Friends and colleagues,

I have heard a number of leaders start discussions about AI with the question, “Are you for or against AI?”

I suspect there are many points in history where similar questions have been posed: In 1889, “Are you for or against electricity?” In 1953, “Are you for or against atomic energy?” In 1995, (this one I remember), “Are you for or against the World Wide Web?” I wonder how long it took for the question to change from “Are you for or against?” to “What are the risks and benefits we need to manage at the same time?”

I am in a small club of people who have seen a number of AI-powered products in classrooms and talked to hundreds of users. What I have learned prevents me from choosing one camp or the other—AI is both a huge opportunity for learning and a huge risk to learning at the same time. AI in schools is not one thing; it is a source of power to do many things—some great, some horrible. My speculations about what will be good or bad are often wrong; I have learned that we have to observe real use to see what adds value and what undermines learning.

The biggest risk to learning I saw on the tour came from a tool that few people would call “AI”: Google Search. I didn’t see students turning to ChatGPT or Gemini in school often (though I heard a lot about students using these tools at home). But, I did regularly see students open their search browser, ask a question, and copy the “AI Overview” summary without questioning it or—as far as I could tell—usually even reading it. I have seen screenshots of computers in New York and LA (where most AI is banned in K–12 schools) that show the AI summary in Google Search still alive and active. The bans do not seem to have ruled out the biggest risk to learning I saw on the tour.

But the bans do seem to have ruled out some remarkably helpful uses of AI that I saw. Based on test scores and reports from educators, we are not preparing kids to be great writers. To become a great writer, a student needs a lot of practice and a lot of feedback, which is constrained by the time and effort it takes teachers to grade. I saw targeted instructional tools use AI to give students massively more practice and feedback on their writing than the most industrious human teacher could deliver in a week. It didn’t eliminate the need for teachers to also read and provide feedback on student work, but it did give teachers a huge assist in a laborious instructional job.

Foundational reading skills are essential for student opportunity. I saw targeted tools use speech recognition and AI analysis to pinpoint the letter sounds that students were struggling with and give them immediate instruction on those learning needs. These tools did not eliminate the need for teachers to hear students read, but they did give teachers a much better way than static worksheets to increase foundational skills practice during a station rotation than a teacher could offer with a worksheet.

As I see policy efforts to try to regulate AI in schools emerge, I worry we are headed toward a quagmire. It is hard to come to a consensus definition about what AI even is. Even if you could get large groups of people to agree on definitions, the technology is changing so rapidly that they’d need near constant revision. Different types of AI are increasingly being integrated into everything (e.g., the new iPads have AI-powered features in the keyboard). It is becoming (or maybe has already become) virtually impossible to isolate and completely expel AI from all the different types of technology that most kids interact with.

Rather than focusing on whether or not AI is present as the approach to regulation, we’ll do better for students if we look at the types of interactions we want technology involved in and the types we don’t. As a parent, I want the tech my kids use to be safe and their data to be secure, but I’m less interested in the particular technology in the tool (e.g., is this predictive AI or generative AI?) than the job the tool is trying to do—and how effectively it does it. I want the very best technology available to help my kids get practice and get feedback on writing, but I don’t want any technology giving them spiritual counsel. I want my civic and educational leaders to focus less on whether AI is good or bad or whether or not AI should be allowed. I want them to wade into the specific kinds of interactions and uses we want to encourage and discourage.

Like our forebearers had to do with electricity, finding the right guardrails will require all of us to be learners and to watch for both access and safeguards at the same time. By 1930, nearly nine in 10 urban homes had electricity, but only one in 10 farms did because private utilities saw little profit in running lines to rural families. It took the Rural Electrification Administration in 1935 to begin closing a gap the market had left open for half a century. Private markets did not manage for equal access to a major enabler of quality of life. Safety lagged in ways that were harmful: New York strung live wires over crowded streets for years before a series of electrocutions in 1889 forced the city to bury them. The rules that made a difference came from watching real use: the National Electrical Code, first published in 1897—many years after cities began introducing electricity—grew out of close study of the fires and failures that actually happened in homes and factories.

Policymakers should act quickly on what we know about safety and data privacy but, when it comes to instruction, we have a lot more to learn before we know how to regulate. What policymakers can do is set up a framework that will support that learning. They can make sure that school leaders and product leaders are in classrooms actively watching how product use is going and use those observations to drive improvements. They can make sure that state leaders and district leaders are carefully studying innovations and sharing what they are learning.

I have been in about 1,000 classrooms over the last decade. Those observations have shown me just how hard it is to teach. Kids have a lot of learning needs right now, and we need every tool in our toolkit to deliver great learning to them. What I saw on the tour convinced me that AI-powered products can offer a lot of value alongside the risks and that we have a lot more to learn to know how to navigate it.

One step at a time, together,

Emily