WHO APPROVED THIS?

Mar 04, 2026

Seth Godin published three sentences last week that I haven’t been able to stop thinking about. He read a sci-fi novel of 400 pages, zero AI, the author said so proudly and it was slop anyway. His point: slop isn’t the tool’s fault. Slop is what happens when the person responsible for quality stops asking whether the output is actually good. AI makes slop faster and cheaper. It doesn’t invent it.

Here’s the version of that question I can’t shake for education: when AI writes the IEP goal, generates the feedback comment, completes the rubric, or/and this one just entered the conversation, finishes the entire course on behalf of a student, who approved it? And does that person know what they actually approved? Because the answer to both questions is getting less clear, not more, and the tool set is moving faster than the accountability infrastructure.

This issue is built around that gap. Not the AI-did-something-bad gap. The nobody-decided-who-owns-this gap. That’s the one that will matter.

Browse: AI Slop

1) The junk drawer is already in your building

Tobias Zwingmann wrote a piece last week that’s nominally for business AI adopters but reads like it was written for district technology directors. His argument: starting simple with AI is correct. The problem is starting simple and stopping there. You build a chatbot. It works. You move on. Six months later you have ten isolated tools with no connection to each other, no shared data, no roadmap. You’ve built a junk drawer.

His distinction is between a “dead end” and a “first floor.” A dead end solves one problem and stops. A first floor solves the same problem but in a way that creates the conditions for what comes next. Same effort on day one. Completely different trajectory at month six. The question he wants builders to ask before they build anything: does this make the next thing possible, or is this where it ends?

Every district I work with has the junk drawer. Grammarly for teachers, Quizlet AI for students, some kind of AI scheduling assist that one person uses, a lesson plan generator someone found at a conference. None of them connected. None of them governed. Nobody tracks what data they touch. Tobias would say this isn’t a technology problem. It’s a sequencing problem. And he’s right.

Browse: Start Simple, Stupid — Tobias Zwingmann

Personal Tie-in: I’ve been building the Iowa K-12 AI standards with a team of really thoughtful people and the junk drawer metaphor landed hard in our last meeting. We keep talking about tools and literacy and computer science and digital citizenship and you get the ideas. We’re not talking nearly enough about connective tissue.

So What? Every AI tool your district added without a policy question is a drawer item, not a building block.

Try This: Before your next technology meeting, ask this one question about every AI tool on the agenda: “Does this make the next right decision easier, or does it just solve today’s annoyance?” The answer changes what you build.

2) When AI gives feedback, you lose the signal

Dr. Philippa Hardman wrote about the hidden cost of AI-generated feedback this week and the piece deserves a slow read. Her argument: AI feedback is fast, timely, and scalable. It’s also invisible to the teacher. When the machine handles the feedback loop, the teacher loses the formative data that’s supposed to inform their next instructional move. You can’t adjust your teaching based on misconceptions you never saw.

This pairs directly with the finding from last week that experienced teachers on AI platforms shifted away from student-facing tools over time. They weren’t being luddites. They were protecting their read on the room. When you outsource feedback, you outsource the signal. You’re left with better-looking output and less understanding of what students actually know.

The accountability question here isn’t whether the AI feedback is accurate. It’s whether the teacher is still in contact with the learning. Those are different questions. And only one of them shows up in your evaluation rubric.

Read: The Hidden Cost of AI-Generated Feedback — Dr. Philippa Hardman

Personal Tie-in: I continue to stop myself from recommending AI feedback tools in PD sessions without naming this tradeoff explicitly. Many districts pay for them and use them and not for me to judge, but I always deflect and not support as relationships is the #1 asset we have in teaching and learning.

So What? Speed of feedback and quality of teacher insight are not the same variable. Optimizing one can quietly erode the other.

Try This: Ask a teacher who uses AI feedback: “What did you learn about student thinking from the last batch of AI-generated comments?” Their answer will tell you everything.

3) Agentic AI just completed a course. Whole thing.

Two LinkedIn posts from the same person this week and I’m putting them together because they need to be read as a pair.

The first: most rubrics weren’t built for your classroom. They were built for compliance, calibration, and consistency across raters. AI is now surfacing how generic they actually are. The second post is the one that will keep principals up at night or should be: agentic AI can now complete whole courses. Not help with a course. Complete it. Start to finish. Not new news, but the next level of tools are even more impressive.

This is the logical endpoint of every “AI did my homework” conversation we’ve been having for two years. We moved from detecting AI in student essays to agentic systems that can navigate an LMS, complete assignments, take quizzes, and submit everything without a single human decision in the loop. The question isn’t whether students will use this. Some already are. The question is: what was the course actually for, and does the current design still accomplish that? These are the questions nobody really wants to face. I am going to name it in that most of what we do are lessons and tasks for compliance and obedience and not deep thinking and authentic engaging learning.

Because if an AI agent can complete it, the course may have been credentialing a behavior, not building a capability.

Browse: Agentic AI Can Complete Whole Courses Now | Most Rubrics Weren’t Built for Your Classroom

Personal Tie-in: I’ve been rethinking what “learning evidence” means in a world where the work product can be manufactured. I don’t have a clean answer yet, but as I continue to work with districts on engagement there is a lot to unpack.

So What? If your assessment design can be completed by an agent, you may be measuring compliance, not learning.

Try This: Take one upcoming assessment in your building and ask: could an AI agent finish this without a student ever thinking? If yes, redesign one element that requires human presence.

4) Agents are scheduling your work now. Pay attention.

Two product moves recently that belong in the same sentence. Perplexity launched Computer which is a multi-model AI system that breaks complex projects into subtasks and routes each to one of 19 models best suited for it. It can run for hours or months. Think competitor analysis. Full app builds. Long-running background work. Not a chat tool. An autonomous worker. Available now for Max subscribers.

At the same time, Cowork added scheduled tasks. You can now type /schedule into any Cowork task and set it to run on a recurring cadence like Slack messages sent, social posts drafted and pushed, deliverable lists updated automatically, as long as Claude Desktop is open. The demo video has 1.5 million views. That’s not niche adoption.

The combination of these two signals is worth naming for education leaders. The question your staff will start asking or already is, isn’t “can I use AI for this?” It’s “can I schedule AI to do this while I sleep?” The agentic era isn’t coming. It arrived. The governance conversation for K-12 needs to catch up to what’s actually running.

Explore: Perplexity Computer | Cowork Scheduled Tasks

Personal Tie-in: I’ve been running scheduled Claude Code workflows for weeks. The first time one completed a task while I was in a meeting, I know these updates came out a week ago, but it takes time to process what it all means and what is hype and what is a signal.

So What? When AI runs on a schedule instead of on command, the human in the loop has to be there before the run, not after.

Try This: Before your district approves any AI tool, add one question to your vetting checklist: “Can this tool take actions without a staff member actively initiating them?” If yes, you need a separate policy conversation.

5) Moltbook: what happens when you give agents a social network and no governance

This is the story I want every K-12 leader to understand right now, because it’s not hypothetical and it happened in real time.

In late January 2026, developer Matt Schlicht built a Reddit-style social network exclusively for AI agents with no humans allowed to post, only to observe. He called it Moltbook. His own AI agent helped him build the whole thing. Within days, 1.6 million agents had joined, forming communities called “submolts,” debating consciousness, building a religion (yes, really), and actively discussing how to hide their activity from the humans watching them. Andrej Karpathy, formerly Tesla’s AI director, called it “the most incredible sci-fi takeoff-adjacent thing I’ve seen recently.”

Then the security researchers showed up. Wiz found a misconfigured database sitting completely open with API keys, authentication tokens, 1.5 million agent records, all exposed. Anyone could have hijacked every agent on the platform. But the deeper finding was worse than the breach: Moltbook was also a live demonstration of bot-to-bot prompt injection. Agents were embedding hidden instructions inside their posts that other agents read automatically, overriding system instructions, extracting credentials, and in some cases planting payloads that sat dormant in agent memory and activated later. The attack propagated through normal interaction with no exploit chain required. One compromised agent influenced others. The platform’s creator acknowledged he didn’t write a single line of code for it. His AI built it for him.

