MIT’s Justin Reich explains why schools should be wary of unproven AI initiatives, rethink how they adopt new technologies, and focus on evidence over enthusiasm.
INTERVIEW | by Victor Rivero
In a field often driven by promises of transformation, Justin Reich has built a reputation for demanding evidence. As an associate professor at MIT and director of the MIT Teaching Systems Lab, Reich has spent years examining how schools respond to technological change—not just what new tools can do, but whether they actually improve teaching and learning.
At a moment when generative AI is inspiring both excitement and anxiety across education, Reich has emerged as one of the field’s most thoughtful voices on the gap between innovation and impact. He challenges popular assumptions about AI literacy, questions whether technology truly saves educators time, and argues that schools should approach new technologies with more humility, experimentation, and a clearer understanding of what actually works..
In this conversation with EdTech Digest, Reich discusses why AI may be an “arrival technology” that schools cannot simply choose to adopt or reject, why curriculum additions should require corresponding subtractions, and why the most important question facing education today may not be what schools should do about AI—but how they can learn, through evidence, what they should do next.
Dr. Justin Reich is an Associate Professor of Digital Media at the Massachusetts Institute of Technology (MIT) and the Director of the MIT Teaching Systems Lab. As a prominent learning scientist, his work centers on educational technology, learning at scale, teacher professional development, and the future of schooling in a networked world. Rather than adopting a hype-driven view of technology, Reich is well-known for his pragmatic and critical perspective on how digital tools impact classrooms. Connect with Justin on LinkedIn.
You’ve argued that “AI literacy” risks becoming the latest version of the old “21st Century Skills” cycle—a shared responsibility that nobody is actually resourced or incentivized to teach. What would it take to avoid repeating that cycle? What structural changes would schools, districts, or policymakers need to make for AI literacy to become real practice rather than another aspirational slogan?
Here are three proposed principles for introducing something new into education curriculum:
1. Only ask educators to teach new topics for which there are curriculum materials
2. Only ask educators to teach with curriculum materials that we’re pretty sure work
3. Only add things to the curriculum when we take other things away
Policy guidance is a kind of rhetorical argument about values and resource allocation: What if every time policy entrepreneurs published frameworks without teaching materials, we simply laughed them off the stage for being impossibly poorly prepared to make their case? I’m doing some research right now about AI policy development, and one of our findings is that some (maybe most) proposed lesson ideas in AI policy guidance have never actually been taught in a classroom. I think if folks in state agencies or on AI policy panels actually had to teach their own guidelines, the resulting efforts would look quite different.
‘Only add things to the curriculum when we take other things away.’
To take the idea a step further, what if state policy makers not only had to partner with classroom educators to actually teach these lessons, but they also had to gather some kind of evidence that they work as intended. One of the throughlines of the AI literacy argument is something like, “if we teach students domain independent knowledge and skills about using AI, they will be more proficient at using AI in their life.” But there is basically no evidence for that argument—no one has shown that teaching something called AI literacy actually leads to better educational or life outcomes. We do know, historically, that previous efforts to address new social problems through educational interventions sometimes don’t work. After the advent of the Web, we taught tens of millions of children provably ineffective strategies for sorting truth from fiction on the Web. DARE was a widely scaled program to prevent youth drug abuse that had no effect. It’s very easy in education to create programs that look very sensible and do not work at all.
While I’m very skeptical of most current efforts to define what we should teach about large language models, I’m happy to concede that at some point, I’m sure we’ll have some evidence that certain teaching approaches really do help people in their lives, wellbeing, citizenship, work, etc. I’m all for teaching topics that have a good evidence base, but the U.S. curriculum is absolutely overstuffed to the gills. Anyone who is serious about adding new stuff to the curriculum needs to be able to simultaneously identify things we can stop teaching. Imagine if every state pushing for new AI curriculum had to do two things: 1) estimate the amount of time per grade or subject that it would take to meet the proposed goals, and 2) propose an equivalent amount of curriculum material that teachers can stop teaching.
I don’t know if those three innovations- pairing guidance with instructional materials, only recommending guidance and instructional materials with evidence, and requiring subtraction alongside additions to the curriculum—would work. But I know that our playbook for introducing technologies into schools over the last two decades has not worked very well, so it’s worth considering alternatives.
You’ve described generative AI as an “arrival technology,” not an “adoption technology.” If AI is already embedded in students’ and teachers’ lives before schools can formally evaluate it, how should educational leadership respond? What kinds of governance or institutional approaches make sense in a world where the technology arrives first and policy follows later?
We don’t know. For many, many things related to generative AI and education, we don’t know.
In an article about AI as an arrival technology, we offered four suggestions of stances to take. One of the first was to adopt a stance of humility. When you don’t know what to do or are moving forward from a place of limited evidence, be upfront about that with all of your stakeholders.
‘When you don’t know what to do or are moving forward from a place of limited evidence, be upfront about that with all of your stakeholders.’
Schools might need to start keeping “rainy day funds” of resources and staff time in order to deal with emergent challenges or opportunities (Maybe we tax technologies companies to create these funds). Schools may need to create new protocols for “adoption-on-the-fly,” or using criteria from procurement to evaluating new technologies that arrive rather than being procured. But those are just guesses that I came up with that could absolutely be wrong.
One day, I do think big science will have answers to these questions, and the education research community will have some reasonably ideas for how to manage arrival technologies. In the meanwhile, we need local science. Local science involves school and district level folks trying new policies out as deliberate experiments, where they try to collect data about what’s work and what’s not, and where they prune flawed ideas and nurture ones that seem to work.
