EdTech Companies Are Racing to Build a GitHub Copilot for Teachers. This Will Not Be Easy.

Generative AI has produced an extremely useful tool for software developers. Can it do the same for teachers?

Here is Sal Khan with a concise summary of the hopes for generative AI in education. He predicts transformation for students and teachers.

But I think we're at the cusp of using AI for probably the biggest positive transformation that education has ever seen. And the way we're going to do that is by giving every student on the planet an artificially intelligent but amazing personal tutor. And we're going to give every teacher on the planet an amazing, artificially intelligent teaching assistant.

I have focused most of my commentary to date on chatbot tutors, trying to describe the actual needs of students and how chatbot tutors can and can’t meet them. Let’s talk about teaching here instead. What impact should we expect generative AI teaching assistants to have on the work of teaching?

Essentially, we’re asking whether or not teachers can have their own version of GitHub Copilot.

What is GitHub Copilot?

GitHub Copilot is a piece of software that developers add on to the environments where they write code. With GitHub Copilot, when you’re writing your code, it’ll autocomplete whatever stubby thing you’ve written and you can accept or reject the autocompletion.

Or you can type a question or prompt as a comment in your code and GitHub Copilot will insert an answer.

Reviews from software developers are generally very effusive, though some developers complain about suggestions that won’t compile or about excessive time spent reconciling the difference between what they wanted and what GitHub Copilot gave them. GitHub’s internal research indicates that software developers feel more productive and complete tasks faster and more successfully with Copilot than without.

A GitHub Copilot for Teachers?

A long list of companies would like to build something similar for teachers. Most of them have created an interface that connects to an existing Large Language Model like ChatGPT so that when you tell the teacher copilot tool you want a lesson plan for a topic, it sends your query along to the LLM wrapped up with some additional prompt engineering. The LLM won’t just tell you some facts about the topic, rather it will structure those facts as a lesson plan or worksheet or whatever else you asked for.

As one example, MagicSchool.ai recently raised $2.4M across several different investors to scale up its teacher copilot tool which will help you create rubrics, lesson plans, and review sheets for your assignments. 

In the image below, I am using MagicSchool’s “conceptual understanding generator” to generate an activity to help my students develop a conceptual understanding of graphs in algebra with an emphasis on sports application.

Jennifer Carolan is the co-founder of Reach Capital and one of MagicSchool’s new investors and she recently described her hopes for this “teacher copilot” space to the Wall Street Journal:

Teachers spend about eight to nine hours on average per week creating content, organizing it, presenting it, getting it ready to deliver to their students, and generative AI has just such great potential in helping teachers do this. And then also in assessment, the ability for generative AI to automatically grade or assess content is also very exciting because that’s another task that teachers spend a lot of time on.

I would love nothing more than for teaching to become much easier for teachers. It would be wonderful if teachers could spend less time outside of their contract hours planning, grading, creating materials, etc. But I think teacher copilot tools face three strong challenges that make it very unlikely they’ll see the usage and success of GitHub Copilot.

Teachers do more kinds of work than software developers.

This isn’t to disparage software developers, rather to point out that the role of “software developer” has been constructed with much more focus than the role of “teacher.” For example, in this survey, more than 75% of a software developer’s working time can be defined by three tasks. Because a software developer’s work is so tightly defined and focused, a single tool can impact a huge percentage of it.

Meanwhile, a 2022 survey of teachers by Merrimack College found that teachers spend their non-teaching time on a much broader distribution of tasks. 

You have to add up the top six tasks to reach 75% of a teacher’s non-teaching time. If you’re developing a teacher copilot tool, you don’t have the luxury of targeting one or two of a teacher’s many jobs. You have to focus on many smaller ones instead if you want the same impact as GitHub Copilot.

Teachers do their work in more places than software developers.

Not only do software developers exhibit more focus in the kinds of work they do, they do much of that work in the same piece of software—their development environment. GitHub Copilot has developed integrations with several of the most popular development environments, including one that is used by 74% of software developers. 

