AI: The State of Play in Language Teaching in 2024

Understanding ChatGPT: What’s Under the Hood?

The key to understanding this is to first understand the acronym GPT:

1. Generative: This refers to the model's ability to generate text. Text that is coherent and contextually relevant to what the person types in the input field.
2. Pre-Trained: Before being released ChatGPT was trained on a large dataset - sitting at hundreds of billions of words. To put this in context, a commonly used corpus ‘The British National Corpus’ comes in at just under 100 million words.
3. Transformer: This is a type of neural architecture used in machine learning. This feature enables these tools to recognize the context and relationships among words.

Think of ChatGPT as a version of Google's autocomplete for conversation, building a response word by word. Unlike autocomplete, which is limited to sentences, ChatGPT has the capacity to write at a higher level, including essays and beyond.

Wolfram (2023) sums it up nicely here:

It’s [ChatGPT] just saying things that “sound right” based on what things “sounded like” in its training material.

When it’s writing an essay, it is asking over and over again ‘given the text so far, what should the next word be?' — then adding another word. Keeping in mind these predictions are based on the scraping of billions of webpages, this process highlights the model's reliance on existing data rather than an understanding of the text.

[Scraping sounds more pejorative than training, but maybe it’s a better word to explain what happened. The data 'scrapped' by these language learning models could be considered the greatest theft of all time... but that's for another article]

To sum up, ChatGPT is like a corpus on steroids. For a thorough understanding of ChatGPT's workings, I recommend reading Wolfram's article (which happens to be 18,777 words long, so you can always ask ChatGPT for a summary!). If visualizations are more your style, this visual walkthrough by the Guardian is excellent.

Now we have an idea of how it works, let's move on to what it can do.

AI in language teaching: What can it do?

According to Ethan and Lilach Mollick (2023), ChatGPT is capable of performing the following tasks:

1. Write text at the level of a human college student.
2. Conduct online research and produce reports with accurate outside references.
3. See images and interpret and work with that information.
4. Read documents and PDFs, and answer questions about the content.
5. Produce realistic images.
6. Do math, code, and solve problems and,
7. Explain concepts, correct mistakes, provide feedback, and more.

A teacher could find all of these general functions to be helpful. More specifically in the language classroom, AI can excel at role-playing conversations in a second language, building quizzes, tests, and lessons from curriculum materials, and reducing friction for non-English speakers. For example, it can grade a newspaper article so it’s readable to a beginner student of English.

Speaking of understanding, it's time to add a few more terms to your AI lexicon.

Key Terms in AI: A Glossary for Educators

In this part, I'll define significant terms that have deepened my understanding of LLMs and may do the same for you.

Prompting

Prompting is in a state of flux. As soon as I write something about prompting, it's outdated before I even finish writing the sentence. But, putting that aside, let’s look at where prompting was, where it is now and where it's going in the future.

It's highly likely that prompt engineering will be one of the shortest-lived professions in history. One year ago, when AI was first introduced, it required carefully constructed input statements or questions in a specific format to elicit an accurate and detailed response.

Although structured prompts still hold value, they are no longer a requirement. Despite this, it remains valuable to comprehend the elements that make up a good prompt at this point of time. In Mollick's (2023)opinion, a good prompt should incorporate the following features:

1. Role and Goal: By implementing role and goal-based constraints, the AI is restricted to a more relevant set of responses, utilizing its natural language understanding abilities. Give the AI your teacher contexts and clear goals for the task you would like it to complete.

2. Step-by-step instructions: Simpler instructions have a higher chance of being understood by the AI, making accuracy more likely. Deliver your instructions to the LLMs like you do to your students. Include sequencing language and examples.

3. Expertise or Pedagogy: This is where the AI's expertise is determined. When teaching a language, AI could be trained to respond through the lens of the Dogme approach. You could also consider training AI to evaluate essays using the IELTS criteria.

4. Constraints: These are the rules or conditions that guide the behaviour of the AI in its interactions with the user. For example ‘Wait for the student to respond.’

5. Personalisation: Prompts that help the AI get more information from the user.

Interested in trying out structured prompting in education? Head to this post on OpenAI's blog for some helpful examples.

The key to structured prompting is providing the AI with context. Mollick (2023) describes it like this:

You can (inaccurately but usefully) imagine the AIs knowledge as huge cloud. In one corner of that cloud the AI answers only in Shakespearean sonnets, in another it answers as a mortgage broker, in a third it draws mostly on mathematical formulas from high school textbooks. By default, the AI gives you answers from the center of the cloud, the most likely answers to the question for the average person. You can, by providing context, push the AI to a more interesting corner of its knowledge, resulting in you getting more unique answers that might better fit your questions.

Imagine if you wanted the AI to act as an IELTS examiner. In this case, you might upload the IELTS marking criteria for writing and include a range of marked essays as well. Also, you could state the type of feedback and return format you prefer. Furthermore, you can tell the AI they are an expert, friendly IELTS examiner who helps students with complex topics (or have a bit of fun and say they are an expert teacher with the personality of Terence Fletcher from Whiplash!)

This will lead AI to the specific area of its knowledge that focuses on IELTS evaluation.

Circling back to my initial statement in this section, it's important to recognize that prompting is in a temporary state of affairs. As AI increasingly gets better the need for engineered prompts decreases.

Case in point, DALL E3. In ChatGPT, you can now use conversational prompts to create images. Rather than engineering overly complex prompts to get the image you want, you just talk to the AI about the art you have in mind. The prompts are then produced by the AI.

Currently, using conversational prompts can likely help you achieve about 80% of what prompt crafting can do.