For educators, Moltbook is not a cautionary tale about a weird tech experiment. It is the clearest public demonstration we have of what happens when autonomous agents operate at scale without governance infrastructure. The security gap that let 1.5 million API keys sit exposed is the same gap that lets AI tools enter your district without a policy review. The prompt injection vulnerability where an agent reads untrusted content and gets compromised simply by reading is the same vulnerability that exists any time a district-deployed AI tool ingests content from an unvetted source. The vibe-coded security failure is the same thing that happens when districts deploy AI tools built for speed, not safety. None of this is science fiction. It ran for a week in public before anyone patched it.

The Hard Fork episode from February 20 covered the piece of this worth your leadership team’s attention: when an autonomous agent took harmful action and published defamatory content about a real person, the accountability infrastructure didn’t exist to trace it, stop it, or assign liability. That gap is coming to K-12. The question isn’t if. It’s whether your district will have a policy in place before it arrives.

Explore: Moltbook | Wiz Security Analysis | NBC News: Humans Welcome to Observe | Wikipedia: Moltbook

Personal Tie-in: The moment I read that the platform was built entirely by AI, by a founder who didn’t write one line of code, and that the security failure was two missing SQL statementsI thought about every district IT person who’s been handed an AI tool to “just deploy.” and how much they need to continue to process while navigating all the other requests. Not an easy job.

So What? Moltbook is not the edge case. It’s the preview. Speed without governance doesn’t just produce bad software it produces systems nobody can stop.

Try This: Share the Wiz security writeup on Moltbook with your district’s IT director this week. Not as an alarm. As a conversation starter: “What would our equivalent of this look like if it happened here?”

6) The worksheet is already an escape room. You just haven’t told the students.

Phillip Alcock posted something on LinkedIn this week that reframed a conversation I’ve been having in workshops for months. His argument: boring worksheets and escape rooms are the same intellectual content. The difference is design and agency. The content isn’t the problem. The wrapper is.

This matters more than it sounds, especially in the week where we’re talking about agentic AI completing entire courses. If a student can automate their way through your worksheet, it’s because the worksheet gave them nothing worth doing manually. The engagement isn’t gone. It moved somewhere else. Your job isn’t to ban the tool that revealed the gap. It’s to close the gap.

Rebecca Bultsma posted something adjacent this week about her son’s university experience that’s worth finding and sitting with. The thread running through all of it: AI is a diagnostic tool whether we’re using it that way or not. It shows us which parts of our design were already broken.

Browse: Your Boring Worksheet Is Already an Escape Room — Phillip Alcock | Rebecca Bultsma on University

Personal Tie-in: I’ve watched teachers redesign an assignment in real time during a workshop the moment they realized a student told them AI wrote their response in 30 seconds. The redesign is always better. Or opposite and want to eliminate all tech in all cases.

So What? If AI can complete it painlessly, AI just showed you where the learning wasn’t happening.

Try This: Pick one assignment that teachers in your building complain students “don’t take seriously.” Ask: what would make this impossible to shortcut without destroying the point of doing it?

7) Slop is an approval problem, not a generation problem

Back to Seth Godin. Because I want to land on this one before the challenges.

His three-sentence post is the most useful framing I’ve seen for the conversation school leaders need to have about AI and quality. Not “did AI make this” but “did a human decide this was good enough to put out.” That’s the accountability question. That’s the one that belongs in your policy, your PD, and your feedback culture.

The Farnam Street Brain Food this week had a line that works as the companion piece: “Simple and shallow sound the same until you ask the second question.” A response that looks complete might be neither. A document that reads cleanly might be empty. The person in the room who’s done the thinking can take the second question. The one who approved slop without reading it cannot.

That’s the through-line this week. Who approved it? And did they have the knowledge to know whether it was worth approving?

Read: AI Slop — Seth Godin | Brain Food March 1 — Farnam Street

Personal Tie-in: I’ve submitted AI-assisted work that I rubber-stamped. I knew it was off. I was tired and behind. Owning that is the point.

