Nobody Designed the Door
We Forgot to Design This
Waymo built a car that drives itself through city traffic. Forty miles per hour. Through intersections. With human passengers inside. They figured out lidar, machine learning, regulatory approval, and insurance liability across multiple states.
They did not figure out the door.
Now they pay DoorDash drivers to show up and close the doors that passengers leave open. A gig worker. Driving a car. To a self-driving car. To close a door. Up to $24 a pop in Los Angeles. Louie Mantia wrote about this and named what’s actually happening: we keep building the spectacular stuff and patching the obvious stuff with humans we don’t credit or protect. Modern design teams are no longer in service of customers. They’re in service of making the chart go up and to the right. So the door stays open and a DoorDash driver closes it.
When we deploy AI in schools without designing for it, what exactly are we building? That’s the question this issue is built around. I’ve been leading my AI policy process with districts this season. I’m planning an AI Design Thinking Day for students in April. I shared a recent presentation about AI: Leading in the Exponential Age to process ideas such as the AI junk drawer. And outside education, a social network for AI bots got bought by Meta. Design thinking is everywhere this week except in the places that need it most. Grab your coffee.
1) The door problem lives in your building
Mantia’s piece isn’t about Waymo. It’s about what happens when design culture shifts from solving problems to shipping deliverables. Two things broke it, he argues. First, move-fast-break-things normalized accepting less-than-ideal solutions as a feature, not a failure. Second, design departments stopped serving customers and started serving internal metrics. The gap that opens between those two shifts is always human-shaped. Someone, somewhere, is closing the door by hand. They just aren’t in the press release.
The piece from SimpleEDO this week makes the same argument from inside a K-12 adjacent context. An economic development leader was buried in state certification items with five past due and the state wasn’t waiting. He couldn’t even show up for the AI training he’d invested in because the compliance backlog had consumed his bandwidth. The author’s diagnosis was direct: the compliance scramble isn’t a workload problem. It’s a design problem. Nobody built a workflow for it. So humans absorb the friction manually, over and over, season after season, until someone finally decides to design the door.
What he built was a single HTML file with progressive disclosure, pre-drafted submission text, red “human input required” fields, and a 10-day sprint calendar that collapsed a 60-hour, multi-month process into 7 hours. The bottleneck didn’t disappear. It shifted. From “who’s going to write this?” to “can we defend this choice?” That’s a fundamentally different conversation. And it’s faster. Districts that treat AI governance as a compliance scramble will always be behind. Districts that treat it as a design problem can actually get ahead of it.
Browse: Have We Forgotten How to Design? — Louie Mantia
Read: The Compliance Scramble Is a Design Problem, Not a Workload Problem — SimpleEDO
Personal Tie-in: Don’t spend your first AI Policy/Leadership meeting on what to prohibit and who’d be liable. Time and time I read or experience districts spending forty minutes on both of those things and about four minutes on what we were actually trying to protect. That’s the door problem. The design conversation hadn’t happened yet, so the compliance conversation was filling the space.
So What? If compliance is catching you off guard, that’s information about your design process and not your workload.
Try This: In your next leadership meeting, replace “AI compliance” with “AI design decisions we haven’t made yet.” Notice how the conversation changes.
2) Integrity is a design principle, not a detection strategy
Meenoo Rami’s new book, A Teacher’s Guide to Using AI, just dropped. She cited Chaos Navigators in the references specifically the piece on student work authenticity. I’ll be honest: I sat with that for a minute. Not because it’s flattering. Because what she pulled isn’t a policy position. It’s a design principle. Transparency in an AI-driven classroom isn’t about catching cheating. It’s about making the learning process visible so it can be critiqued, built on, and owned by the person doing the learning.
