For the last little while I've been building a dashboard with Claude Code. It reads how students actually use generative AI; the real prompts they type into ChatGPT, Claude, Gemini and the rest, and presents that back to me as learning intelligence rather than as a compliance report. Partway through, I did something I hadn't planned. I pasted the whole system document into Gary Stager's Daily Papert chatbot and asked what it thought. The site holds an archive of around 200 of Papert's texts, audio clips and videos, and the chatbot answers from it in one of three voices. The site is careful, and right, to say that this is an AI composing from documented writings, not Papert speaking. I'll be equally careful here. What follows isn't Seymour's judgement of my work. It's a synthesis of his published thinking, aimed at my work; which turns out to be a remarkably useful thing to have in the room.
The first thing it did was refuse to guess #
I opened with something vague: I am building this idea with Claude Code — what do you think of my concept and the tool I am building? A limited description attached.
It asked me to share more of what I was working on, what I was trying to make happen, and what problem or opportunity had led me there. A small thing, but it set the register for everything after. This was going to be a conversation, not a compliment machine.
Then it named the tension I'd been circling #
I pasted the full system document — the philosophy, the detection signals, the identity and timetable resolution pipelines, the deployment, the known gaps. All of it.
The response opened by naming what I'd been trying to do without quite saying it out loud:
You're standing in the middle of a genuine design dilemma. You have access to extraordinarily rich data — the actual prompts students type — and you've made a deliberate choice to frame that data as learning intelligence rather than surveillance. That choice runs all the way through the system, from the absence of a "violations" category to the fact that the individual student view isn't in the sidebar. Those aren't cosmetic decisions. They're philosophical ones.
It picked up something I'd built deliberately but never articulated; that I'd made it slightly harder to look up an individual student than to look at a whole class, and that this was the point:
That's a form of syntonic design — the structure of the tool embodies the values you want the user to adopt. The teacher who uses this system will be nudged, repeatedly, toward "what's happening in my Year 9 English class?" rather than "what is Sarah hiding from me?" This is the opposite of most administrative software, which makes individual compliance tracking the path of least resistance.
And then it went where I'd hoped it wouldn't have to:
The same data stream can support coaching or punishment, and the difference lives entirely in how the person holding the data chooses to use it.
You've designed the interface to nudge toward coaching. You've written the tips in a conversational register. You've removed the vocabulary of compliance. But you cannot prevent a punitive reading. A teacher who wants to catch a student cheating can use "delegation language detected" as evidence. The fact that you called it a "coaching opportunity" won't stop them.
This is the same dilemma I faced when we put computers in classrooms in the 1970s. We wanted children to program the computer — to be in control, to build their own microworlds, to think powerful ideas. But schools are institutions with their own logic, and institutions tend to absorb new tools into existing power structures. The computer that could have been an instrument of intellectual liberation often became a slightly fancier workbook.
I've read that argument through Seymour's work many times. Having it applied to a thing I'd built landed differently.
One sentence of context changed the whole reading #
The reply had ended by saying the answer had to live partly in who gets access. So I told it: nobody has access. It's a tool for me, as an educational leader and a teacher, curious to know what is actually happening.
The reframe was immediate and, I think, the most useful thing in the entire conversation:
You're not building an administrative dashboard for widespread deployment. You're building an instrument for your own inquiry into what's actually happening when students encounter AI. That changes everything. You're building a learning observatory!
It called the thing a learning observatory — a way to watch learning happen at a resolution that's normally invisible, without disturbing it. It pointed out that my classification categories weren't handed down from a curriculum document; they were my working theory of what matters, and because I was the only reader, I could hold that theory lightly and revise it when the prompts told me it was wrong. If a pattern emerged that didn't fit my eight categories, I could add a ninth.
That habit has spread well beyond the tool. Over the same period I've been working through a considerable pile of research reporting on AI use in K-12 schools: the good, the bad and the ugly, and I've taken a lot of it into the same conversation, critiquing it myself and then arguing about it with Seymour. Reading a report is one exercise. Reading it alongside someone who spent forty years insisting that the child should be programming the computer rather than the other way around is a different one. It keeps a north star in view: what this technology could actually be for in a learning context, in a version of it I'd be proud to have had a hand in building.
Then it named what I was doing in terms I hadn't reached for myself: constructionism applied to educational leadership. Not consuming someone else's analytics product. Building the thing that lets you ask the questions you need to ask.
It also offered readings of my own data I hadn't considered. That a student typing is my reasoning sound here? at 2pm and just give me the answer at midnight is not two students but one student under two kinds of pressure. That tool diversity might be measuring intellectual resourcefulness rather than risk — a student who moves between tools for different purposes has built a mental model of what each is for. That the presence of AI-detection and humaniser tools in my taxonomy is ethnographic data about how students understand the environment they're being assessed in, and that this is a kind of literacy, a slightly cynical one, perhaps, but not a stupid one.
