A few weeks ago I wrote about Claude for Teachers and why the free year of access is worth signing up for. Since then, two things have caught my attention: a closer look at what "FERPA compliant" actually means when an AI company says it, and a batch of new research on what actually makes AI tutoring work in a classroom. Both are worth five minutes before you connect any AI tool to student data or hand it to students directly.
The Fine Print Behind "FERPA Compliant"
Anthropic and OpenAI have both rolled out free K-12 teacher programs this year, and both describe their privacy protections in slightly different language: a Data Processing Addendum "written to comply with FERPA," data handling "aligned with FERPA," and so on. These aren't interchangeable legal claims, and privacy researchers have flagged the gap. Amelia Vance of the Public Interest Privacy Center makes an important point that's easy to miss in the marketing copy: FERPA obligates school districts, not the ed-tech vendors selling to them. A vendor's compliance language, however it's worded, doesn't transfer that legal responsibility away from your district.
In practice, this means a teacher signing up individually and uploading a class roster, a gradebook export, or IEP notes to get a "personalized" lesson plan can create a compliance problem the district never approved and may not even know about. Before any teacher in your school connects an AI tool to real student information, check with your IT department or administration first. That's true whether the tool is Claude, ChatGPT, or anything else, and it's true no matter how confidently the vendor describes its own compliance.
What the Research Actually Shows AI Tutors Can Do
Setting privacy aside, does AI tutoring actually work? Two recent studies suggest yes, but with an important condition attached.
A Harvard physics course ran a randomized controlled trial comparing a custom-built AI tutor against an active-learning classroom, the gold standard for in-person instruction. Students using the AI tutor learned more than twice as much in less time. The tutor wasn't a generic chatbot, though. It was built through careful, content-specific prompt engineering around the exact learning objectives, problems, and feedback structure of the unit being taught.
Separately, the World Bank ran a six-week randomized trial in nine public schools in Edo State, Nigeria, giving roughly 800 secondary students after-school English tutoring through Microsoft Copilot. The results were striking: gains equivalent to one and a half to two years of typical schooling, with the strongest effects among girls in lower-scoring schools. But again, this wasn't students left alone with a chatbot. Students worked in pairs under direct teacher supervision, with a teacher introducing each topic, guiding the AI interaction, and leading a reflection exercise at the end of every session.
The Common Thread: Pedagogy, Not the Model
Neither study found that AI alone produced the gains. Both embedded the tool inside a deliberate instructional design, tight alignment with what students needed to learn, a defined role for the teacher, and structure around how the tool was used. Strip out that design and hand students an open-ended AI chat window instead, and there's no research suggesting you'd see anything close to these results.
This is the same conclusion I reach with most of the nonprofits and schools I work with: the AI tool you pick matters less than how deliberately you build the process around it. That's true for AI tutoring platforms, and it's just as true for AI use in grant writing, donor communication, or administrative work.
The Piece Still Missing: Behavior Change
One more thing worth knowing if your school is building out an AI literacy unit alongside a tutoring pilot: teaching students to understand how AI works doesn't automatically change how they use it. Educators surveyed on AI in the classroom continue to report that academic integrity is their top concern, and the researchers studying this note that the more useful question isn't just "did a student use AI to cheat," but whether real learning happened either way. Understanding concepts and changing habits, like verifying AI output instead of accepting it, are two different things, and curricula that teach one don't automatically deliver the other.
What This Means for Cochise County Schools and Nonprofits
If your school or organization is considering an AI tutor or classroom AI tool, three things are worth doing before you sign up:
First, get the vendor's exact data privacy language in writing and have your district's administration or counsel confirm what it actually commits to, rather than taking a marketing page's word for it. Second, if you're piloting AI tutoring, build it around your actual curriculum and keep a teacher actively involved rather than handing students an open chat window and hoping for the best. Third, if you're teaching AI literacy, pair it with real practice verifying AI output, not just lessons on how the technology works.
None of this is a reason to avoid these tools. The research is genuinely encouraging when the tool is used well. But "used well" is doing most of the work in that sentence, and that's exactly the part worth getting right before you roll anything out. If you'd like help thinking through a pilot for your school or organization, use the contact form to start a conversation.
