Introduced Species

Artificial Intelligence deployment in schools.

My school district recently bought all of us access to a teacher-facing AI platform for an unknown, but likely large, sum of money. I am starting to worry that we are introducing a species of sorts into the classroom ecosystem, and that predicting the outcome will be difficult.

In June 1935, the Bureau of Sugar Experiment Stations dropped tens of thousands of cane toads into Australia. They were supposed to eat a bunch of beetles that were eating the sugar cane. Cane toads do eat those beetles, in some places, but toads are also famously short little guys, and in Australia the beetles mainly hung out on top of the sugar cane. It also turned out that the cane toads have no natural predators in Australia and are extremely poisonous. The native critters don’t even know they’re poisonous, eat them, and die. This led to the case study for invasive species and subsequent ecological disaster.

In retrospect, the causal chain can be gleaned, it’s not magic, but it can be difficult to predict just how wrong things can go in advance. Why don’t we just kill all of the ticks and mosquitos that are disease vectors? Almost anybody can tell you that something eats ticks and mosquitos, and if you kill their food, you kill them too. An unwanted outcome. Some may go slightly further and say there’s actually something called the food web, and in fact all species in an ecosystem are interdependent, and killing off one node in that web can trigger cascading effects that reach far beyond one predator-prey relationship.

There are more or less blunt instruments available. Scientists have been pretty successful releasing sterile mates for nasty insects like the Screwworm Fly to reduce their population. Much more finesse than just dumping toads from planes. The animal kingdom is not a black box, but we are far from finished mapping its machinations. Living things ultimately do not respect borders, and definitely do not respect the niches we assign them.

I have attended not one, but four professional development sessions about using AI as a teacher this year. The guts of a Large Language Model, much like a living ecosystem, can produce what appears to be unpredictable results. The inner machinery of these models are understood, but their behavior is very difficult to predict exactly. Much like a living community, the contribution of different factors is easier to piece together in reverse than it is to predict in advance.

There are a wealth of essays about the effect generative AI is having on student learning. I have plenty to say about that, but something that gets less attention is teacher use of AI. Full disclosure, I complain about these AI trainings not because I instinctively reject them, but because I am already using it for dozens of time consuming tasks at work. A power-user, if you will. This is not a polemic against AI.

The platform our district purchased is basically a guard-railed wrapper around Anthropic’s Claude, and works as well as any other platform at generating quizzes, lesson plans, accommodations etc. The pitch from the district was basically “Make your job easier, automate the slog”. Fair enough, the main barrier between the teacher I am and the teacher I could be is time. There is not enough of it to do all the things I am supposed to do.

When generative AI models like GPT hit the scene and started to get good, it changed education forever. These models were still problematic and creaky when I was in school, but shortly after, they were robust enough that the task of “assigning an essay” became a quaint gesture. A huge amount of undergraduates report that they use AI for virtually all of their assignments.

At the most recent AI focused professional development meeting, a colleague asked an interesting question.

If the kids are using AI to write, and we’re using AI to grade, why don’t we all just go home and collect a check?

It made me laugh out loud because it rhymed with a famous dirty philosophy “joke”. I beg your forgiveness for this reference, but it is maybe the most apt available. Slavoj Žižek said, and I’ll paraphrase to keep things PG-13: perhaps the ideal physical relationship between two lovers is for both parties to bring their own electronic devices, plug them both in, and leave them to entertain each other. Then:

leave all the fun to this ideal couple, with us, the two real human partners, sitting at a nearby table, drinking tea and calmly enjoying the fact that, without great effort, we have fulfilled our duty to enjoy.

So, why don’t we all just go home and let ChatGPT grade the papers it just wrote? Žižek is proposing that maybe the machines which, on the one hand, threaten authentic human connection, could be properly channeled to actually allow us to attend to real human need. Teaching, hopefully, cannot be distilled down to the assigning of grades. Measuring brains and sorting the future workforce into buckets of variable aptitude. There is something about the actual agency a teacher brings into the room that may be hard to quantify, but is very much real. In fact, the average person doesn’t remember lessons, they remember the people who delivered them. Imagine:

I knew I wanted to be an astronaut when Grok told me it was a high-status, high IQ profession.

Systematically introducing AI tools into the teaching profession is like introducing a species into a troubled ecosystem. With great care, perhaps we can target this species to gobble up the problematic actors in this ecosystem, and will end up with a more harmonious, balanced school. The introduction of the Cane Toad did actually work to control crop pests in some countries. And this ecosystem is, after all, plenty troubled. I’ll save the enumeration of these troubles for other times, but the obvious one is time. Maybe this introduced machine agent will automate all of the time consuming bureaucratic tedium and allow us to finally sit down with a student and ask “how are you doing? do you enjoy this topic?”. Maybe all of the impossible demands teachers try to meet will finally be possible, and the spiritually nourishing aspects of teaching and learning will finally reemerge.

Or, we end up like the Cane Toad in Australia. The intervention runs away from us. We defeat the tedious task pile, but we also defeat ourselves. Maybe, we discover that what appears to be the simple transfer of information from one mind to another, is actually an entangled meeting between two complex beings, and the societal benefits that it yields are from latent interactions that only happen when humans teach other humans. It’s a romantic view of education, but it’s one many people share.

So, the question is then: how well can I be simulated? I can prompt any LLM “explain cosmological redshift in terms a 10th grader can understand. List relevant vocabulary and provide sources for further reading”. It will likely do quite a good job. How will AI perform when faced with a student who sees their teacher, and all authority for that matter, as an extension of their father or mother with whom they have conflict, and therefore will not be wrestling with the mysteries of cosmology today? Any resolution that may occur there is mysterious to me. To the extent that a teacher does help these psychological barriers to learning it is usually difficult to define what combination of words or micro expressions earned their trust.

From the vantage of capital, the school is a limping gazelle. Its inertia is aimed toward deprofessionalization in many states, the field offers unattractive pay, and enterprise licensing is the biggest bag a software company can get.

This makes it an attractive target for ed-tech companies who want to hop on the gravy-train that is “AI enabled”. There are many e-learning platforms out there that are benevolent actors (Khan Academy for example), but I would not say my default stance is trusting concerning the ed-tech landscape. Beyond the tech hype, the entire field of public education is constantly under threat. Many in the United States openly call for its elimination. Privatization is where the wind is blowing in my state.

If we are going to outsource some of the labor of educating our youth to LLMs, it’s important to remember two things.

First: AI models are by design incapable of true novelty. They carefully mimic the data they were trained on. Their output, though impressive, is always the shadow of something that’s already been done. This is great for generating code, summarizing texts, and throwing together a lesson plan when I’m sick, but it’s far from making novel discoveries. I’ve felt many emotions interacting with AI models, but true surprise is not one of them.

Second: the education ecosystem is fragile and not well understood. Universal, free public education is young. The school is a nexus of culture, political economy, labor, intellectual production and more. Trying to explain how anything works is impossible, it’s a miracle we get anywhere close. I get the same feeling of unease when I try to really understand what’s happening inside a school as I get when I read too long about the brain. The nonlinear effects, folded on top of one another, the complexity of it all is staggering.

Maybe that’s a reason to take a breath before the mass deployment of generative AI.