What Data Actually Leaves Your Classroom When You Use AI Grading

"We use AI to grade" is a sentence that raises a fair question most teachers don't get a straight answer to: what, exactly, does that mean is leaving your classroom? Not in the abstract, privacy-policy-boilerplate sense — the actual data flow, the one time a student's writing goes somewhere outside the app. This post is that answer for Writing, no kidding, written the way a colleague would explain it rather than the way a legal document does.

This follows on from two things already covered here: what the EU AI Act's transparency rules require of a tool like this from 2 August 2026, and what "teacher always has the final say" actually means under GDPR Article 22. Both of those are about decisions — who makes them, who reviews them. This one is about data — what moves, where it goes, and what's deliberately left out.

What actually gets sent when a teacher clicks "grade" — and what's left out

There's one moment where a student's writing leaves the app: when a teacher triggers an AI feature — AI writing correction or AI rubric scoring — on a submission. At that point, two things go to the AI model:

• The student's written submission text.
• The rubric criteria the teacher defined for that assignment, when the assignment uses a rubric — the criteria titles, the levels, and the points, not the raw assignment prompt itself.

That's the request. Nothing is sent automatically or in the background — a teacher has to take an action that triggers it, submission by submission. There's no batch job quietly running overnight, no submission that gets AI-processed without a teacher having asked for that specific piece of work to be looked at.

The more interesting list is what's not in that request. No student name. No email address. No account identifier of any kind. The model that drafts a correction or scores a rubric criterion has no way to know whose writing it's looking at — it sees the text and the assessment context, and nothing that would let it, or anyone downstream, connect that text back to a specific student.

This isn't an accident of how the feature happens to work — it's the deliberate design, the same principle already touched on when we covered why "teacher decides" is only half the story: the AI shouldn't need more of a student's data than the task requires. Grading text for grammar, vocabulary, and structure doesn't require knowing who wrote it, so who wrote it isn't part of what gets sent. It's a narrower version of a principle GDPR calls data minimisation: don't collect or transmit more than the purpose actually needs, even when it would be more convenient to.

Where that request is actually processed

The model behind AI correction and rubric scoring is Google Gemini 2.5 Flash, called through Firebase Genkit using Google's standard Gemini API. Google is the processor for that specific model call; Writing, no kidding is the data processor for the platform as a whole, acting on behalf of the teacher, who remains the data controller for their own students' information. That chain matters if a data protection question ever comes up at your school — it's the same three-party structure a school's existing Google Workspace or Google Classroom use already sits inside, not a separate arrangement layered on top. It's governed by Google's Cloud data processing terms and Standard Contractual Clauses, the same mechanism that covers the rest of the platform's use of Google Cloud infrastructure.

One thing worth being straightforward about, rather than glossing over: this post isn't going to name a specific processing region and call that the final word, because that's a live, more detailed question than a blog post should try to settle on its own. If a named region for every hop of a request matters to your school's own data-residency policy, that's a fair, specific question to put to any AI grading vendor — including this one — and the answer should be a documented commitment, not a general "we use Google Cloud" reassurance.

What gets stored afterward, and for how long

Two different things get retained, on two different schedules, and it's worth not conflating them.

Student submission text — what a student actually wrote, along with the AI-drafted corrections and rubric feedback a teacher has reviewed — is kept for as long as the teacher's account stays active. It's part of the ongoing classroom record, the same way a marked-up paper essay would be until a teacher decided to clear it out.

AI usage logs are a separate, much thinner record: token counts, which model was used, and a timestamp — not the submission content itself. Those exist for billing, auditing, and quality monitoring, and every one is tagged at creation with a two-year deletion date, by design, as a data minimisation control rather than an afterthought — the intent is that old billing metadata doesn't accumulate past the point it's actually useful.

Why this is a GDPR question, not just a technical one

Two things make this arrangement hold up under GDPR rather than just being a reasonable-sounding design choice. First, the legal basis: generating AI corrections and rubric scores runs on the performance of the teacher's or school's contract with the platform — it's part of what the service is for, not a separate use of the data. The thinner usage-log record (token counts, model, timestamp) sits on a different basis, legitimate interests, since it's about billing and system monitoring rather than the assessment itself. Second, none of this changes what was already true from the Article 22 side: no AI output reaches a student without a teacher reviewing and confirming it first, so the data question here sits alongside the decision question already covered, not instead of it. Minimising what's sent to a model is a meaningfully smaller step than deciding a grade, but it's part of the same overall discipline — send less, decide nothing without a person, and be able to show both are actually true rather than just claimed.

None of this is presented as a permanent, unchangeable state of affairs — it's what's true today, documented on the Privacy Policy page rather than only described here, so it stays checkable independently of any single blog post.

What to check if you're evaluating a different tool

Not every AI grading tool will answer these the same way, and that's fine to find out before you commit a term's worth of student work to one. A few concrete questions, distinct from the decision-making ones already covered here:

• Exactly what's included in a single AI request? Not "student data" as a category — the actual fields. If a vendor can't list them, they likely haven't audited their own prompt construction closely enough to tell you.
• Is anything sent that could re-identify a student on its own? A name is the obvious one; a full email address or a unique ID string embedded in a prompt does the same job less obviously.
• Where is that request processed, specifically? "The cloud" isn't an answer. A named region is.
• What's the actual retention period, and is it enforced automatically or manually? A stated policy and an automatically enforced deletion schedule are not the same level of assurance — the second one doesn't depend on someone remembering to run a cleanup job.
• Does anything get sent to the model without a teacher having triggered that specific action? Background processing, pre-emptive grading, or batch jobs change the shape of this question considerably.

These are narrower than "is this GDPR compliant," and that's the point — they're answerable in specifics, the same way the questions worth asking about human oversight are answerable in specifics rather than reassurance.

The bottom line

If you're choosing an AI grading tool, or you already use one and haven't asked, the useful question isn't "is it GDPR compliant" — every vendor will say yes. It's narrower and more answerable: what specifically gets sent to the model, and what's deliberately excluded. For Writing, no kidding, that answer is: submission text and the rubric criteria — no names, no identifiers, no raw assignment prompt — with a teacher's action required before any of it happens at all.

---

Read the full data-handling breakdown. Read the details