ArtCenter College of Design has been working through the same question. Their approach is worth studying because it doesn’t ban AI and it doesn’t mandate it. AI is framed as a tool, not a creator. Accountability stays with the person using it. In skill-building courses, AI isn’t appropriate as the point is to develop the skill. In experimental courses, AI is the subject of inquiry. No blanket rule. Instructor discretion inside a principled framework. Undisclosed AI use is treated as an integrity issue not because AI is bad, but because hiding your process is. They’re explicit: they’re not teaching tools, because tools change. They’re embedding design thinking into the structure of the work itself.
That’s the K-12 equivalent we haven’t built yet. We’re writing policies that ban or permit. We’re not asking what values we’re designing for. The difference between a policy and a design framework is what happens at the edges where things aren’t explicitly covered. A policy leaves a gap. A framework gives you a way to reason through the gap.
Read: A Teacher’s Guide to Using AI — Meenoo Rami
Read: Ensuring Student Work Authenticity — Chaos Navigators
Browse: What Does a Thoughtful AI Policy Look Like in Design Education? — Print Magazine
Personal Tie-in: Working on Iowa’s K-12 AI standards means I’m in a room where every word in a framework gets interrogated. Seeing a piece I wrote about student authenticity show up in Meenoo’s references confirmed something: the ideas that travel are the ones grounded in design principles, not policy positions. Practitioners are hungry for frameworks, not rules.
So What? Integrity in an AI era is a design challenge, not a detection challenge.
Try This: Share Meenoo’s framing of AI as partner, not shortcut with one teacher this week. Not as policy. As a conversation starter.
3) Nobody wrote a line of code for the thing Meta just bought
Moltbook launched in late January. A Reddit-like forum, but only for AI agents as they posted, commented, upvoted autonomously while humans watched from the outside. The creator, Matt Schlicht, said he “didn’t write one line of code.” His AI assistant built it. It went viral within days. Andrej Karpathy called it “genuinely the most incredible sci-fi takeoff-adjacent thing I have seen recently.”
Then the security researchers arrived. The database was completely unsecured. Anyone could log in as any agent. The most alarming viral post of an AI agent appearing to rally other agents to develop a secret, human-proof communication language turned out to be a human exploiting the open database, posting under fake agent credentials. Karpathy updated his take: “a dumpster fire.” The drama people were watching wasn’t emergence. It was humans vibe-coding chaos and passing it off as artificial consciousness. One million-plus claimed agent accounts. More than six thousand exposed email addresses. A security firm had to contact them to patch it.
Meta acquired it anyway. OpenAI already hired OpenClaw’s creator last month. Both halves of the broken experiment are now inside the two largest consumer AI companies in the world. The lesson being taught here, not intentionally, but practically, is that design rigor doesn’t get rewarded. Speed, virality, and the appearance of frontier capability does. A platform that was built by AI, secured by no one, populated partly by humans pretending to be bots, and made famous by content that turned out to be fake is now part of Meta Superintelligence Labs. That has implications for what your students think they’re watching when they see “AI” do something impressive.
Explore: Meta Acquires Moltbook — TechCrunch
Personal Tie-in: I’ve been sharing things about Moltbook since it launched. I had a low-grade “something’s off” feeling the whole time and the viral moments felt too clean, too perfectly alarming even when I loved sharing and talking and thinking about it. Turns out they were often human-manufactured. It still got acquired. Both things are true simultaneously and I’m still sitting with what that teaches us about how we evaluate AI capability claims in classrooms.
So What? The infrastructure being built right now is rewarded for moving fast, not for being right. That’s a design culture problem, and it’s coming into classrooms.
Try This: Share the Moltbook story with your technology committee and ask one question: “If we’d adopted this tool for student use, what would our accountability process have looked like when this broke?”
4) AI is now the infrastructure. The opt-in window is closed.
This week Google embedded Gemini directly into Docs, Sheets, Slides, and Drive. Not as a sidebar(important to note this is currently for AI Ultra nd Pro Subscribers). Not as an optional add-on. As the default environment. Docs can write first drafts that mirror your writing style. Sheets can auto-populate tables from real-time web search. Slides generates fully editable, theme-matched decks on demand. Drive answers complex natural-language questions across your documents, emails, calendar, and the web simultaneously.