GitHub Copilot can support an enormous percentage of the work that an enormous percentage of software developers perform on a daily basis! Meanwhile, different teachers perform the same task—planning and preparation, for instance—in environments that differ widely between them.

Because of those differences, most teacher copilot tools don’t attach themselves to the environments where teachers work. They ask teachers to come over and attach their work to the copilot. Paste in your teaching objective. Select your desired reading level for the text you pasted. Copy and paste the output back into wherever you make your worksheets, slides, or tests.

That creates a difficult working relationship between you and your teaching assistant. These teacher copilot tools may create less work for the teacher on net but they create more work initially.

Teachers need more than text.

Large Language Models are generated from huge bodies of text. They return text in response to user queries. It is very convenient, then, that software developers do the majority of their work in text—highly-structured text that computers can execute called “code,” but text nonetheless.

Teachers work with lots of text also, especially teachers in the liberal arts. But most teachers need more than text to support their work. For many teachers, the text these teacher copilot tools return is insufficient.

It takes work to turn LLM-generated text into a worksheet or to add graphical models to whatever word problem the teacher copilot generated or even to create blanks where students can answer the questions. In the example about sports above, the teacher copilot has given the teacher some text, but the teacher also needs data sets, illustrations of the different sports, space for answers, etc.

Given the effort involved in taking your request over to the teacher copilot and the effort involved in polishing the text output into a form that is useful to students, we should not be surprised when teachers turn to downloadable worksheets or whatever their colleague down the hall has instead.

Who Will Win (and What Do We Mean by “Win” Anyway)?

If I were building one of these teacher copilot tools, I’d build a plugin that integrated with Google Apps for Education, a tool which has something like 60% of the K-12 market. I’d focus the copilot tightly on the needs of liberal arts teachers—English, language arts, social studies, etc—since they make the most use of text. All the copiloting would happen inline with the teacher’s existing work without any need to visit an external site and engage in roundtrip copy-paste. In the entire space of teacher copilot tools, I’m the most interested in Brisk Teaching, which attempts to do exactly this. (No conflict of interest to report here.)

Is there a successful business here? Maybe. A bunch of these tools are clearly grasping at cash in a fundraising environment where cash is tough to grasp. A bunch of them will merge or close down in the next several months. Among the survivors, I don’t think there is anything close to a GitHub Copilot-sized business, if only because of the challenges I have described above.

Commercial prospects aside, is there actual teacher impact here? I think it’s very possible that the best of these tools will help teachers reclaim an hour or two of time every week. But I wouldn’t disband the teacher’s union just yet. The benefits of increased productivity in post-war America don’t typically accrue to workers. They accrue to management who invests them in more and different kinds of work rather than, say, a four-day work week or shorter work days.

I don’t expect these tools to reverse dismal teacher retention trends either. The most common complaints teachers have about their jobs include salary, benefits, societal respect, and more support staff (p. 59). Generative AI is powerful, but the biggest challenges to teacher retention have political solutions, not technological ones.

What Else?

• The social media landscape in edtech is full of people trying to grab a little piece of the attention economy for themselves, whether as an expert or (ahem) a skeptic. For those reasons, I love tapping into little Reddit threads where everyone is a little more obscure and everyone shares their thoughts with a little less social media gloss.

• “AI Startup Buzz Is Facing a Reality Check,” reports the Wall Street Journal, describing flatlining user growth and even layoffs at some of the buzziest generative AI startups. I am trying to be mature about this.

• I interviewed Kristen Moore about generative AI in math education for the Math Teacher Lounge podcast. She’s more excited about the technology than I am. At one point, I took the list of nineteen high leverage practices from TeachingWorks and told her the ones I didn’t think AI could enhance. She pushed back with some interesting possibilities.

• OpenAI, the company behind ChatGPT, released their Teaching with AI guide. Amanda Bickerstaff, the founder of AI for Education, is unimpressed.

• My colleagues Allison McCulloch and Jennifer Lovett have just released a new book I’m excited to read: Exploring Math with Technology: Practices for Secondary Math Teachers.