Leon Furze

Chances are slim by now that you haven’t heard of ChatGPT. Every major news outlet from the Guardian to the Herald Sun has run articles on OpenAI’s latest offering, and it has featured heavily on social media newsfeeds across Twitter, LinkedIn, and Facebook.

I’ve written a lot of posts lately about ChatGPT in education, and I’m not alone. Others have pulled together great content to help teachers get their heads around ChatGPT, like Nick Jackson’s great free course for teachers.

But sometimes, amidst the hype, it’s good to take a step back. I live in a self-imposed online bubble where it’s easy to assume (like I did at the start of this post) that everyone has heard of ChatGPT, experimented, and drawn their own conclusions about the potential risks and opportunities in education. But that’s not true at all.

This post is a “back to basics” for anyone who is struggling to keep up with the rapid changes happening with AI – and specifically ChatGPT – right now.

If you’re more interested in a deeper exploration of ChatGPT, check out some of my earlier posts like this one. I’ll put a list of links at the end of this post.

I built an AI Section C generator and the world did not end

What is AI?

The term “Artificial Intelligence” was coined by John McCarthy for a 1956 Dartmouth College conference. Early work in AI was based in computer science, but also influenced by philosophy, mathematics, literature and history. The term “Artificial Intelligence” refers broadly to machine intelligence that is comparable to that of a human.

It’s a term that’s been contested since its inception. After the first wave of hype in the 50s-70s, many of the grand claims of AI being able to match human intelligence fell flat. Early AI was largely built on “expert systems” models that could specialise in one area but relied on vast amounts of human-entered rules and data.

The lack of “intelligence” in AI has led to a fork in definitions: Narrow or “weak” AI, and Artificial General Intelligence (AGI) or “strong” AI. Strong AI is the kind we see in the movies and science fiction: it’s Hal from 2001: A Space Odyssey, or the sentient machines in Isaac Asimov’s 1950 collection I, Robot. Strong AI does not exist outside of fiction, and though many people are pushing towards its creation there’s no sign of it just yet.

Narrow AI, on the other hand, is everywhere. The term covers everything from the algorithms that govern your newsfeed to the systems that are already driving autonomous cars. Narrow AI also covers a specific type of AI which has spurred the recent media frenzy: Large Language Models.

What is a Large Language Model?

A Large Language Model (LLM) is a type of narrow AI that has been trained on a huge amount of written text. These models are able to generate natural-sounding text, complete sentences or paragraphs, and even write coherently structured articles or creative writing.

Like many types of AI, the more data these models have, the better they get. In recent years we have seen models trained on “scrapes” of vast chunks of the internet. GPT-3, for example, includes a dataset named the Common Crawl, which includes billions of webpages. The process of training a model involves using a set of labeled data to adjust the model’s parameters so it can accurately predict the labels for new, unseen data.

After a model is trained on a dataset, it can be used to make predictions or generate outputs based on new input data. This means that the model can function like a powerful “autocomplete”, like the predictive text on a phone. More than that, the model can create entirely new text based on just a few words. The text the user enters is the “prompt”.

What is ChatGPT?

ChatGPT is a chatbot built on top of OpenAI’s GPT-3 LLM. It is basically a “user interface” that makes interacting with the LLM more friendly than trying to use the model directly. ChatGPT has also been “fine tuned”, meaning it has been refined with additional, more specific data and training methods to reinforce the correct kinds of responses.

ChatGPT launched late in 2022 and at the time of writing is free to access. This “free to access” element has made it incredibly popular. It has also given OpenAI a vast amount of free data from its users to continue fine tuning and developing the model.

ChatGPT works by passing any prompt entered into the chat window to the GPT-3 model, and returning the output from the chatbot. Because it is designed to be conversational, it is possible to have a “dialogue” with ChatGPT, like this:

However, because it is also a window into the powerful GPT-3 LLM, it’s possible to do any number of other things, such as generating code, brainstorming, or creating spreadsheets and tables:

What are the implications for education?

The impact of ChatGPT in education is already being felt. From district-wide bans in NYC to top universities in Australia reverting to pen-and-paper exams, there have been some strong reactions to the technology.

This is because of ChatGPT’s capacity to generate sophisticated, cohesive responses to a number of familiar assignment styles, including essays and short answer questions.

There is a widespread fear – fuelled by the media coverage – that ChatGPT will be used as a “cheating machine”, particularly by secondary and tertiary students. Some commentators are more optimistic, claiming that ChatGPT and other Large Language Models might bring about an end to the worst aspects of our education system.

The truth is, we have little idea of the impact the technology will have on education. Students haven’t started back yet in most places, and very little of the news coverage has included their opinions. Some states are still deciding whether to ban the technology outright, while others try to grapple with the ethical and academic implications of permitting its use.

