Three Tests AI Needs to Pass Before It Can Start to Transform Teaching

ChatGPT, a chat bot powered by artificial intelligence released in late 2022, continues to transfix the internet. Ask it to write you some code. It will oblige. Ask it to write you some legal boilerplate. It's yours. Major search engines are rushing to merge its technology.

Some technologists have argued that innovations in artificial intelligence like ChatGPT are increasing fast enough in quality, speed, and scale that they may soon replace teachers, or at least transform the work of teaching. 

Though I work in education technology, I’m skeptical. The more classrooms I observe and teach, the more I understand the process of teaching to be too complex across too many facets to reduce to a transfer of questions and information between a teacher (human or AI) and a learner.

I know I’m not alone in this skepticism. And it’s to my fellow skeptics that I want to pose the question:

What hurdles would AI have to clear in order for you to believe that it had, if not eliminated a teacher’s job, at least transformed it so as to be unrecognizable?

I have three proposals here.

1. AI needs to produce 99% accuracy across 99% of the questions asked in a curriculum.

In mathematics, ChatGPT can currently answer with 100% accuracy only a very narrow selection of questions from a given curriculum: the kind with highly structured inputs and outputs and no surprises.

Ask ChatGPT, “What is the solution to 2x - 3 = 6?” and it offers an answer that is correct and well justified:

Ask ChatGPT a question that is more difficult by a single term, “What is the solution to 2x - 3 = 5x + 6?” (a question that middle school students routinely answer in the United States), and it falls apart:

It combines the proof diction of a PhD (“therefore!”) with an approach to algebraic operations that’s less thoughtful than a kitchen blender. ChatGPT transforms math teaching here only insofar as it gives math teachers a new source of interesting errors for their students to analyze and learn from. 

Once we broaden our scope beyond highly structured problems like these ones toward problems where students have to answer interpretive questions about multiple connected representations of data, it’s clear that ChatGPT doesn’t yet deserve the trust and authority that even a novice teacher enjoys.

2. AI needs to understand and build from your incorrect answer, not just model a correct answer.

It’s one task for an AI chatbot to offer an explanation for finding the solution to the equation 2x - 3 = 6. But it is another task entirely, and a much harder one, to understand the different routes students might take toward understanding that question and its answer, including the many partially correct answers. This is called pedagogical content knowledge, as defined by Schulman, Grossman, and many others, a special kind of knowledge that AI hasn’t yet mapped.

For example, if you ask a teacher of even moderate skill, “Why is the solution to 2x - 3 = 6 not x = 1.5? What did I do wrong?” they’ll recognize that the student may have subtracted 3 from 6 and then divided by 2, rather than undoing the subtraction by adding a 3 to both sides of the equation. That pedagogical content knowledge helps the teacher connect the student’s path to other paths, including one that leads to a correct answer. When I asked that same question of ChatGPT, though, I received the following response:

Essentially, ChatGPT has said, “You were wrong because you weren’t careful. Here is how I would do it.” No part of ChatGPT’s response refers to the ways I might have been thinking about the equation. No part indicates the sense behind my thinking (I performed the same operation to both sides of the equation), leaving me to believe I am 100% incorrect. 

All of this means that if I am to learn from ChatGPT’s response, I need to draw connections from my thinking to expert thinking on my own, as a novice. This is a cognitive tax, one that skilled teachers and tutors generally pay on behalf of learners.

3. AI needs to pass the “Caring Test.”

An AI has passed the “Turing Test” if it can have a conversation with a human that the human finds indistinguishable from another human being. An AI has passed the “Caring Test” (let’s call it) if the human wants to continue that conversation another day.

This kind of social affinity is important for learning. Researchers have found that a student’s belief that they are taking a socially relevant action creates conditions for greater learning and engagement. If you think you are interacting with a human, someone with whom you feel kinship, you’re more thoughtful about your explanations and put more effort into their preparation.

Even if I know the AI will give me consistently correct answers to questions I ask, how much do I care that it thinks my own answers are correct? A whole raft of academic objectives essentially ask, “Can you persuade other people?” Do students care that they have persuaded ChatGPT? Before we can say AI has transformed teaching or tutoring, it needs to convince students that their interactions are socially meaningful.

Obviously AI hasn’t yet cleared any of these hurdles in education to say nothing of clearing them all at the same time. It’s possible that teaching is so irreducibly complex that AI may simply hop into a long line of technologies—including television, the radio, the graphing calculator, and even the printing press–that contemporaries believed meant the end of education as they knew it then.

Rather than wonder “Will AI replace teachers?” educators and technologists might instead wonder how teachers and artificial intelligence can partner in productive ways, for example with AI offering recommendations to teachers that grow more relevant and more accurate over time. If past technologies are any indication, this hybrid model of human and machine is likely to retain the best of teaching’s past while also pointing towards a happier, more productive future.