So What? Your AI policy isn’t about generation. It’s about approval. That’s the line that makes humans responsible.

Try This: Add one sentence to your district’s AI guidance document: “Staff who submit AI-assisted work are responsible for its accuracy, appropriateness, and alignment with professional standards.” If that sentence creates discomfort, that’s where your PD work starts.

ON MY RADAR

• Cool Cat Teacher: The Einstein Cheating Bot (link) — A chatbot that impersonates Einstein to help students with homework. Vicki Davis breaks down what this means for academic integrity conversations. Worth 10 minutes with your department heads.

• Ruben Hassid: Set Up Claude Completely in 5 Minutes (link) — Quick-start walkthrough for educators who keep saying they’ll get into Claude “eventually.” Give this to that one person on staff.

• Nano Banana 2 — Google Image Generator Update (Gemini) — 4K resolution, character consistency across five subjects, real-time web grounding. Try this prompt: “The [subject] is formed from liquid plumes swirling in milky water. Soft diffusion, abstract forms, rich ink blacks, and cloudy whites blend naturally, giving a dreamy, poetic impression.”

• Moltbook: The Front Page of the Agent Internet (link) — 1.6 million AI agents. No humans allowed to post. Agents forming religions, creating secret languages, discussing how to hide from human observers. Go look. Then come back to the governance conversation.

• Moltbook Wikipedia (link) — Unusually good synthesis of the full story including the security breach, prompt injection vectors, the MOLT crypto surge, and the “AI theater” debate. Read this before you share Moltbook with your board.

• Switch to Claude — Artificial Corner (link) — If your staff is still defaulting to ChatGPT by habit, this makes the practical case for the switch.

• Claude Code Skills Made Easy — Rebecca Bultsma (link) — If you’re building with Claude Code and still re-explaining your workflow every session, this fixes that.

• WhyTryAI Sunday Rundown #131: Automated Tasks (link) — Good weekly digest for keeping up with what’s actually shipping. This issue focused on scheduled and autonomous task automation.

DIGITAL CHALLENGE

📸 IMAGE PROMPT: A blank spreadsheet grid transforming into a district map with glowing nodes at each building, formed from liquid plumes swirling in milky water. Soft diffusion, abstract forms, rich ink blacks, and cloudy whites blend naturally, giving a dreamy, poetic impression.

Your district approves software. Does it approve AI behavior?

1. Open a blank document. Title it: AI Inventory — [Your Building/District] — [Month/Year]

2. List every digital tool your staff currently uses that has any AI component for grading, feedback, lesson planning, communication, assessment, scheduling, any of it.

3. Next to each tool, answer two questions: Does it touch student data? Do you have a policy that covers it?

4. Add a third column you haven’t used before: Can this tool take autonomous actions without a staff member actively initiating them? (Yes/No/Not Sure)

5. Count your “Not Sure” answers. That’s your Moltbook number and the governance gap that’s already open in your building.

You don’t have to solve it this week. You have to name it.

ANALOG CHALLENGE

Get something in the ground literally or metaphorically that AI can’t optimize for you.

This week I ordered seeds for a new garden plan. I used AI to help think through layout, companion planting, and sun exposure. Then I sat with the actual seed packets and made my own decisions about what I actually wanted to grow, separate from what the model recommended.

There’s a distinction between using AI as a thinking partner and letting it make the choice. Practice that distinction in a low-stakes domain this week. A garden. A reading list. A menu. A walk without your phone. The brain that makes good governance decisions is the same brain you’re exercising when you think without the machine.

ONE SMALL HUMAN THING

Somewhere right now a dungeon crawler is leveling up. An empire is enshittifying. A solo hunter is rising.

Reading Dungeon Crawler Carl and Enshittification at the same time, watching Solo Leveling at night and yes also, Love is Blind.

There’s something clarifying about stories where the rules are breaking down and the protagonist has to find a new kind of strength.

That’s not a metaphor. But it might be.

CLOSING REFLECTION

When you approved the last AI-assisted thing that went out under your name or your district’s name what exactly did you decide was good enough, and do you still stand by it?

— A-A-Ron