Demystifying AI for Educators
A Guide to Understanding AI in Tertiary Education
In this post I will go over some of the basic terms related to AI, assuming no prior knowledge.
I'll look it from the point of view of students, teaching staff and administrative staff, and the benefits and drawbacks of incorporating it into how you work.
I've met people who work in Education who use it every day, and people who are completely against it. It's a shifting landscape which had a large impact on education on its release, but which is now affecting multiple industries and raising massive societal questions.
We (teaching staff, administrators and students), need to be aware of AI and the questions it raises, even if we don't want to use AI ourselves. So an overview of how these tools work and what they can do it useful.
I write about how I think AI can be used positively in this post How to Leverage AI in your Lifelong Learning Journey. I also wrote about how AI can negatively affect writing skills in Where will Young Writers Learn their Craft in the Age of AI. This is equally relevant for students.
In this post:
- AI Basics for Educators
- Artificial General Intelligence
- What is an AI Agent?
- How to end up in a world of pain with AI
- Ethical Considerations of Using AI
AI Basics for Educators
We constantly hear and read about AI, but what does this actually mean in the context of education. Let's go back to the basics.
AI - Artificial Intelligence - was a term which was coined in 1955 by computer scientist John McCarthy. So the natural question to ask is, this system isn't self-aware, so how can we use the term "intelligence".
The argument was that it was concerned with the outputs of human intelligence, such as writing, making images, driving a car… it's the outputs the term refers to, not the process which is taken to arrive at them.
Let's go through the AI family tree until we arrive at the "AI" which most concerns educators - Large Language Models.
At the top level, there are different types of AI. Some are programmed and some use what is called "machine learning".
Machine learning is where the system learns from data over time without being explicitly programmed by a human. It can learn with a teacher, learn on its own and learn by trial and error. There are different types of machine learning and a "neural network" is one of them.
A "neural network" is inspired by how the human brain works, using layers of interconnected "neurons."
At the core of a neural network, is simple maths.
I borrowed this analogy from the YouTube video from IBM - Neural Networks Explained in 5 minutes.
Say a neural network is looking at whether to go surfing. Let's give it three categories. Good surf (1) as it's true. Far to travel (1 as true), sharks in the water (o as not true). Then you apply what they call "weights" to this. Good surf is important - the weight is 5 our of 5. I don't care how far I have to go - the weight is 3. I'm not worried about sharks - 1.
Multiply these (1x5) + (1x3) +(0x1) - 8. The baseline is 3. We're going surfing.
There are different types of neural network. One type of neural network is a large language model or an LLM. This is the one that concerns us most in education. These are apps like ChatGPT (From the company OpenAI), Google Gemini, Microsoft Copilot and Claude (from the company Anthropic) which can write essays.
A Large Language Model (LLM) can respond to a prompt that you give it and produce text. The window you type into and get answers in an LLM is called the context window. A prompt is your question or guidance for the LLM.
The LLM is an advanced text predictor. It is trained to look at a string of text and calculate the most statistically likely next word - to a degree. They found that when they used the most likely next word all the time, it didn't read as very human. So they added a degree of randomness. (This setting is called "temperature').
Once you write your prompt and get an answer, the system does nothing. It doesn't think in-between these actions. It is a static system.
A term you might have heard is that LLMs "hallucinate". It's going to give you an answer, and so it can make things up. Like, really make up a whole lot of nonsense. This improves as the system uses your data to "ground" its responses. Interestingly, there's some work being done on introducing a degree of awareness of uncertainty into the system.
LLMs like ChatGPT create text, photos, videos… this is what is called "Generative Artificial Intelligence (GAI)". This is different to "Artificial General Intelligence (AGI)".
Artificial General Intelligence
Artificial General Intelligence (AGI) is what a number of companies like Google are trying to achieve at the moment. They are trying to create a system which is as generally capable as a human across all tasks. This is the holy grail that large companies (hyperscalers) are investing billions into.
If you want to see a very interesting documentary about this, I would recommend "The Thinking Game", which you can view on YouTube. It is about Demis Hassabis, the founder of Deep Mind, and the journey towards AGI.