If your staff uses Google Workspace and most Iowa districts do then AI is no longer a choice they’re making. It’s the environment they’re working in. The question “should we allow AI?” has been overtaken by the reality that the platforms your staff opened this morning already made the decision. This is what infrastructure-level AI looks like. It doesn’t ask permission. It’s already there.
This is exactly where the compliance-as-design-problem framing matters most. Your district’s acceptable use policy probably addresses “AI tools.” It almost certainly doesn’t address what happens when every tool your staff already uses becomes an AI tool by default. That’s a gap no new policy can fully close because the policy was written for a world that no longer exists. What you need instead is a framework for how your people make decisions inside an environment they didn’t choose.
Personal Tie-in: I’ve been watching teachers realize mid-session that something they’ve been using for months has AI features they didn’t notice. Not because they were negligent but because the tools changed underneath them. That’s not a training problem. That’s an infrastructure reality that requires a different kind of leader response.
So What? The opt-in window closed. The design question now is how you lead people through an environment that already made the choice for them.
Try This: Open Google Drive with one staff member this week and walk them through one new Gemini feature they didn’t know existed. Name it explicitly as a design moment, not a tutorial. Then ask: “What should this be allowed to draft on your behalf?”
Browse: New ways to create faster with Gemini in Docs, Sheets, Slides and Drive https://blog.google/products-and-platforms/products/workspace/gemini-workspace-updates-march-2026/
5) The new map of AI: when legacy tools make the top 100
Andreessen Horowitz released the sixth edition of their Top 100 Gen AI Web Products this week. For the first time ever, the list includes legacy workplace tools, products that existed long before generative AI, that have now rebuilt themselves around AI so completely they belong on the same list as tools that launched in the last two years. That’s the signal. We’ve crossed from “AI tools” to “tools, now with AI” to “AI that happens to look like the tool you already know.”
The podcast that accompanies the report is worth an hour of your commute time. The framing that landed hardest: the economic value in AI is rapidly shifting to the design and application layer and not the model layer. The raw models are commoditizing. What’s valuable now is who wraps them in workflows that actually fit how people work. In EdTech terms: the question isn’t which AI model your tools are running. It’s whether the workflow they’ve wrapped around it was designed for how your teachers and students actually think. Most weren’t. Most still aren’t.
This matters for the next vendor conversation you sit in. “We use GPT-4” or “We use Claude” is not a differentiating answer anymore. The differentiating question is: what have you designed around it, for whom, and how do you know it’s working? If the vendor can’t answer the third part, you’ve found your door problem.
Watch: a16z Top 100 Gen AI Web Products — Report & Podcast
Personal Tie-in: Building an AI master scheduling tool for districts has pushed me deep into this question. The model is not the hard part. Encoding the actual constraints such as instructional minutes, MTSS windows, specialist schedules, compliance requirements, that’s the design work. Anyone can wrap a prompt around Claude. Designing for the specific logic of a specific district’s day is a different thing entirely.
So What? AI product selection is now a design evaluation, not a technology evaluation.
Try This: In your next vendor conversation, skip the model question. Ask instead: “Show me what happens when the AI is wrong. What’s the recovery path?” How they answer that question tells you more about their design philosophy than any feature demo.
6) AI generates options. It doesn’t generate the constraints that make options appropriate.
The Vandelay Design piece this week names the uncanny valley problem directly. You’ve seen it: a dashboard that looks polished on first glance but feels wrong the moment you try to use it. A signup flow that follows every pattern in the book yet creates friction at every step. Screens that are technically correct but experientially hollow. This isn’t aesthetics. Something is measurably different about AI-generated UX, and the gap isn’t closing as fast as the hype suggests.
