On Tuesday, Anthropic, the company behind Claude, signed a legally binding agreement with the American Federation of Teachers and the United Federation of Teachers about how it handles student and educator data. K-12 Dive reports that Anthropic must use that information only for authorized educational purposes, that schools control how long data is kept and when it is deleted, and that the terms apply to Anthropic’s existing agreements with districts. Microsoft signed on to the same union-drafted standard in September.
I think this is good news for student data privacy and AI in schools. Unions pushing vendors into written, binding commitments is exactly the kind of pressure our kids’ data deserves. AFT president Randi Weingarten called it “one step closer to an industry standard.”
But I want to ask superintendents and IT directors a harder question, the one I’d want a board member to ask me: a promise from a vendor is a promise. Who actually holds the data?
Key takeaways
- Anthropic’s new agreement with AFT and UFT, effective November 10, 2026, commits it to using student and educator data only for authorized educational purposes and lets schools control retention and deletion.
- Vendor privacy agreements are valuable, but they are terms you rely on, not controls you operate. The data still sits on someone else’s systems.
- A district-owned, on-site private AI server keeps prompts and logs on district hardware inside the district network, so the district sets the rules directly.
- Districts can use both: run open-source models locally for most work and connect commercial models by API only when and for whom they choose, under that provider’s terms.
What the agreement does, and what we don’t know yet
Here’s what has been reported. According to K-12 Dive, the agreement is aligned with the union-drafted “AI Safety and Privacy Standard for Schools.” Anthropic must include safeguards for schools and give schools and families clear information about its products, the data they collect, where they fall short and what protections exist. The Daily Caller’s coverage adds that student data cannot be sold and describes the deal as a memorandum of agreement among AFT, UFT and Anthropic.
Here’s what I haven’t seen in the coverage: specific retention periods, audit mechanics, or what happens if a term is broken. The standard itself, as K-12 Dive describes it, bars companies from training AI models on student and educator data and from selling it or using it for advertising, but the reporting describes the standard rather than spelling out each Anthropic term. Not everyone is cheering, either. Michael Marinaccio of the Center for Responsible Technology told the Daily Caller these agreements are “always too little, too late.”
I land in the middle. This is real progress. It’s also a reminder that, with almost every AI product a district buys, privacy is something you negotiate for rather than something you control.
Student data privacy and AI: a promise is not the same as possession
I’ve been working with schools on technology rollouts since 2009, starting with 1:1 laptop programs. The lesson that stuck with me wasn’t about devices. It was that whoever owns and supports the platform ends up making the decisions. AI is the same, and a privacy agreement doesn’t change where the platform lives.
Think about the questions a board member might ask after reading this week’s headlines:
- Where are our students’ prompts stored right now? With a vendor product, the answer is “on the vendor’s systems, under the contract.”
- Who can change the rules? Terms of service, default settings and features change on the vendor’s schedule. I wrote about this when Google tied Search to AI earlier this month.
- Can we prove deletion? You can request it. You can’t walk into the server room and see it.
On a district-owned private server, those answers change. Prompts and logs stay on district hardware inside the district network. The district decides retention. The district decides which roles (administrators, teachers, students) can use which models and tools. That design supports FERPA alignment; it doesn’t make anyone “automatically compliant,” and no honest provider should tell you it does. What it does is put the decisions where your board already expects them: with the district.
A concrete example: the middle school research assistant
Picture a middle school science department that wants a research assistant students can use during a unit on local water quality. On a private server, the IT director creates a student environment that only sees an open-source model running locally. Students’ questions, half-formed ideas and spelling mistakes never leave the building. The teachers’ environment, separate from the students’, can also reach a commercial model for lesson planning.
That commercial connection is the part I always want to be precise about. When a district connects a commercial model like Claude by API, those specific requests go to that provider under its API terms. Agreements like this week’s make that choice safer, and it’s worth asking every provider whether such a commitment covers your API use. But the district decides when that happens and for whom. For most student use, the answer can simply be “not at all.”
On cost, quickly: the server is a one-time purchase with no per-user fees. If a district adds commercial models, it pays the provider directly for tokens (small chunks of text, roughly three-quarters of a word each), and that usage is not included in the server price. Our Education Token Price Comparison shows how those per-token rates stack up.
What district leaders can do this month
- Inventory where student prompts live today. List every AI tool students or staff use and write down where conversation data is stored and for how long.
- Ask your current vendors about the standard. Microsoft and now Anthropic have signed on to the AFT standard. Ask the others, in writing, whether they will.
- Read your contracts for retention and training language. Look for who decides deletion timelines and whether student data can be used to improve models.
- Separate student and staff use in policy. Decide which tools students may use at all, and which data may ever leave the district.
- Bring a one-page privacy brief to your board. Put this week’s news in plain language: what vendors have promised, and what the district actually controls.
Why a private server
Vendor commitments raise the floor. Ownership raises the ceiling. A district-owned AI platform means your data, your rules, your models and your costs, without waiting for the next agreement or the next change in terms. On-site is the deployment we recommend because it keeps the hardware, and the data on it, inside your walls.
It also matters who helps you get there. The private server comes through LeadershipAIplus, a nonprofit, educator-led program: educational leaders supporting educational leaders, with districts from Kansas to Washington to Berlin already running their own servers. The purchase includes a year of virtual professional development run by LeadershipAIplus, because a privacy policy only works if people know how to live it.
If student safety is the lens your community cares about most, Tricia Friedman’s conversation on moving beyond screen-time bans on the Shifting Schools podcast is a good companion to this post. And I’ve written about why the AI off switch was never a strategy; ownership is the strategy.
FAQ
Does Anthropic’s agreement with teachers unions protect my district’s student data?
According to reporting, its terms apply to Anthropic’s existing agreements with school districts and commit it to using student and educator data only for authorized educational purposes. Check your own contract and ask Anthropic directly how it applies to the products and API use in your district.
Is a private AI server more private than a vendor with a strong privacy agreement?
It’s a different kind of protection. A vendor agreement governs how someone else handles your data. A district-owned, on-site server keeps prompts and logs on district hardware, so the district controls retention, access and deletion itself.
Can we still use Claude or other commercial models with a private server?
Yes. A district can connect commercial models by API alongside local open-source models. Those requests go to the provider under its API terms, and the district pays the provider directly for the tokens used.
The bottom line
Binding vendor promises are welcome. Owning the platform is better. One protects your data on someone else’s systems; the other keeps it on yours.
So here’s my question for your next cabinet meeting: if every AI vendor signed the standard tomorrow, would you know where your students’ prompts are stored?
If you’d like help answering that, schedule a Systems Review and we’ll walk through your current tools, data flows and options together.
Sources
- Naaz Modan, K-12 Dive, “Anthropic enters student AI privacy agreement with teachers unions,” November 10, 2026.
- Dylan Kresak, The Daily Caller, “Teachers Unions, Anthropic Strike Deal To Protect Student Data,” November 10, 2026.
- LeadershipAIplus, “Lead with Confidence in the Age of AI,” accessed November 13, 2026.
- PrivatEDU AI, “Private AI server for school districts,” accessed November 13, 2026.