From My Planning Partner

The Margin

What's real in AI for your classroom this week, and what's just loud. Compiled by a chemistry teacher turned AI engineer. Five minutes, every week.

Issue No. 01Thursday · 5 min
How this is made The Margin is compiled by AI and edited by a human. A pipeline scans a curated list of research, journalism, and lab sources, ranks what might change a teacher's mind, and drafts it. Then I read every word, set every verdict, and cut what doesn't hold up. See exactly how it works, with the data →
This week's read
Real

Your students use AI less than they think their friends do.

The read A survey of 338 University of Chicago undergrads found 60% say they use AI tools, while 90% assumed the average student does. So a 30-point gap between what's actually happening and what students think is happening. The researchers chalk it up to social-desirability bias, the same reason people lowball how much they drink and assume everyone else parties more. It's a college sample, so the 60% itself doesn't transfer to a middle school. The gap is the part worth holding onto, because it should make you suspicious of any self-reported AI number.
What it means for your room A lot of the panic in schools right now rides on the assumption that everybody's already using it, so we may as well give up and police what we can. This study is a small, useful crack in that assumption. The kids are looking around at each other and guessing high too, which is part of how a thing like this snowballs. Worth asking yourself, before the next policy meeting, whether you actually have numbers for your own building or whether you're running on the same hunch your students are.
The rundown
Real

Teachers' unions are moving to limit student-facing AI for young kids

NYSUT passed a resolution urging no 1:1 devices in PreK-2, no student-facing chatbots for kids under 16, and classroom AI that stays educator-led. A resolution isn't law, but it's usually the first place you see what'll land in district policy and contract talks a year later. And it's worth noticing this isn't coming from people who don't get the technology. It's coming from teachers. Chalkbeat →

Real

Brookings: five teacher-built rules for AI that acts human

Teachers worked up a set of lines for AI that simulates friendship, counseling, or a relationship with a kid, plus clear labeling when an AI is playing a historical figure. Alberta has already steered around 50,000 educators away from this stuff. The companion-bot thing is showing up in classrooms way ahead of any guidance on it, and this is at least something concrete to hold. Brookings →

Watch

NYC delayed its school AI guidance after 6,500 public comments

The biggest district in the country pushed its final AI policy to the fall after the draft blew up: a "traffic light" framework (assessment red, brainstorming green) drew heavy backlash, a moratorium petition, and complaints that it barely touched what students themselves are allowed to do. Word is the next version gets stricter for the youngest grades. Whatever NYC is wrestling with now tends to reach the rest of us eventually. Chalkbeat →

Watch

Google shipped class-grounded "study notebooks" in Gemini at ISTE

Google announced Gemini "study notebooks" that build lessons from your actual class context, plus a Classroom integration. That's at least pointed at the thing that matters, a tool that knows something about how your class works instead of spitting out generic photosynthesis worksheets. But it's a launch claim and nobody's run it in a real room yet. Try it on one lesson before you believe the keynote. Google →

Horse in the race, as always: this is my category. I'm building My Planning Partner to close this exact gap, so take my side-eye at a competitor's launch with a grain of salt and judge the thing on your own lesson, not on what either of us tells you.

A read on trending research
Real

AI in K-12 has almost no rigorous evidence behind it yet.

Stanford reviewed 800+ studies on AI in K-12. Only about 20 are rigorous enough to prove cause and effect, and none were run in a U.S. classroom. So when a vendor says "studies show," ask which studies.

800+
studies reviewed
~20
causal · rigorous
0
in U.S. K-12
Source: Stanford SCALE, 2026 review of the K-12 AI evidence base.
Apply it

Two groups of kids used the same model to study. The ones whose chatbot just gave them answers ended up 17% worse on the test than kids who'd had no AI at all. The ones whose tool was built to give hints instead were fine. The model didn't decide that. The design did.

Plain GPT-4
answers on demand
During practice
+48%
On the exam (no AI)
−17%
worse than no-AI students
Guardrailed tutor
hints, not answers
During practice
+127%
On the exam (no AI)
±0
same as no-AI students
Source: Bastani et al., PNAS 2025, ~1,000 high-school math students.

Your one move this weekPick one place you're letting students use AI and ask whether the tool is doing the part you actually want them to struggle with. If their work comes out polished but they never really had to think, that's the 17% group, and the fix usually isn't banning anything. It's letting the AI fetch and draft and check, and keeping the thinking on them.

That's it for this week. See you next Thursday. Katie
The Margin · A studio for the lesson
Compiled by AI, edited by a human
My Planning Partner · @myplanningpartner