But the gap is closing…

One last point on prompting, interacting with ChatGPT at times can be a bit weird. Massaging your prompt in certain ways can get better results. Language learner, Mathias Barra (2023), shares some of the weirdness in his blog post ‘A few strange ways to get better answers from AI’. Phrases like ‘Take a deep breath and think step by step’ and ‘You would be a lifesaver if you could do this’ can get better results. Also using ALLCAPS or generating the response multiple times can get well-crafted answers.

The next term I want to define is agent.

Agents

How do AI agents differ from prompting?

The key difference between the prompting I highlighted before is that AI agents can think and function independently, while in the previous examples humans are always initiating the process. With AI agents you only need to provide a goal, and they'll create a task list and start working on the project autonomously.

Bill Gates (2023)sees a future where agents are:

Capable of making suggestions before you ask for them. They accomplish tasks across applications. They improve over time because they remember your activities and recognize intent and patterns in your behavior. Based on this information, they offer to provide what they think you need, although you will always make the final decisions.

This seems would be a powerful tool to have, but there is still a computer-generated elephant in the room... hallucinations. Are we really willing to unleash autonomous agents into the world or language classrooms, despite their tendency to create false information?

Hallucinations

ChatGPT and other large language models (LLMs) have the unfortunate habit of telling you incorrect, misleading, or nonsensical information. Maggie Appleton provides a great analogy for what is taking place:

In ways, it’s like a terribly smart person on some mild drugs. They're confused about who they are and where they are, but they're still super knowledgeable.

Now this is a problem. Particularly if you're a teacher or student who doesn’t like to check their facts…

Additionally, Appleton highlights the drawback of these models being solely focused on language. This is a problem because language is only a small part of how a human understands and processes the world.

LLMs resemble lobotomized individuals. While LLMs do have language, they are missing crucial aspects of human interaction with the world, including spatial reasoning, time perception, touch, taste, and memory.

All that's left are words.

Furthermore, the words used to populate the data of these LLMs are very monoculture.

They primarily represent the generalised views of a majority English-speaking, western population who have written a lot on Reddit and lived between about 1900 and 2023.

Maggie Appleton

As we have seen before, an increase in autonomy for LLMs could result in the introduction of hallucinations without human supervision. That's your elephant, but there are strategies available to minimize the risks of hallucinations in the 'Jagged Frontier'.

The Jagged Frontier

No one really knows the best ways to use AI or how and why they fail. While AI can excel at offering feedback on writing, it can fall short on seemingly simple tasks like constructing anagrams, frequently omitting necessary letters.

To describe this phenomenon Ethan Mollick (2023) coined the term ‘the jagged frontier’.

Imagine a fortress wall, with some towers and battlements jutting out into the countryside, while others fold back towards the center of the castle. That wall is the capability of AI, and the further from the center, the harder the task. Everything inside the wall can be done by the AI, everything outside is hard for the AI to do.

Mollick suggests in your industry - language teaching - to use AI a lot. For example, planning a lesson - use AI; giving feedback - using AI; grading  writing - use AI.

I think you catch my drift.

As you progress, the contours of the 'the jagged frontier' in language education will start to take shape. You will start to understand what AI is scarily good at and also where you shouldn’t use it.

Appleton believes AI works as a skill leveler. Creating higher floors and ceilings. Why not for teaching  too?

But a word of warning: don’t fall asleep at the wheel! If you stop paying attention to what AI is doing then you might fall victim to those pesky hallucinations.

A better way to work with AI might be in a  ‘centaur’ or ‘cyborg’ type relationship.

Centaurs and Cyborgs: Integrating AI in Teaching

Centaurs and Cyborgs are two ways you can work with AI. They both keep the ‘human in the loop’ and mitigate the risks of hallucinations, but in two different key ways.

The division of labor for centaurs is explained by Mollick (2023) as follows. A centaur - like the mythical creature half human half horse - keeps a clear line between person and machine. There is a strategic division of labor; With the human doing the work they are strongest in themselves, and then giving tasks off to AI based on what it does best in the 'jagged frontier'.

While on the other hand, the deep integration of technology and humans defines cyborgs. Humans in this sense don’t delegate tasks. They switch back and forth over the 'jagged frontier', working in tandem with the AI.

To clarify let’s apply these two concepts to planning lessons in ESL. If we are using a centaur-like model, the curriculum and lesson plans would be developed by the teacher, with AI supplementing the learning experience by offering student-specific resources, tasks, or games.

On the flip side, a teacher who adopts a cyborg mindset with AI would integrate AI into the planning process, employing advanced algorithms to evaluate student data and generate personalized lesson components.

When it comes to my relationship with AI, I tend to gravitate towards a centaur approach. At this stage of AI's development I think it's better to use 'teacher in the loop' systems rather than delegate too much to AI.  Appleton builds on this and says dealing with AI we should:

1. Protect human agency,
2. Treat models as tiny reasoning engines, not sources of truth and,
3. Augment our cognitive abilities but not replace them.

I tend to agree.

Traversing the ELT Jagged Frontier

Ethan Mollick proposes that you spend at least ten-hours using AI in your day-to-day tasks. He says ‘Just use it and see where it takes you’.

So, that’s what I’m going to do.

For the next half-year, I'll be immersing myself in the 'jagged frontier' of ESL teaching, channeling my inner centaur.

The series called 'Centaur Elters' will look at lesson planning, language output, vocabulary, grammar patterns, and feedback.

The content will be a mix of articles and video walk-throughs. [Remember to subscribe to be notified when I release each new post].

By scouting the limits of LLMs in language teaching I hope it will guide you where you can and can't use AI in your own teaching.

Before you 'Go', let's return to 2016.