LLMs in teaching? Why?

The current issue of Nature has a 'Feature', i.e. not a peer reviewed article, called 'ChatGPT enters the classroom'.¹ While the feature is well-researched journalism and contains a great deal of interest, it is uncritical at key points. I want to look carefully at this paragraph (p.475):

Using LLMs to read and summarize large stretches of text could save students and teachers time and help them to instead focus on discussion and learning. ChatGPT’s ability to lucidly discuss nearly any topic raises the prospect of using LLMs to create a personalized, conversational educational experience. Some educators see them as potential ‘thought partners’ that might cost less than a human tutor and — unlike people — are always available.

The first claim is:

Using LLMs to read and summarize large stretches of text could save students and teachers time

So if students and teachers are spending time summarising texts now, we need to ask first why they are doing that? While it might be because the texts are badly written and can be summarised without loss, I am going to set that aside because it only suggests a use for LLMs in publishing, not in teaching and learning. Similarly, LLMs could be used to write abstracts for texts which do not have abstracts (conceived here as providing the information needed to decide whether you should read the text or not), but that too is a publishing task.

Teachers might summarise 'large stretches of text' because they want to extract the important points for their students. What is the criterion of importance here? The learning outcomes. How is importance evaluated? By the teacher's knowledge and understanding. An LLM will summarise the text without exercising judgement, including in the summary points that are irrelevant or unhelpful (all texts include those!) and excluding some details the teacher may have thought important. Given the fact that when a teacher summarises a text they exercise their judgement, if they left an LLM to do the job, they would still need to read and understand the text and check that the summary met their specific pedagogic requirements. That doesn't sound like much of a time-saver to me.

Students might summarise texts as part of their note-taking. In fact, we sometimes set it as an exercise, because it is an important skill. Why is it important? Because it doesn't merely demonstrate understanding, it is part of the process of coming to understand. If we want to teach students to understand complex ideas, we have to get them to read those ideas and 'summarise' them in their own words. If they just read the summary, they get at best an illusion of understanding. Using LLMs would save time but impede learning in many cases.

ChatGPT’s ability to lucidly discuss nearly any topic

'Lucidly'? It provides grammatically correct, syntactically sophisticated, structured responses to questions. Those are formal properties. Do they amount to lucidity? Of course not: there can be grammatically correct, syntactically sophisticated, structured nonsense. Ask Lewis Carroll. In fact, what this conflation of form and content does is weaken a student's critical abilities, because it teaches them to trust superficial markers of understanding and knowledge, the superficial markers highly valorised in elite education, over careful thought about the content of what has been said. It teaches them to fall for intellectual con artists.

Some educators see them as potential ‘thought partners’ that might cost less than a human tutor

They might cost less than a human tutor, at least they might have a lower cost at point of use, but they use up much more energy. We should not forget that educated, informed people may be slower than LLMs to answer questions, but they are considerably more energy efficient. How many calories has my brain used writing this sentence? If it took 30 seconds on 2000kcl/day, then ~0.6kcal. And how much energy would ChatGPT4 use? We don't know, but estimates are that it uses 50-90x more energy than a traditional Google search ,² and between 3x and 7x as much energy as a human (who is also supporting other useful, and energy intensive, functions at the same time).

They may be some future in which the energy and environmental costs of LLMs are comparable to human tutors (well, the marginal costs amyway), but rather than spend billions reaching that point, why don't we spend the money training more teachers? We know how much (or little) that costs and how effective it is. We shouldn't buy into this fantasy in which humans are replaced by machines without asking whether we could have achieved the same outcomes without the machines but with a better politics instead.

“One-on-one tutoring is the single most effective intervention for teaching, but it’s very expensive and not scalable,” says Theodore Gray, co-founder of Wolfram Research, a technology company in Champaign, Illinois. “People have tried software, and it generally doesn’t work very well. There’s now a real possibility that one could make educational software that works.” Gray told Nature that Wolfram Research is currently working on an LLM-based tutor but gave few details.

That is a perfect example of answering the wrong question. Wolfram Research clearly has a lot of capital to invest in AI and they are using this to create AI tutors which - if successful - will be sold to governments around the world. I.e. they want governments, especially in poorer countries, to spend their limited educational budgets on AI rather than paying more teachers. And they are trying to persuade us that this will ’ameliorate the human condition’ because the AI will provide what we can't afford now. But they haven't told us if we can afford the AI yet.

The right question to ask is: Why aren’t we investing that capital in training and employing more teachers around the world?

Doing that would create millions of fulfilling jobs and stimulating classrooms, and we know it improves learning. It would also mean the money was spent in the local economy where the children lived, keeping that economy buoyant and giving the children a brighter future. As opposed to extracting that wealth from those countries to the countries where the owners of Wolfram Research live.