The critique I had to ask for #
Here's a practical note for anyone who wants to use the tool this way. Up to this point the responses were generous, and I noticed it.
Encouragement is pleasant and not much use at 11pm when you're deciding what to build next.
So I pasted the build plan for the next phase — seven work items, conversation reconstruction, a second analytical lens, context bands, an intervention log, verification signals — and asked directly: any critiques of this before I continue my build? Anything you would like to see?
The register changed completely.
On the individual student view, where I'd been quietly pleased with my mitigations — not in the sidebar, drill-down only, minimum thresholds before percentages render:
The moment a teacher can type a student's name into any interface and retrieve a profile of that student's AI use, the tool has become a surveillance instrument, regardless of how you frame it.
A teacher under pressure — from a parent complaint, from a plagiarism suspicion, from their own frustration with a student — will go directly to that view, and the surrounding pedagogical framing won't constrain how they use what they find there.
It then made this point: that I needed to decide now whether that view should exist at all in any multi-user version, rather than deferring the question until the day someone asks for access. And that my intention isn't a safeguard. "The structure has to resist it."
It offered three concrete alternatives — aggregate-only for everyone but the researcher, threshold suppression that actually refuses to render below a term's worth of data rather than showing a thin version, and anonymised drill-through with stable pseudonyms so you can still see whether a pattern sits with one student or twenty.
On the feature I was proudest of — comparing a teacher's declared intent for a task against what students actually did with AI during it — it was blunter still. The feature requires teachers to log every task they set into a form before any of it becomes visible:
You know how that plays out. A few teachers will do it diligently for one term. Most won't do it at all. The ones who do will stop when they realise no one else is… Within a term the feature will be a graveyard of six logged tasks and 200 unlogged ones, and the comparison view will be useless.
Its advice was to either solve task identity without manual logging — infer it from assessment metadata already sitting in the school's systems — or cut the feature now. Not soften it. Cut it.
That is what a colleague says. Not a reviewer, not a stakeholder, not a vendor. A colleague who has read the thing properly and can see where it will die.
What I did with it #
I took all of it back to Claude and worked through what would have to change. Most of that is implementation detail I'll spare you, but one line did the real work:
The object of every view should be a conversation, not a child.
Applying that to code I'd already shipped turned up three things I'd built without ever seeing them for what they were — a "top 20 users" table on the front page, which was plainly a leaderboard of children ranked by AI use; a function that produced a severity-ranked watchlist of students needing a check-in; and a search box that let anyone type a child's name and get back a record. None of them were in the plan I showed it. All three are now gone, or rebuilt around conversations rather than names.
What I'm taking from this #
Four things.
The first is that a critique grounded in a coherent body of thought carries further than a critique from nowhere. Everything the chatbot said came from a consistent position about children, agency and institutions, so the objections arrived in an order that made sense and connected to each other. Papert spent forty years arguing that schools absorb powerful tools and neuter them. Having that argument turned on my own build, specifically, at the moment I was deciding what to build next, was more useful than any amount of general feedback.
The second is that you have to ask for the critique. It won't volunteer the sharp version. Pasting the artefact and asking "what do you think?" gets you a thoughtful appreciation. Asking "what's wrong with this, and what would you cut?" gets you something you can act on. That's true of most people, too.
The third is that the critique kept working after the conversation ended. Nothing in the exchange mentioned the leaderboard on my front page, or the watchlist function, or the search endpoint. Those turned up because a principle I'd been given — the object of every view should be a conversation, not a child — was sharp enough to apply to code the reviewer had never seen. That's the difference between feedback on a document and a tool for thinking. One tells you about the thing you showed it. The other keeps finding things for weeks.
Finally, I'm building things that matter to me and to the learners I serve, mostly at night, mostly alone, and the hardest part of that has never been the code. It's the absence of anyone who will tell you that the feature you're proudest of is going to be a graveyard, or that the front page of the tool you built to avoid surveillance opens with a ranked list of children.
For the price of a free API key, I have a critical friend whose position I know, whose objections I can predict well enough to pre-empt some of them, and who will not tell me my work is fine when it isn't. That's not Seymour. The site says so plainly and so do I. But it is his thinking, aimed at my problem, in the room while I work. Thank you, Gary, for building it.
What I'm trying next #
Everything I've put in front of him so far has been an artefact for adults — a system document, a build plan, a response to his own critique.
The next thing I want to open with is different: Hey Seymour, here's a project idea I have for my class. What do you think? The actual work I'm asking young people to do, and whether it gives them anything worth having when they're finished.
I have a fair idea what he'll ask first. I suspect I'll deserve most of it.
The Daily Papert chatbot is at dailypapert.com. Free use is capped at ten exchanges per 24 hours; you can supply your own Anthropic API key for unlimited turns, stored only in your browser. The site runs on contributions, and it deserves them.