ChatGPT’s Terms and Conditions prohibit people under 18 from signing up, precluding the use in class by secondary students. However, there are many ways teachers might use ChatGPT (I’ve covered a few of my ideas about that in the posts linked at the bottom of this article), and it is almost certain that many students will be using the technology. This means that one of the biggest factors in education should be the discussion of the technology’s ethical and appropriate use.

What are the ethical implications?

The most prominent ethical aspect of the technology being discussed in the media is the use of ChatGPT for cheating on assignments. Whether or not ChatGPT is going to turn our students into compulsive plagiarists is still up for debate, but there are far more concerning ethical conundrums facing AI as a whole.

Algorithmic bias: Probably one of the most reported on ethical issues with AI (other than cheating) is the capacity of large language models to produce biased, discriminatory, or harmful text. The datasets are scraped unceremoniously from all corners of the internet, and in many ways reflect the worst of humanity’s online presence.

This means that racist, sexist, homophobic, ableist and other discriminatory language makes its way into the model and is then produced as output. Efforts are underway to curb this, but it is a slow and painful process.

Environmental impact: The environmental impact of training and running large language models like ChatGPT is a significant concern. The process of training these models requires a large amount of computational power, which in turn requires a significant amount of energy. This energy consumption results in a significant carbon footprint, contributing to climate change.

The energy consumption occurs in two stages: the training stage and the inference stage. During the training stage, the model is fed with large amounts of data and the parameters of the model are adjusted to minimise the error. This process requires powerful hardware and large amounts of electricity to run the processors and cool the servers. The energy consumption during the training stage can be substantial, depending on the size of the model and the amount of data used for training.

Power dynamics: Less talked about but still important, is the way that Large Language Models can reinforce and exacerbate existing power dynamics. The power structures reflected in the datasets become encoded in the models, meaning that any output reinforces those structures. Because much of the data in the training set is produced from a white, male, English-speaking perspective, the data is heavily skewed to reflect those structures.

For example, if a model is trained on a dataset that is heavily skewed towards texts written in English, it may not perform as well when generating text in other languages, or when dealing with cultural concepts that are not present in the training data. This can lead to a reinforcement of the dominance of English and Western cultural norms.

Data and privacy: As mentioned earlier, AI is reliant on huge amounts of data. Some of this data is personal or private, and has been scraped indiscriminately along with everything else. Earlier models were shown to output private and identifying information from people in the dataset. Additionally, as models become more and more powerful there is an imperative to supply them with increasing volumes of data. This creates an economic imperative for the LLM creators to collect data in ways which are sometimes unethical or invade people’s privacy. There are already lawsuits underway over the copyright infringement of AI text-to-image generators like Stable Diffusion and Midjourney – it might not be long before we see similar suits against LLMs.

Entering text into ChatGPT also adds that input to the data, meaning that every conversation with the chatbot is stored and potentially can become part of future iterations of the model. Teachers and students using the technology should be wary of the data they provide.

TL;DR:

By way of example, and to provide an easy overview of this post, I copied the whole thing into ChatGPT and requested a dot-point summary. Here it is:

• “Artificial Intelligence” (AI) refers to the simulation of human intelligence in machines that are programmed to think and learn like humans.
• “Large Language Model” (LLM) is a type of AI that is trained on vast amounts of text data to generate human-like language.
• ChatGPT is a specific LLM developed by OpenAI that is trained to generate natural language text, and it can be used for a variety of tasks such as text completion, translation, and conversation generation.
• The impact of ChatGPT in education is already being felt with some strong reactions to the technology from district-wide bans to pen-and-paper exams
• ChatGPT’s ability to generate sophisticated, cohesive responses to assignments raises concerns about cheating and ethical implications
• Widespread fear about ChatGPT’s impact on education fueled by media coverage
• Little is known about the true impact of the technology on education as students haven’t started back yet in most places and opinions are not yet available
• ChatGPT’s terms and conditions prohibit people under 18 from signing up, precluding the use in class by secondary students
• ChatGPT can be used by teachers in various ways, but it is important to discuss the technology’s ethical and appropriate use.
• Ethical concerns include algorithmic bias, the environmental impact of LLMs, reinforcement of power structures, and concerns around privacy and data security.

Here’s a list of my earlier posts about ChatGPT which go into many of these aspects in more depth:

I built a Section C generator and the world did not end
Beyond Efficiency: AI can be more than just “Edtech 2.0”
Socrates Against The Machine: Can Looking Back Help Us to Think About the Future of Education?
Prompt Whispering: Getting better results from ChatGPT
Can an AI critique human writing?
A New Level of AI Essay

Or click here to go to all posts tagged with “ChatGPT”

Whatever happens next, there are plenty of educators and academics out there grappling with the implications of LLMs in education. If you’ve got a question, concern, or want to talk directly about ChatGPT and other AI, then get in touch: