From My Planning Partner

The Margin

What's real in AI for your classrooms this week, and what's just loud. A weekly digest of the most practical findings and headlines for the K-12 education leader.

Issue No. 21Friday, November 13 · 6 min
How this is made The Margin is compiled by AI against a bar a human set. A pipeline scans a curated list of research, journalism, lab, and policy sources, scores each against a rigorous rubric for whether it should change a leader's mind, and publishes only what clears the bar. I set that bar, audit what goes out, and cut what doesn't hold up. See exactly how it works, with the data →
This week's read
Real the signature is real; the document is not public

Anthropic signed the teachers' union AI privacy standard. The terms apply to contracts districts already have.

The read The American Federation of Teachers and its New York branch, the United Federation of Teachers, announced on Tuesday November 10 that Anthropic has signed an agreement aligned with the AI Safety and Privacy Standard for Schools. K-12 Dive reports that the agreement is legally binding, that it took effect that same Tuesday, and that its terms "will apply to Anthropic's existing agreements with school districts." Anthropic is the company behind Claude for Teachers. Four requirements are named in the reporting. The company must use student and educator information "only for authorized educational purposes." Schools, not the vendor, "control how long data is kept and when it is deleted." The products must "include safeguards for schools." And the company must give schools and families information "on its products, what data is collected, where it falls short, and what protections are in place." The underlying standard, which the unions drafted in September, "bars tech companies from using student and educator data to train AI models and from selling or using that data for advertising purposes," and "also requires those companies to implement strong security protections and to design their tools to avoid harmful or manipulative experiences for students." This is the second signature on that standard. Microsoft signed in September, its agreement "went into effect Nov. 1," and K-12 Dive reports that any district nationwide that wants to add the standards to a new or current Microsoft contract "can do so without being required to renew or renegotiate." Elizabeth Kelly, head of beneficial deployments at Anthropic, said AI can only "protect and enhance" teachers' work "if teachers, students and school systems trust us," and that "These agreements provide a foundation for that trust." Randi Weingarten called the Anthropic signature "one step closer to an industry standard" and said the union hopes the protections will eventually be "embedded in law," adding: "In the almost total absence of federal and state AI regulation, we want to make certain no teacher or parent ever has to choose between AI tools and the safety and privacy of their children." Michael Mulgrew of the UFT named the reason this exists at all: "Ideally, the federal government would step in, but since it hasn't, teachers unions are working to hold AI companies accountable." Now the three places the claim breaks, because they matter for how much weight you put on it. First, every term above reaches you through the announcement and the reporting on it rather than through the signed document, which is not public. This newsletter has not read the contract text. Second, the reporting does not describe the enforcement mechanism, so who can act on a breach, on what grounds, and with what remedy is not established anywhere in it. Third, and this is the one that will get muddled fastest: on every term K-12 Dive describes, the standard governs data handling. Nothing in it, as reported, claims that any of these products helps a student learn.
2
AI companies now signed on to
the union standard: Microsoft
in September, Anthropic Nov 10
0
Contract renegotiation a district
needs to add the Microsoft terms
to an agreement it already has
0
Things in the standard about
whether any of these products
helps a student learn
The 2 counts the two signatories named in K-12 Dive's November 10 report: Microsoft in September and Anthropic on November 10. The first 0 restates that report's sentence that a district wanting the Microsoft standards added to a new or current contract "can do so without being required to renew or renegotiate," so it is a restatement of reported contract mechanics and not a measurement. The second 0 is this newsletter's characterization rather than a reported figure: across the four requirements and the underlying standard's terms as K-12 Dive describes them, every provision concerns data use, security, deletion control, product safeguards and disclosure, and none concerns learning outcomes. The signed documents themselves are not public and this newsletter has not read them, so all of the above describes the agreement as reported.
What it means for your schools There is a real action attached to this one and it is short, so start there. If your district uses Claude for Teachers, the reporting says these terms apply to agreements Anthropic already has with districts, which would mean they reach a contract you have already signed. Get your vendor contact to confirm in writing that yours is covered, because that sentence is the whole value of the announcement and it is the part worth having on paper. The reason to confirm rather than assume is Issue 13, where a vendor's public promise turned out to require the district to go and ask for it. If you are a Microsoft district you are in the asking version, and the useful detail is the one in the reporting: you can ask without reopening the whole agreement. So the item for your next cabinet meeting is a list rather than a debate. Which of our AI contracts have this standard attached, and which do not. Whoever holds the vendor relationships can answer that in an afternoon. Then notice the larger thing, which is who wrote the rule. In the absence of federal or state regulation, two major AI vendors have now accepted terms drafted by a teachers' union. Weingarten wants that in law eventually, and until it is, a negotiated clause is doing the work a statute would do. I think that is worth saying plainly rather than being reflexively sour about, and that is my read rather than anything the reporting claims. One caution, and hold it firmly, because it is the thing most likely to go wrong in the next six months. A signed privacy standard is going to start appearing in sales conversations as though it were evidence. It is not. It tells you a vendor will not train on your students or sell their data, which is worth having. It tells you nothing at all about whether the product teaches. When somebody answers a question about whether the product works by pointing at the signature, that is a different answer to a different question, and you are entitled to say so in the room.

Disclosure, kept short but owed to you here. Anthropic sells a teacher-facing AI product, which puts it adjacent to my category, and I am building My Planning Partner in that lane. This item hands a competitor-adjacent vendor a Real stamp, which is the opposite of the usual conflict and worth flagging for exactly that reason. The stamp is on the existence and terms of the agreement as reported, not on the product. Read it knowing I have a horse in this race.

K-12 Dive →

The rundown
Watch
one correlational survey, one set of 29 interviews

Two reports landed the same week on teenagers using AI for comfort. The one that found worse mental health cannot say which came first

Two separate pieces of research reached the press on November 10. The first is a report from The Jed Foundation, written by Kayla Anderson, the foundation's vice president of research and evaluation, analysing a student survey "collected between October 2025 to April 2026 from 5,560 middle and high school students in 21 schools in the United States." Education Week states the finding this way: "Students who use AI for friendship or companionship or who seek help from AI when they're struggling emotionally are more likely to have symptoms of anxiety and depression, loneliness, and a history of suicidal thoughts and attempts than students who do not use the technology for those purposes." It also found that many students use AI when they feel "sad, stressed, lonely, or overwhelmed." Education Week's language is associational throughout, so the direction of the arrow is not established, and the clearest statement of that comes from the researcher who ran it. Anderson's own reading is that "It's not a black and white conversation," and she then names the confound herself, prefacing it with "This is very early research": "what is it also replacing, or what is it about young people who are using AI in this way? Maybe there's something there, they're maybe already vulnerable." That is the devil's advocate reading, volunteered by the author, and it is the honest frame. A young person already struggling is more likely to reach for a chatbot in the first place, which would produce exactly this correlation with no harm caused by the chatbot at all. Her other finding is the one worth carrying into a staff meeting: "More than a majority of young people reporting that they're using AI when they're experiencing difficult emotions, but only a really small proportion say they're using it to get help." The second study, covered by The 74, is a Rithm Project report following up a survey the group published in March of "2,383 young people ages 13 to 24," in which "18% of this group use AI for personal and emotional support and 15% talk to AI characters and personas." Researchers went back to those two groups and analysed "29 conversations with participants ages 16 to 24." The researchers themselves caution that these interviews are not representative and rest on a self-selected group, so read them as case studies rather than a statistical portrait. What they describe is a mechanism. The shift from routine use to reliance came, in almost every case, at what the researchers call a "gateway moment," typically a rupture in a human relationship. The report's own framing, as The 74 renders it, is that AI does not have to be trustworthy to get used, only less risky than a real person in that moment. Roughly one in four of the interview subjects said they had little or no awareness of AI's problematic behaviours, one of them saying "I don't think AI's ever been off or wrong." Alison Lee of the Rithm Project on what young people are hearing from adults: "What they're hearing from adults is either 'Just don't use it, dummy,' or 'Go all in and learn all the things,'" with, she says, no meaningful third way. The headline version of all this will be that AI is harming teen mental health, and neither study can support that sentence. What they support is narrower. A meaningful number of your students talk to a chatbot when they feel worst, most of them do not think of it as asking for help, and roughly a quarter of the ones interviewed reported little or no awareness of AI's problematic behaviours. That last one is a teachable fact, it takes five minutes, and it is true whichever way the causation runs.

Education Week, the Jed report →   The 74, the Rithm report →

Watch
a complaint is an allegation, and the sourcing on the alert is unclear

A lawsuit over a 13-year-old's mental health crisis names a school-issued Chromebook. The alert reportedly went to a counselor who had already left

On November 4 the EdTech Law Center and Qureshi Law filed suit against Google in Santa Clara County Superior Court in California, on behalf of the parents of a 13-year-old girl. Education Week reports that the parents are "saying the tech giant is liable for causing their daughter to experience a psychological crisis that almost resulted in suicide and pinning at least part of the blame on her use of a school-issued Chromebook." Andrew Liddell, a lawyer for the EdTech Law Center: "As far as we know, it's the first case targeting Google's Chromebooks, which run on software that Google controls and can update when it wants to." According to the lawsuit, the girl began engaging with the chatbot after an "ordinary search," then in December "started to heavily engage with Google's Gemini-powered AI Mode for Search," and over the following months "became convinced, through her chats with the AI chatbot, that she had psychological disorders." Education Week reports that "Google officials did not immediately respond to a request for comment," and notes that "Chromebooks are the most popular laptops used in schools in the United States." Start with what this is not. A complaint is an allegation. Nothing in it has been tested, no court has found anything, and the causal story it tells is the plaintiffs' story. Treat the legal claim as unproven, because it is. But one detail in the same article is the reason this item is here, and it is worth being exact about where it comes from. Education Week writes, directly after a passage it attributes to the lawsuit and without saying whether this part also comes from the complaint, that "All of this was discovered by adults when the school's third-party monitoring software notified a former counselor at the school (who had left the school but was still receiving alerts). The counselor then notified the girl's parents, and they intervened." Read that twice. On that account the monitoring system did its job, the flag went to somebody who no longer worked there, and it reached the parents anyway because that person acted on a message they should no longer have been receiving. Whoever established it, that is a description of luck rather than of a safeguard, and unlike the rest of this story it is checkable in your own building without waiting for a court. Liddell on the parents' position: "It's horrifying," he said, "how this could happen to your daughter, without even knowing that she ever had access to a chatbot at school, much less that she was in such emotional distress." You cannot act on the lawsuit, which will take years. You can act on the alert list tonight. Every district running monitoring or filtering software has a notification configuration in an admin console somewhere, it was set up once by somebody, and staff have left since. Pull the list and read the names against your current roster.

Education Week →

Hype
on the pedagogy claim; the licensing dates are precise

Microsoft calls its new Classroom learning-science-backed and cites no learning science. The A1 retirement is the thing actually changing

Microsoft published "Future-ready education: Evolving Microsoft 365 Education" on November 5. The sentence that will travel is this one: "Microsoft Classroom brings learning science-backed tools and AI-powered experiences into one customizable environment for K-12 or higher education." Nothing in the post supports the middle of that sentence. There is no citation attached to "learning science-backed," no study named, no sample size, no comparison group and no outcome reported anywhere on the page. Set that beside what the post does state precisely, because the concrete part is a procurement event rather than a pedagogical one, and it has a date on it. Microsoft writes that most portfolio changes, "including retirement of Office 365 A1 (OA1), the Student Use Benefit program, and free Exchange Online for Alumni, take effect at an institution's first contract renewal or extension date on or after January 1, 2028, giving institutions time to plan." A new no-cost option, "Microsoft 365 Education A0," is scoped to "devices running Windows," with paid student and faculty offerings "available from Microsoft 365 Education A2-A7." Microsoft Classroom "will be available across Microsoft 365 Education A2-A7." The AI sits at the top of the stack: A7 "brings together our most advanced AI, security, governance, and device management capabilities" and the post names Microsoft Copilot, Microsoft Agent 365 and Microsoft Entra Suite. On price, the post states that "Microsoft 365 Education A2-A5 offerings provide discounts of more than 80% compared with commercial pricing" while "Microsoft 365 Education A7 offers discounts of up to 60%," so the tier with the most advanced AI is also the tier with the thinnest education discount. Note that Classroom and its "AI-powered experiences" span A2 to A7, so A7 is not the only tier with AI in it, just the top of the range. Storage changes too: "Paid student licenses will contribute to storage for the first time, and faculty contributions will increase across paid suites." Microsoft adds that "Existing agreements, pricing, licenses, and entitlements remain in effect until the applicable renewal or extension." Two documents are inside this post and only one should reach your board. The learning-science claim is marketing until somebody produces the learning science, and you are entitled to ask for it by name: which studies, what sample, measured against what. The part that belongs in a calendar is the retirement of Office 365 A1 at your first renewal on or after January 1, 2028. If you are running on A1 today, that is a budget line arriving on a known date. The post does not say A0 replaces it, and A0 is scoped to devices running Windows, which may or may not describe your fleet, so work out what your A1 users would actually land on.

Disclosure, and it is the one I always make. AI lesson planning and teaching materials are exactly my category, I am building My Planning Partner in that lane, and Microsoft Education is a direct competitor in it. I am the one awarding a Hype stamp to a competitor's pedagogical claim, so discount that accordingly. Note also that every licensing fact above is quoted from Microsoft's own post, specifically so you can check it without taking my word for any of it. Read it knowing I have a horse in this race.

Microsoft Education blog →

Watch
self-reported, and the 900 is arithmetic

Three quarters of high schoolers say they misuse technology during school hours. The 900 minutes a week is derived, not measured

The Thomas B. Fordham Institute released survey results this autumn, written up by Brian Fitzpatrick, Fordham's national research manager, "based on a nationally representative survey of over 2,000 current students and recent graduates, fielded between February and March 2026." K-12 Dive reports the headline: "Just over three-quarters of high school students report misusing technology during school hours," including "about 1 in 5 students who use artificial intelligence without teacher permission." Among students using AI, as K-12 Dive presents the figures, the report gives the uses, and this is the part worth keeping: "48% saying they use the tech to solve math problems, 36% to summarize books or avoid reading them, 29% to write first drafts of essays or reports, and 26% to help during quizzes or tests." On screen time, "Some 62% of surveyed students said they spent more than half of in-class time on school-issued devices," which K-12 Dive reports "would amount to 900 minutes a week, according to the report." Fitzpatrick's own summary names four problems at once: "rampant distraction, widespread AI misuse, heavy dependence on screens, and grading practices that make it easier to earn high marks without mastering the material." Four caveats, none of them fatal. Everything here is self-reported, which cuts in both directions on a question about rule-breaking. The 900 minutes is what K-12 Dive reports that 62 percent answer "would amount to", so it is derived from the answer rather than clocked, which makes it a plausible estimate and not an observation. The fieldwork closed in March, so the data is roughly eight months old and predates this autumn's district restrictions. And the author states his prior rather than hiding it, writing of those restrictions that "Though we might quibble about the details, we at Fordham agree with the impulse," with the report recommending stricter device bans, paper-and-pencil exams and firewalls on school devices. A think tank that favours restriction surveyed students and found facts that support restriction. That does not make the numbers wrong, and it does mean reading the recommendations as advocacy and the percentages as data. Strip the framing out and one number here is worth a staff meeting, and it is not the three quarters. It is the breakdown of what the AI is being used for, because it tells you where assignment design is failing. Maths practice first, then summarising books to avoid reading them, then first drafts. Those are three different problems with three different fixes, and calling all of it cheating guarantees you solve none of them.

K-12 Dive →

A read on trending research
Watch a university research review, so no Gate score

Stanford went looking for causal evidence on AI and learning. It found 52 studies, and 10 of them were in US K-12.

