Student data: private, protected, and never used to train AI
Schools are right to be careful. In a 2025 national survey, 69% of parents said they worry about the privacy and security of the student data their child's school collects, and 23% of teachers reported a large-scale data breach at their school ([CDT, "Hand in Hand", 2025](https://cdt.org/insights/hand-in-hand-schools-embrace-of-ai-connected-to-increased-risks-to-students/)). Here is how The Learning Library handles student data, which laws we follow, and what we don't claim.
We don't train AI on your data
AI writes each shelf's study material from the book, and reads a student's written answers to give a guiding question back. None of it trains a model:
- Anthropic, which makes the Claude models we use, does not train on inputs or outputs from its commercial API (Anthropic).
- Search uses Amazon Bedrock embeddings of book text only, never student data. Bedrock doesn't use that content to improve its models or share it with model providers (AWS).
- We don't train models ourselves.
We also send less to begin with. A request carries the book's text and the student's answer, never a name, email or account id, and we strip emails and phone numbers a student types before it's sent. We keep the score of an answer, not what was written.
FERPA
When a school uses an online service, the US Department of Education's guidance says the provider can receive student records as a "school official" only if it does a job the school would otherwise do, stays under the school's direct control, and uses the records only for that job, without passing them on (PTAC, 2014). That's how we work: student progress is used to teach that student and inform their teacher, nothing else; it's never sold or used for ads; students and parents can download or delete it; and every teacher view is logged.
COPPA and younger students
COPPA protects children under 13. The FTC lets a school consent on parents' behalf for a service used solely for school, with no other commercial use (FTC COPPA FAQ). The FTC's 2025 update to the rule left this school-consent question to future rules, so it still rests on that guidance (Federal Register, 2025).
We take the stricter path for now: we ask every learner's age group first, and for anyone under 16 we store no learning history until a parent or guardian approves by email. Until then they can still read shelves, use flashcards and print tests. That also covers stricter rules for under-16s in Europe and the UK.
SOC 2: what it is, and where we are
SOC 2 is an independent CPA firm's report on a company's security, availability, confidentiality and privacy controls (AICPA). It isn't a certificate; a Type 2 report covers how controls worked over a period, usually 6 to 12 months. We haven't completed our own SOC 2 audit, and we won't say we have until a report exists. What we do have:
- Encryption in transit (HTTPS with HSTS) and at rest (database, backups and files).
- Private storage: uploaded books are never public.
- Least privilege for every part of the system; keys in a secrets manager.
- An audit log of teacher views and exports, and of history exports and deletions.
- Daily backups, kept 7 days.
What you can do
Students can download their history, delete it, leave a class, see when a teacher viewed their progress, or delete their account, all from their account page. Schools can ask us for a data privacy agreement.
How this connects to our solution
- Our privacy promises on Our solution: this post is the long version, and the Trust page is the reference.
- In the product: the join-a-class consent screen, "Who viewed your progress" and Delete account on the account page.
- Call to action: Read the Trust page · Questions? privacy@thelearninglibrary.org