Not so long ago Google DeepMind was trying to figure out how to win at Pacman, now it is doing things like solving intractable problems in biology like protein folding, for which Hassabis won a Nobel Prize.
What is an AI Agent?
So now we know what an LLM is, let's get into that a little and expand on what it can do. I look at it as three levels.
The first level is where you use an LLM for chat. You put in your question, you get an answer.
The second level is what some call "builder mode", although this is an unofficial term. You can add documents, get data from the web, co-work on a piece of work together. You are still involved in guiding it at every step. It increasingly uses your data for "grounding", so that its responses are better.
The third level is an agent. An LLM cannot do anything on its own, so they wrap it in a programme called a harness. You can give an instruction to an agent and it will go off and do it. You can instruct an agent to go off and do something in the real world.
Some of these wee agents have caused absolute mayhem.
How to end up in a world of pain with AI
There are serious data issues to consider.
If you aren't paying for it, you are the product. If you use free versions of models like Google Gemini, your data may be used for training the model. Also, human reviewers can access it. It is generally only Enterprise accounts where your data is safe from training.
One example of this. Google Gemini had a "share" facility, so you could share answers to your prompts with other people. However, these shared results actually got their own webpages. And so the text became publicly available in search results.
Another important point for students. If you use an LLM to check your entirely original essay, say you put it in Google Gemini, this can cause problems. Gemini can ingest it as data, so that if someone checks your essay using Turnitin, it can deliver a "written by AI" result. AI models ingest and start using data extremely quickly.
If others on your class also use AI to write an essay, they can end up being quite similar.
What is the best way to use it to improve learning? Paying to do a course, spending precious time and money, only to degrade your writing and reasoning skills by using AI to do the work for you, doesn't seem to be a good deal.
As these tools get more sophisticated, so does hacking. There are many ways these tools are used for hacking - from writing convincing phishing emails, to more targeted correspondence with a particular person (spear-fishing).
There are a lot of "unknown unknowns". For example - prompt engineering. This is when there is malicious (or sometimes not) code on a website which tries to get your AI app to take actions. AI is getting better at recognising and not acting on this, put this is the tip of the iceberg.
Ethical Considerations of Using AI.
Students are in general very aware of issues with AI. There is a considerable societal backlash against it at the moment for different reasons. Here is a YouTube short of Graduates booing commencement speeches which reference AI (From 404 Media).
The first, is that the creators of these tools have been hitting the narrative hard that they will destroy enormous amounts of jobs. Why they thought leading with this as a story would be good is a mystery. They aren't talking about it as much now because… who knew… people don't particularly like it when their working opportunities are destroyed.
A Large Language Model has been trained on an enormous amount of data. This is an important ethical point - a portion of that data was used by companies such as Meta and ChatGPT without permission.
A live issue is the environmental cost of AI with the building of data centres. AI search uses a significant amount of energy and water - there are no exact figures (there should be!), but a low estimate would be 1.26 millilitres per query (about 5 drops of water). And that doesn't count other sources like upstream power generation. However, we need to be aware that other activities like watching YouTube and Netflix also use considerable water.
The honest answer is that a ChatGPT query consumes somewhere between 3x and 10x more water than a conventional search, depending on the model, the prompt complexity, and what you count. (Source - Article on website ModulEdge - How much water does AI use.)
One of the areas the AI is being incorporated into by states is warfare and surveillance. How do we feel about the products of companies that we use, being involved in these areas? It is morally complex.
There are some competing narratives as to the future direction of AI. Will it be a cheap utility like electricity. Or will certain companies (and countries) be able to choose who uses it and when. There are already example of the US weaponising AI by not allowing countries outside the US to use certain models, for a period.
The Rest of Politics podcast goes in these questions in depth. There is a series on AI which is for paid subscribers.
Paul_OS
There is one final question I'd like to raise here. The use of AI for cheating in academic work is well documented. On the other side of this, how much support should students be given in developing their AI literacy and skills with AI tools? Students will enter a world of work where these tools are increasingly used. What is the best way to support them?