The good, the bad and the ugly (ways of using AI in education)

Should we use AI in education?

There is an ongoing debate on using AI in universities: the question is whether we should allocate budget to LLM tokens for students, for researchers, for administration and so on. After all, why should universities, research bodies that are publicly funded, not take advantage of the “productivity boost” that apparently LLMs can provide? Do we not want public money to be spent efficiently?

What we need to realise is that there are two ways in which this can go: one is good for students, researchers and admin, and not good for AI salesmen; the other is good for AI salesmen and strictly bad for everyone else. The difference is in the incentives: let me explain.

A great way to understand my point is the example given by Cory Doctorow about the use of AI in a (not so) hypothetical radiology laboratory: the radiologists want to read the lab results as accurately as possible, but the lab shareholders do not: they want instead to bill as many consultations as possible to the patients, because they are mainly, if not solely, motivated by profit.

When the AI salesman knocks at the door, they can pitch two systems:

1. The first is a system that reads all the lab results alongside of the radiologists already doing so and flags suspicious files so that, if the radiologists missed it, they go back to it and avoid a mistake. This system, compared to the status quo, does not cost much, but it improves accuracy.
2. The second is a system where 90% of radiologists are fired, AI is the only one reading the lab results, and the remaining 10% of radiologists do nothing but take responsability for the work of AI, which they claim to proofread but can of course not - there is 10 times as much as they could read before!¹

If you are a patient and you have to choose which AI system your local radiology lab implements, I have little doubt that you would go with Option 1. Of course you would: if the deal is to maybe pay 2% more to halve the mistakes in a radiology practice, that is a good deal. Most people would take that deal.

What if you are an AI salesman though? You are not interested in healthcare quality as much as you are in profit, and profit must come to you by the shareholders, who are also interested in money and little else. If that is you, then Option 2 is a much sweeter deal! You get to tell the shareholders “Hey, if you fire 90% of your guys my LLM can² replace them; it will only cost half of their current salaries, and you get to keep the other half!”

What are the options?

The same decision scheme might soon be applied, without public awareness and consent, to higher education and research bodies. Before we look at what is currently being done, let us consider what would be a “good way” of using AI in education. What would Option 1 be in this context?

I am honest, I can not think of many applications of AI to education that I would consider good, at least on the side of students. Some on the side of the lecturer and the TA, though, I can certainly name:

• writing exercise sheets in LaTeX;
• transcribing handwritten notes into LaTeX;
• writing scripts to produce examples/illustrate the use of some software, etc

These might be mundane applications, but typing notes and exercise sheets used to take me hours. Now it takes me minutes to handwrite the exercise and instruct an LLM to transcribe it. Last week I wanted to 3D-plot a parametric surface for my Calculus course: without using an LLM I would have spent I don’t know how long trying to remember how to use Matplotlib, navigating StackExchange and skimming manuals. Instead, it took me two literal minutes to produce a plot, which was of course much neater than what I would have made anyways. This gave my students a better product than what I would have been able to make, and saved me time that I could spend on other teaching duties.

I argue that is good use of technology. What is better, transcribing notes in LaTex and producing quick Python scripts sits very comfortably in the realm of things a free/very cheap LLM can do. This way of using AI in education has a negligible cost. Which is, of course, the reason it is not what AI salesmen are selling to universities.

What are they selling instead? One of the things I am aware of is “AI TAs”: if a student struggles to understand a concept, now certain universities are telling them “here, ask your question to our chatbot”. Aside from sounding incredibly sad, this removes much of the value of a TA from a course. Students are less incentivised to go ask a person, who was a student like them once and has had a decade to understand that concept; they are less incentivised to attend exercise sessions, to actually solve exercises, to actually try and understand concepts. If five years ago asking a question to a TA was guaranteed to give you a solution, it also required that a student put a minimum of effort in their attempt beforehand. To ask a TA for help meant that I had tried to solve the problem myself, and no matter what I tried I had to admit (to myself and to the TA) that I could not. This pushed me to at least try, before asking for a solution. I do not know if this “admitting defeat” thing is the best way to learn, or if it is desirable at all, and maybe for some students it is not. But it gave me a good motivation to actually work through problems, and taught me to be independent. If students are not only capable, but encouraged to ask a chatbot first, what actual incentive are we giving them to try? Why should we expect students to actually struggle and get better if the message we send is “there is no point in trying, the machine has all the answers”?

No matter how little value, if any, this brings the students, the AI salesman are happy to sell this to administrations. Of course they are: 30 students asking the same questions over and over for hours a day consume way more tokens than one lecturer doing OCR + light coding for one hour every week, so of course an AI company offers such a product to faculties. How is it funded? Well, for now it might be “generously” paid for by AI companies budgets, or offered at a heavily discounted rate.³ But in the future, will universities really not be expected to pay way more for these products? And once a budget is tight, and a contract with a LLM provider has to be honoured/renewed, where are the cuts going to be made?

This is where we are now.

The effect, inevitably, will be budget cuts for staff. Remember the radiologists?

After all, who needs to hire PhD students anymore if most of the value they offer to the administration (being TAs) is (read, “appears to be”) cheaply replaced by a chatbot? The salesman knows that this is how the administration thinks, and they know that the fewer PhDs are hired, the more tokens the university will be able (and willing) to pay to them.

Universities are not for-profit instutitions,⁴ and they should not be run like ones. Their purpose is advancing human knowledge and forming generation after generation of critical thinkers, artists and informed citizens. These purposes necessarily have to be centred around human interaction, and have to be led by humans. They cost money, because they are a public service, and should be run to be effective, not cheap.