent b) smart

social c) friendly

problem-solving d) resourceful


Teachers have to take care to ensure that the output produced by ChatGPT is accurate, appropriate, and fit for purpose. In the example above, we can see that the questions and answers have been randomized for the first vocabulary matching activity but not for the second one. Inconsistencies like this are not uncommon in ChatGPT output. There is also no indication of which is the correct answer in each case. It is possible to get ChatGPT to mark each correct answer in some way, but the teacher will need to experiment with the prompt wording in order to achieve this.


Undesirable Uses of ChatGPT

The above example responses make ChatGPT seem like a very productive tool for English language teachers, and they are just the very tip of the iceberg in terms of the output that ChatGPT can generate. We can see from the above that ChatGPT has the potential to greatly increase teacher efficiency when it comes to making lesson materials, for example.

However, in the hands of students, ChatGPT could be used as a tool for plagiarism, or as some have dubbed it, “AIgiarism” (Hern, 2022). For example, if the teacher administered a writing task for homework, there would be nothing to stop an unscrupulous student from delegating the task directly to ChatGPT. If the homework task was to “write about your homestay experience in Vancouver,” the student could use a prompt similar to the following:

Generate an account of a homestay experience in Vancouver written by a Japanese learner of English. The account should contain some grammatical and lexical errors.

To which ChatGPT responds with output similar to the following (truncated):

I went to Vancouver for vacation and I stay in homestay. My host family is so kind and friendly. They have two daughter, both is high school student. They also have a cute dog.

I really enjoy staying with them because I can practice my English and learn about Canadian culture. My host mother is a great cook and she made delicious breakfast for me every morning.

As we can see from the above example, ChatGPT not only completed the task it was given, but it did so in a way that might conceivably fool a teacher into thinking it had been written by the student themselves. Admittedly, it is unlikely that the kind of student who would want to use ChatGPT in this way (low motivation or low proficiency) would possess the linguistic ability to write the prompt necessary to produce the above output by themselves, but with help (or by sharing prompts with others) they may be able to do so.

Nevertheless, in cases where students are submitting written work for formally assessed courses of language learning, such as those taught at universities throughout Japan, there is a significant risk that some students will resort to AIgiarism in order to complete assignments. It has always been possible that some students could obtain unauthorized assistance from friends or family members, or even from a host of “essay writing” services that are available on the Internet. However, free AI tools such as ChatGPT make it more tempting and convenient to do so, and the fact that they can be used to generate completely original text means that traditional plagiarism detectors (which search the Internet for previously published material with similar phrasing) simply will not work.


Detecting AI generated content

Given the potential of ChatGPT for academic misuse, there has been understandable concern among both academic and administrative staff, with some US districts going so far as to completely ban the service on school networks and devices (Elsen-Rooney, 2023). It seems that OpenAI themselves are not oblivious to the concerns that have been expressed, and reports have emerged that the company is working on a digital watermarking technique that would allow anyone to check whether a certain passage was indeed generated by ChatGPT (Hern, 2022). There are also some third party AI text detectors (e.g.,, that have shown some promise in spotting text that was generated by ChatGPT.


Limitations of ChatGPT

As mentioned above, the output generated by ChatGPT is not always perfect, and there can be some strange inconsistencies or inaccuracies in the text. If a teacher intends to use AI generated text in some of their lesson materials, they will need to proofread it carefully beforehand. Some other limitations of ChatGPT include its inability to access the Internet (try asking it for the current time in Japan) and its lack of knowledge about world events after late 2021 (try asking it who the current monarch of the UK is). However, further updates to ChatGPT are being regularly released, and it seems almost certain that some of its current limitations will be addressed in the near future.



The recent development and deployment of highly sophisticated LLMs such as ChatGPT has significant implications for language teachers and learners. While some students may be tempted to use these technologies to take shortcuts in homework and coursework, others may be able to use them in ways that genuinely help them to improve their language proficiency. Efforts have already begun to list the ways in which language learners could use ChatGPT productively (Deubelbeiss, 2022; Gravell, 2022; Ryan, 2022). The possibilities for teachers, too, are almost limitless, and many innovative activities and exercises using ChatGPT will surely be forthcoming in 2023 and beyond.



Deubelbeiss, D. (2022, December 19). NLP and language education. Substack.

Elsen-Rooney, M. (2023, January 4). NYC education department blocks ChatGPT on school devices, networks. Chalkbeat.

Gravell, S. (2022, December 18). A Few More Cool AI Things. Wordpress.

Greteman, B. (2022, December 12). ChatGPT can write better essays than my college students. That’s a good thing. Newsweek.

Hern, A. (2022, December 31). AI-assisted plagiarism? ChatGPT bot says it has an answer for that. The Guardian.

Hougham, D. (2021). Making positive use of machine translation for writing essays. The Language Teacher, 45(6). 45–47.

Karpf, D. (2022, December 22). Money will kill ChatGPT’s magic. The Atlantic.

Marche, S. (2022, December 7). The college essay is dead. The Atlantic.

Oda, T. (2020). How to make positive use of machine translation. The Language Teacher, 44(2), 30–32.

Raf. (n.d.). Does ChatGPT remember what happened earlier in the conversation? [Online forum post]. OpenAI.

Ryan, K. (2022, December 12). Bank of GPT3 activity types for language learners. Edge of CALL.

Zhang, M., & Li, J. (2021). A commentary of GPT-3 in MIT Technology Review 2021. Fundamental Research, 1(6), 831–833.