Most schools that have tried AI in the last few years have something in common. They started with the tool. A vendor showed up, the pitch was compelling, the board approved a pilot, and the school rolled it out. And then, somewhere between the rollout and the results, nobody could quite explain what problem they were solving.
David Adams has watched this play out across public education for years. He leads work at The Urban Assembly, a nonprofit in New York City that builds and tests approaches to making public schools work better for kids who need them most. He’s not against AI. He’s building it. But the two tools his team has built, CounselorGPT and Project Cafe, both started the same way: with a problem they already knew how to recognize.
That’s the thing this episode kept coming back to. Not what AI can do. But whether you know your problem well enough to know if AI is actually solving it.
What does it mean to solve a bounded problem with AI?
David talks about AI in education in terms of constraints. Time, knowledge, resources, incentive structures. The question isn’t whether AI is good or bad. It’s whether AI can help reduce a specific constraint that’s keeping students or teachers from getting what they need.
Project Cafe came from a known constraint in teacher development: feedback is the highest-leverage thing you can give a teacher, and it almost never happens frequently enough because it costs too much. A principal can’t be in every classroom. Coaches are stretched. So feedback is sparse, and teachers don’t improve as fast as they could.
Project Cafe lets teachers watch video of their own classroom practice, review the questions they asked, the redirections they made, the ratio of teacher talk to student talk. They can share clips with colleagues. They can track their own patterns over time. The result, according to David, is about twenty percent more respectful language, forty-five percent higher positive climate scores, and meaningful gains in how teachers structure and summarize lessons.
That’s not a coincidence. It’s what happens when you apply AI to a constraint you understand well enough to measure.
How does CounselorGPT help students make better decisions?
Post-secondary counseling in most public schools comes down to a math problem that nobody talks about honestly. There aren’t enough counselors to give every student real guidance. So students make expensive decisions about credentials, colleges, and careers with very little information, and sometimes those decisions cost them years and hundreds of thousands of dollars.
David’s framing for this is sharp: we’re not against college. But seventy percent of people going to college say they’re going to support economic mobility. They deserve to know whether their credential matches that goal.
CounselorGPT doesn’t tell students what to do. It gives them data. Growth paths. Labor market statistics. The cost of credentials and the likely return. What a surgical tech earns. What an air traffic administrator does and where they’re hiring. The idea is to move students from guessing to guidance, so that by the time they sit down with a counselor or a parent, they’re asking better questions rather than starting from zero.
The constraint being reduced isn’t the counselor. It’s the information gap between the counselor and the student. That’s a constraint AI is well suited to close.
Where does AI not belong in education?
David doesn’t frame this as red lines. He frames it as problem clarity.
He gives two examples worth sitting with. The first is bedtime stories. If you think reading a bedtime story is about transmitting words, you could automate it. But bedtime stories are about bonding. Joint attention. Relationship development between a parent and a child. AI can’t do that. Not because it’s incapable of reading words, but because it’s replacing the wrong thing.
The second is jury deliberation. AI could probably process evidence and return a verdict. But a jury of peers carries legitimacy because the jurors have responsibility. They can be held accountable. A machine can’t. The verdict would mean something different, not because the analysis was worse, but because the source of authority was different.
His point across both examples is the same: before you reach for AI, get clear on what the interaction is actually for. If it’s for relationship, for accountability, for the kind of attention that only comes from a person who has skin in the game, then AI isn’t going to solve it. It might make it look like you solved it. That’s worse.
What about AI detection tools in classrooms?
This came up directly in the conversation, and David was as direct as he gets anywhere in the episode.
AI detection tools in classrooms are not effective. They flag student work as AI-generated when students wrote it themselves. And when a teacher tells a student “you didn’t write this,” it’s not just a grade problem. It’s a dignity problem. Students feel accused of something they didn’t do. They feel like their intelligence is being questioned. David hears about it directly from students who come to him with that experience.
There are better approaches. Blue books. Project-based work that draws on specific experiences. Having students submit a draft and use AI to revise it, then discuss both versions. Building a feedback loop where students get more frequent responses on their writing before a final draft is due, so the feedback actually reaches them while they still remember what they wrote.
The common thread is knowing the student well enough to recognize their work and their thinking. That knowledge can’t be outsourced to a detection algorithm. It’s part of what it means to be a teacher.
Use This Today
Before your next AI decision, in your school, your organization, or your team, write down the specific constraint you’re trying to reduce. Not the goal (“improve student outcomes”). The constraint (“teachers have time to observe each other twice a year at most, and feedback takes six weeks to get back to them”).
If you can write down the constraint in one specific sentence, you’ll know whether AI is actually suited to help. If you can’t, that’s the work to do first.
David’s closing thought from the episode: “If you solve problems, you’ll be okay. If you’re just using tools, I can’t vouch for that.”
Connect with David Adams and The Urban Assembly: urbanassembly.org
Free AI Policy template: goodcirclemarketing.com/summer-of-good-ai
Good AI Score self-assessment: goodcirclemarketing.com/summer-of-good-ai