AI in the Classroom: What Stanford's Dean of Education Actually Says

Everyone reacts to AI in the same four stages

Dan Schwartz has shown AI tools to a lot of people. He's the dean of Stanford's Graduate School of Education and a cognitive psychologist who studies how people learn, so he's watched the reaction up close, over and over. It goes the same way almost every time.

First, disbelief. Then: "what's left for humans?" Then: "wait, will there even be jobs?" And finally: "oh no, education needs to change."

That last stage is where most schools, teachers, and parents are stuck right now. Not because the technology is confusing, but because nobody's sure which parts of the old rules still apply. Schwartz spent an episode of Stanford's The Future of Everything podcast working through exactly that, and it's a rare thing: an actual expert talking through specifics instead of vague hopes or fears. Here's the useful part, without the buzzwords.

We've had this argument before, it was about calculators

Every time a new tool threatens to do the "hard part" for students, schools face the same fork: ban it, or redesign what you're actually testing for.

Calculators went through this exact fight. The eventual answer wasn't to ban them, it was to stop testing arithmetic and start testing something a calculator can't do:

• Could the student take a different approach to solve a problem?
• Could they reason about why an answer worked, not just produce it?
• Were they being innovative, not just accurate?

Schwartz thinks AI forces the same reckoning, just faster and everywhere at once. The tool changed. The underlying question, what are we actually trying to teach, didn't.

What happened when he actually let students use it

Schwartz doesn't just theorize about this. He teaches a writing-heavy ethics class at Stanford, and he and his co-instructor allowed ChatGPT outright, on one condition: students had to disclose their prompts and show the initial output before submitting a final version.

The results, after a careful pass by engineer TAs comparing declared use against actual use:

• Grades dropped slightly. About two points out of a hundred, roughly a 90 becoming an 88.
• Time dropped a lot. Students who used it finished in about half the time.
• Some students still tried to game it, claiming they'd only used it for copy editing when the TAs could tell it was more than that.

His conclusion wasn't "ban it" or "it's fine." It was blunter: motivated students who want to learn will use AI as a tool and keep learning. Students who were never engaged in the first place were probably not going to be reachable through the rules anyway.

The mismatch he's actually worried about

This is the part of the interview that stands out. Schwartz isn't losing sleep over cheating. He's worried about a mismatch between what people want from AI and what the edtech industry is building.

"Everybody who's got this tool wants to create with it... Meanwhile, the [edtech] field is trying to push towards efficiency. Can we get the kids done faster? Can we get 'em through the curriculum faster? Can we correct them faster? In which case the kids are going to optimize for being really efficient, as opposed to just trying to be creative, innovative, and use it for deeper kinds of things. This is my big fear."

In other words: individual people use AI to make things. The companies building AI for schools are mostly optimizing for speed and correction. Students, in turn, learn to optimize for whatever the tool rewards. If the tool rewards speed, that's what they'll get good at, not judgment, not creativity.

How he changed what actually gets graded

If AI can produce a passable draft or working code in seconds, grading the draft or the code stops measuring much. So Schwartz has started grading the parts AI is worse at:

• In writing, he grades for novelty and appropriateness, not just correctness. AI tools have seen enormous numbers of examples of the safe, common answer, and it shows. They're noticeably weaker at genuinely original, risk-taking arguments.
• In a programming class, points for code that simply runs went down, and points for explaining why a given data structure or algorithm choice was the right one went up.
• In his PhD statistics course, he follows up with quick conceptual questions with no single right answer, the kind of thing a generic explanation doesn't actually answer.

The pattern: stop grading the artifact AI can produce. Start grading the reasoning behind it.

Where AI is genuinely helping, not just saving time

It's not all trade-offs and workarounds. A few uses stood out to him as real wins:

• The "protégé effect." Students who teach an AI agent how to reason about a topic learn it better than students studying for a test, because explaining forces you to actually understand the material first.
• Project-based learning at scale. A teacher with 130 students can't individually plan 130 separate projects. AI can help design each one, and a dashboard can track whether every student is hitting their milestones, something that was simply impossible to manage by hand before.
• Real-time assessment. Instead of grading a finished essay after the fact, AI can watch how a student works through a problem and adjust the next piece of instruction immediately, which matters more for learning than the final grade does.

What AI still can't touch

Schwartz's other research area is embodied learning, the idea that physically manipulating something (turning your hands to model gears, walking through a space) builds understanding that words alone don't. His example: a colleague who teaches architecture has students design a building, feeds the blueprint into a physics engine, then has the student navigate their own design in a wheelchair. Students immediately understand doorway width and turning space in a way no explanation gets across.

Today's AI, he points out, is still overwhelmingly verbal and abstract, even with video and image generation. The physical, hands-on side of learning is a gap current tools don't fill, and it's worth knowing that going in rather than expecting a chatbot to replace it.

The takeaway

The debate over AI in schools keeps getting framed as allow-it-or-ban-it. Schwartz's actual experience says the real work is elsewhere: figure out what you want a student to walk away understanding, then stop grading the parts a tool can now do for them.

That's also the most useful lens for picking an AI tool as a student, parent, or teacher: is this helping someone build a skill, or just helping them skip the step where the skill would have formed? If you're looking for tools built specifically for studying, writing, and classroom use, browse AI tools for students and teachers rather than reaching for a general chatbot by default.