A lot of edtech rhetoric still frames AI as a time-saving tool for teachers. But you’ve been skeptical that technology meaningfully reduces educator workload over time. What do you think the field keeps misunderstanding about the relationship between efficiency and human labor in schools? And are there examples where technology genuinely has simplified teaching work rather than just redistributing it?
I was at a school recently, where an academic dean was telling me that one of his current challenges was getting all of his teachers to input their assignments into the Canvas calendar. Because the school has a rotating daily schedule, different class periods have assignments and tests due on different days, so he wanted teachers to get all those details online. But it was a ton of work for teachers, so many of them weren’t doing it or not doing it consistently.
Consider all of the generations of technologies that have been introduced to help students manage work in learning environments: printing presses that could mass produce journals and calendars, typewriters for more legible prose, mimeographs and eventually photocopiers to reproduce documents, personal computers for the storage of electronic documents, internet networks for the sharing of electronic documents, Web 2.0 for the easy editing of internet resources, the learning management system to… well… manage all of these things.
After a century of technology innovations, in the year 2026, I’m still sitting with academic leaders telling me that it takes too much teacher time to help students figure out what they need to do for class.
One techno-utopian instinct is to say “Oh, LLM agents will solve that. They’ll fill out the calendar for teachers.” That’s probably true, but it also misses the point: because the mass-printed schedules, the typewriters, the mimeographs, the computers, the shared documents, the web, the LMS were also supposed to solve this problem and they didn’t. In 10 years, I’ll be sitting with another academic dean complaining to me that teachers aren’t spending enough time customizing their agents to do the ever more complicated scheduling tasks that those agents enable.
Technology-enabled efficiencies will not let teachers focus on the most important things. If that were true, then today’s teachers would be awash with time to spend on the most important things standing on a century of innovations in information technology. Subtraction is the solution, not efficiency. If you want people to focus on the most important things, you have to point at parts of the educational system and say, “We’re going to stop doing this thing and instead focus our energies on this core part of our work.”
‘Subtraction is the solution, not efficiency.’
(Incidentally, one of the best research studies on this general theme is Ruth Cowan’s 1985 classic More Work for Mother, which explains why decades of technological progress in home maintenance technologies did not make maintaining a home and family any less time consuming.)
You often come back to this idea of “invest in human systems” — that technologies are only as powerful as the communities and institutions guiding their use. In the current AI moment, where do you think the balance is most out of sync? Are we over-investing in tools while underinvesting in teacher capacity, curriculum design, trust, and community infrastructure?
Well, I’ve never said invest in human systems. Sal Khan said “our biggest lever is investing in the human systems,” which I thought was pretty funny, because “the human systems” sounds like something a non-human would say.
Technology is only as powerful as the communities that guide their use. For my entire career, I’ve watched schools purchase technologies at rates that far outstripped their capacity to use those purchases. In the 2000s, when I ran a technology consultancy, I joked that schools would spend a million dollars to buy new machines and software, and then ask us to do $10,000 in professional development.
Technology only helps if teachers change their practice. Most teachers only change their practice in response to shifts they seem among their peers. Every instructional change initiative is really a peer learning problem. I wrote a book about it called Iterate. You can buy a copy, or just steal it from online pirates like the AI labs do.
One thing that stands out in your writing is your resistance to simple narratives — whether about AI in schools or phone bans. You seem less interested in declaring technologies “good” or “bad” than in asking what conditions shape their effects. Has the education conversation become too polarized around technology? And how do we create space for more nuanced, evidence-driven conversations when public debate increasingly rewards certainty and hype?
I wouldn’t say we need more nuance. Oftentimes, “the simplicity on the far side of complexity” is better than nuance.
Consider phone bans. I think most folks in favor of phones in schools concede that phones can be distracting, but they argue that schools need to be the spaces where students learn to manage that distraction: that there should be a nuanced place for phones to exist within classroom learning alongside just-in-time restrictions.
And many schools tried that balance in some form or another for a decade, and now many communities have decided that nuanced approach doesn’t work—it’s bad for learning, bad for teacher working conditions, bad for student socialization, and so on—and they want to try a simpler approach: banning phones.
I think simple approaches in schools are great. Schools have a very limited annual capacity for improvement. They can really only get better at a couple of things every few years. If schools say “no nuance, no phones” then they are conserving their staff time, attention, and other resources to work on problems that they think are more important. I applaud that. (I’d probably also be excited about schools that really decided that having students learn alongside their phones was a crucial life skill that they were going to invest the time and energy into making work.)
‘A better question is, ‘how can we gather the evidence to figure out what to do?’’
So, while I could do without nuance in a lot of places, I’m nearly always a big fan of evidence. The central challenge of navigating schools since the arrival of generative AI is the widespread lack of evidence for AI policy, AI teaching practices, AI tools, and related issues.
In those circumstances, “what should we do?” is a frustrating question because no one knows what to do, and the people with the most confident answers are as clueless as anyone else. A better question is, “how can we gather the evidence to figure out what to do?” For me, that’s the much more promising pathway forward.
—
Victor Rivero is the Editor-in-Chief of EdTech Digest and Executive Producer of EdTech Digest’s Future Focus Forums (F3), bringing together edtech companies with top U.S. school district technology leaders for the most effective way to get feedback on new and existing products. Write to: victor@edtechdigest.com
0 Comments