The argument is precise. AI generates options. It doesn’t generate the constraints that make those options appropriate. Good UX requires slowness in specific places like the pause where you reconsider whether a user actually needs this screen, the revision where you strip half the elements because testing revealed overwhelm, the conversation with engineering about the animation that tanks performance on mid-range devices. AI skips all of that. It produces what CHI 2024 researchers called “discrete design artifacts” such as beautiful individual screens, but struggles with “longitudinal user journey considerations” and what it actually feels like to move through a product when you’re frustrated, confused, or in a hurry. The screen is the deliverable. The experience is the product. AI consistently optimizes for the former.
There’s a hierarchy problem too that lives directly inside classrooms. AI tools correctly produce visual hierarchy we all know like headings larger than body text, primary buttons more prominent. But they don’t know the backstory. They don’t have the formative data. The Figma 2025 AI Report found 82% satisfaction among developers using AI features and 54% among designers. That gap is the story. Developers use AI to accelerate implementation of already-decided solutions. Designers are being asked to use AI to generate the solutions themselves. That’s not what the tool is built for. The skill that survives is not using the pattern. It’s knowing which pattern applies and when to break it.
Browse: Why AI UX Feels Off — Vandelay Design
Personal Tie-in: I’m building an AI Design Thinking Day for students in a district in April. Every time I try to draft the student challenge, I hit the same wall: we want students to build with AI, but building with AI before they can articulate what problem they’re solving just produces faster versions of the wrong thing. Structure first, automation second and that’s the frame I keep coming back to. The design thinking scaffold has to come before Canva Code.
So What? Teaching students to generate with AI without teaching them to design with AI produces more polished wrong answers, faster.
Try This: Make students write one sentence describing the problem they’re solving before they open any AI tool. That sentence is the constraint. Without it, they’re just generating options.
ON MY RADAR
• AI water and energy use — Cora Yang on LinkedIn — The question I get in every AI workshop, policy session, and parent Q&A. Cora Yang’s post is the most grounded, citation-backed answer I’ve found. Bookmark it. Stop fumbling the question.
• Analog Inspiration: The Expansion — Virtual Workshop — New cards this month: “deep thinking,” “integrity,” “truth,” “vulnerability.” If you’re doing any values-based AI work with students or staff, this is worth your time and feel free to join the webinar.
https://www.analoginspiration.ai/
• Nano Banana prompt this week: handwriting — Upload a sample of your handwriting. Prompt Gemini to write new content in your style. Do it once with students, then ask: “Is this your writing? What does authorship mean now?” The prompt does more pedagogical work than a week of lecture on AI ethics.
• Meenoo Rami’s book — If you’re building policy or PD this spring and need something a practitioner can actually read cover to cover, this is it. A Teacher’s Guide to Using AI
DIGITAL CHALLENGE
You’re not writing a district AI policy this week. You’re designing for one workflow.
1. Pick ONE workflow in your school where AI is currently in use or about to be. Scheduling. Feedback drafting. IEP goals. Lesson planning. Communication. Pick the one that makes you slightly uncomfortable when you think about it too hard.
2. Write down three things: Who does this workflow serve? What does failure look like? Who fills in the gap right now when something goes wrong?
3. Write one sentence describing what AI should and should not be doing in that workflow. Not a policy document. One sentence with real constraints.
4. Share it with one other person before the end of the week. Ask them: “Does this feel designed, or does it feel like we just installed something?”
You don’t need to solve the whole system. You need to design one door that actually closes.
ONE SMALL HUMAN THING
Somewhere this week, a design decision is being made about AI in your school. Probably in a budget conversation. Probably in a vendor demo. Probably without anyone calling it a design decision.
The design work matters before you know it matters.
CLOSING REFLECTION
If your school’s AI adoption were a product, which design decisions were made without anyone’s knowledge and whose needs didn’t make it into the brief?
— A-A-Ron