Every story above turns on something that happened in the last seven days. This is the one that asks what the whole body of evidence actually says. Researchers at the AI Hub for Education, part of Stanford University's Accelerator for Learning, "analyzed 52 randomized controlled trials and quasi-experimental studies that tested the effect of student and teacher use of generative AI on students' academic achievement in K-12 and higher education, perception of their learning experience, and social-emotional outcomes," in an update of a similar review the group ran in March of this year. Two numbers frame everything else. "Of the 1,686 papers in the AI Hub's research repository database, only 52 tested the impact of specific interventions." And "Only 10 of the 52 studies examined were conducted with U.S. K-12 students or educators, and as such the review's findings shouldn't be treated as broadly generalizable, the authors write." Now the finding you can actually use. "When students used generative AI for learning support, like generating explanations of concepts or practice problems, it generally led to better learning gains than when students used the tools to complete their work. Across the studies, though, neither use led to consistent positive outcomes." On the teacher side the picture is genuinely mixed: teacher use "can increase the quantity and quality of student feedback, even leading to improved student outcomes," and "At least one study found teacher AI use reduced student motivation and confidence." Susanna Loeb, a professor at the Stanford Graduate School of Education and an author on the report, refuses the binary directly: "You can use this tool in so many different ways. Some of these ways can be good, and some of them can be bad for student learning and development." Her analogy is the one to keep: "Is paper good or bad? Well, it depends on how you're using it." On chatbots: "Unless we're careful with chatbots, they really can take away from learning, because it is so tempting to give over the learning to the chatbot to produce something." The review also reports that students can think they are learning more with AI support even when they are not, and that when it probed which design features make a tool more effective, there were no clear answers, though "some studies suggest adaptive capabilities, interactivity, and structured guidance could improve learning." Elizabeth Tipton, a professor of statistics and data science at Northwestern who was not involved in the review, put its value in the scarcity it exposed: "Not just for the findings," she said, "but for shining a light that there are so few studies that meet the criteria for the standards we should be upholding for our schools." Her harder question is the one to carry into a board meeting: "Why are we willing to accept this risk in schools if we don't know?" On status, stated plainly: this is a research review from a university centre. This newsletter's own source rules exclude report publishers from its scored research page, so no Research Integrity Gate score was run on it and none is claimed here. Education Week does not state the review's peer-review status, so neither will I. It also pools K-12 and higher education together, and it inherits every limitation of the 52 studies underneath it.

1,686
Papers in the AI Hub's
research repository database
52
That tested the impact of a
specific intervention, and so
entered the review
10
Of those 52 conducted with
US K-12 students or educators
What the review could say
Findings the authors report
Learning-support use beat using AI to do the work
Teacher use can raise feedback quantity and quality
Adaptive features and structured guidance may help
Students can feel they learn more when they do not
Neither student use led to consistent positive outcomes
What it could not
The claim to avoid making
That the findings generalize broadly
That any feature list separates good tools from bad
That teacher AI use is safe for student motivation
That 10 US K-12 studies settle anything
It will be a long time before we can say what the effects are
All three figures are as reported by Education Week on November 10, 2026, quoting the review. The 1,686 is the stated size of the AI Hub's research repository database. The 52 is the number of those papers that tested the impact of a specific intervention and so entered the review. The 10 is the number of the 52 conducted with US K-12 students or educators, which the authors give as their own reason the findings should not be treated as broadly generalizable. The left-hand card lists findings the review reports; the right-hand card lists limits and claims the review does not support, not findings. The closing line on that card is a compression of Loeb's quoted sentence that it will be a long time before we can say what the effects of AI tools on students are. No Research Integrity Gate score is attached to this item: the Gate applies to journal-grade research, and a university centre's research review is excluded from this newsletter's scored research page by its own source rules, so no score was run and none is claimed. The papers that are scored, with their nine-dimension breakdowns and devil's-advocate passes, are on the research page.
What it means for your schools This is the most useful research item of the week, and the reason is one distinction you can act on without waiting for better evidence. Students using AI to help them learn did better than students using AI to do the work. That sounds obvious written down, and it is not how these tools are usually pitched. A product that explains a concept or generates practice sits on the better side of that line. A product that produces the artefact the student was supposed to produce sits on the worse side. Many tools can do both, which means the line is often not in the purchase at all. It is in the assignment and the instruction around it, and that part is yours rather than the vendor's. So the procurement question changes shape. Instead of asking whether a tool works, ask which side of that line it sits on by default, and what it would take to keep it there with a full class and one adult in the room. If the honest answer is constant supervision, you have learned the real price. Then hold two cautions, and they are the reason this review is worth more than its headline. The first is the finding that students can feel they are learning more when they are not, which means your own pilot's satisfaction survey is not evidence, and may well point in the wrong direction. If you are going to pilot something, decide in advance what you would measure instead, and write it down before you start. The second is the count itself, which is the actual headline, and it is Tipton's point rather than her figure. Out of 1,686 papers in the repository, 52 tested an intervention, and 10 of those were in American K-12 schools. You are being asked to make decisions for a system that has been studied roughly ten times. That is not a reason to do nothing, and it is an excellent reason to write down what you expect to happen before you buy, so that you find out whether it did.

