New York City's AI guidance has ten steps for procuring AI tools yet punts on the most important questions about AI and learning.

New York City public schools recently released preliminary AI guidelines for its 1,600 schools, 150,000 employees, and one million students. I’m personally invested because my child is a student at one such NYC public school, and apparently is required to use an AI tutor, which I wrote about at length. But how NYC rolls out AI to kids isn’t just a matter of my personal concern. Because New York City has the largest school system in the country, how it handles AI and education could influence schools across the country. I don’t envy the hard job administrators have here, but I’d like to point out what I think is a major flaw in the guidelines.

Here’s what I want you to imagine while I describe what’s in the AI guidelines: You are a ninth-grade English teacher. You have spent your career refining a curriculum around close reading and argumentative writing. You are a pro in the Herculean and unenviable task of motivating teenagers to persevere in the face of a difficult text until they understand, often against their will, and then prove it to you in an essay (also against their will). Then, in a matter of a few months in early 2023, you watched that methodology, however imperfect it was, completely fall apart. Half your students use AI to summarize their reading and then generate a passable essay in ten seconds. The ones who don’t use AI, meanwhile, struggle to explain why they should bother when AI could do all their schoolwork for them. You try AI detection software, but students use AI humanizers in an arms race that frays trust. AI is starting to make you and your students question the entire value and purpose of education. Yet, you need to show up every day, and somehow figure out, on the fly, how to rethink your assignments, your assessments, your whole theory of what your classroom is supposed to be. You need help.

Here’s what the largest school system in the country produced in response: a procurement document. Ten steps for vetting vendor products and a traffic-light risk framework (green means educators can use AI for scheduling and drafting communications; red means no behavioral monitoring or student surveillance). We know who should submit approval requests, what data agreements vendors must sign, how cloud architecture will get reviewed, and how approval can be revoked.

The document gets the problem wrong. The challenge AI poses to learning isn’t primarily about which AI tools belong inside the classroom. It’s about the AI tools that are already outside it. Students have an undetectable homework-completion engine in their pockets right now, and a vendor compliance process with one hundred steps won’t change that. Many of the mechanisms through which schools have assessed learning for decades (RIP, take-home essay) no longer assess learning—not because the wrong AI entered the classroom, but because AI entered the world.

The guidelines attempt to take a stance on their role in a world saturated with AI. Under a section titled “Beyond NYCPS Authority,” the guidelines state: “This Guidance focuses on what is within our control: how AI is evaluated, governed, and used within NYCPS schools for educational purposes.” How students and families use AI outside of school is listed among the things that “fall outside the scope of what NYCPS can govern.”

I would argue that’s only half right. True, they can’t govern what teenagers do on their phones at home. But the conclusion they draw—that their job is, therefore, only to manage AI inside schools—is flawed reasoning. The existence of AI outside school is precisely what demands they rethink what happens inside it. Thus, the question the guidance should primarily be asking is not how do we vet AI tools for schools but how do we redesign the classroom now that ambient AI in the world has made many of our old methods obsolete?

The NYCPS guidelines pretend the second, harder question is outside their purview. However, in reality, schools have enormous power over what happens inside their walls—I should know because I only found out my son used an AI reading tutor when he mentioned a mysterious AI companion named “Amira” over mac and cheese one night. It simply was not up to me. NYCPS controls the curriculum, the assignments, the technology, the structure of the school day, the nature of the interaction between a teacher and a student. The problem of ambient AI is a problem they are capable of addressing—not by ridding the planet of AI, which they can’t do, but by changing what happens in the classroom.

Take a step back: Although the visible output of a high-school English classroom may be papers and book reports, the true product of education is the invisible changes in the minds—and character—of the students. In this view, it makes more sense to start with the capacities you’re trying to build in future people: the ability to read closely, to argue logically, to hold two contradictory ideas in your mind at the same time, to tolerate and persevere through difficult material, to wrestle and engage with different points of view. And then you ask how your teaching needs to change to cultivate the development of those capacities in a world where a machine can simulate most of the traditional outputs, like take-home essays, that used to serve as proof of them. Perhaps AI tools can help with this overhaul; perhaps not. But that is a secondary question. Arguably, all you need is a teacher, a room, and something difficult to think about or do.

Some may argue that because AI is already a part of the world, then schools have an obligation to teach students to be AI literate. This fuzzy term “literacy” can mean many things, but at its simplest I would agree that students should be exposed to AI in controlled settings to learn about its strengths and weaknesses, its potential harms and benefits. But this level of literacy requires something akin to Driver’s Ed, but for AI, not AI in every moment of school. The stronger case for AI in the classroom is that AI skills will be needed for a future AI-driven economy.1 Although this reasoning is superficially plausible, I think it’s misguided. The point of school isn’t job training for an uncertain economy, it’s building foundational skills that will allow students to thrive, whatever the world throws at them. You don’t need to stop learning to think clearly because there is AI when you graduate any more than you should stop going to the gym just because you can drive a car outside of it.

The educator Andrew Cantarutti has developed this reasoning at length, calling for schools to be “walled gardens,” or spaces where kids can develop sustained attention and deep thinking precisely because they will most benefit from those capacities when facing the much noisier, chaotic world outside. I like that metaphor, especially now. The fact that AI is omnipresent, that students are swimming in the slopstream the moment they leave school, seems to me to make the case for shoring up the walls of the garden, not tearing them down. And this is squarely within NYCPS’s power.

So what would a different NYCPS response look like? I was recently on a panel with Eric Hudson, a former classroom teacher who now advises schools on learning design, and he put it simply: if you treat AI as a technical development, you invite technical solutions. If you treat it as an assault on current lesson designs, you invest in people. I think about the hypothetical high-school teacher struggling to adapt to a reality they weren’t trained for. They don’t need a painstakingly vetted AI tool. They need professional development.

So, I have a modest proposal for NYCPS administrators: for every dollar spent on AI procurement, spend a dollar helping teachers redesign their teaching for a world that has changed underneath their feet, through no fault of their own.