Disclosure, and it is the one I always make. This review includes findings about teachers using AI to build lessons and materials, which is exactly my category, and I am building My Planning Partner in that lane. One of Loeb's own suggestions, as Education Week relays it rather than quotes it, points straight at what I sell: that it is possible developing lessons and materials on their own helps teachers build important skills, and that using AI for those purposes short-cuts that learning process. She also says "We do have evidence that the tools for teachers can save time and produce high-quality work," and then that "if new teachers are using the tools, sometimes the students won't like that as much, or won't be as motivated by the lessons that come out. And we aren't quite sure why." I am passing all of that on undiluted, including the part that is bad for my business. Read it knowing I have a horse in this race.

Education Week →

Apply it

Line this week up and a pattern falls out that is not about AI at all. It is about which questions are getting answered. Anthropic signed a data standard a teachers' union wrote, and the terms reach contracts districts already hold; Microsoft signed the same standard in September. Two surveys found teenagers turning to AI when they feel worst, one of them linking that use to worse mental health, and neither able to say which came first. A complaint filed on November 4 says a school-issued Chromebook was part of how a 13-year-old's crisis happened, and on Education Week's account the alert reached a counselor who had already left the building. Microsoft called its new Classroom learning-science-backed and cited no learning science. Fordham found three quarters of high schoolers saying they misuse school technology, and the breakdown of what they use AI for points at assignment design. And Stanford, looking for causal evidence, found 52 studies worth counting out of 1,686 papers, with 10 of them in US K-12. So here is the pattern, and it is one sentence said twice. The Anthropic standard, on every term reported, governs how student data is handled and says nothing about whether the product teaches. The Stanford review went looking for the thing it says nothing about, and out of 1,686 papers found 52 that tested an intervention and 10 of those in US K-12. Those are two halves of the same problem. A signed standard looks like diligence and reads like proof, and it is only the first of those. So the discipline this week asks for is keeping the two questions apart in your own mouth. Is it safe to run, and does it do anything. Both are fair questions, and this week only one of them had a document behind it.

Your one move this week Pick the thing you can finish tonight. Open your student-monitoring or filtering console and print the list of who receives alerts. Read every name against your current staff roster, then fix the two things most likely to be wrong: an alert routed to an individual's account rather than to a role that persists, and no tested fallback for when the first recipient is asleep, on leave, or gone. While you are in there, find out what happens to an alert that arrives at eleven at night, on a weekend, or in July. That audit costs nothing but the time to read a list against a roster, and this week gave you the reason to run it. Then put one line on next week's agenda, for whoever owns your vendor relationships: which of our AI contracts have the union privacy standard attached, and which do not. You do not have to act on the answer yet. You do have to know it before somebody tells you a signature is the same thing as evidence.
That's it for this week. See you next Friday. Katie
Brought to you by Planning Partner
A studio for tomorrow's lesson.

Quality AI lesson planning that lifts instructional quality and cuts teacher workload at the same time. It closes the gap between your curriculum and student-ready instruction, and gives leaders visibility into how your instructional model is actually being implemented, without piling more onto teachers.

See Planning Partner →
The Margin · A weekly AI-in-education digest for education leaders
Compiled by AI against a bar a human set
My Planning Partner · @myplanningpartner