What 'Teacher Always Has the Final Say' Actually Means in AI-Assisted Grading

"Teacher always has the final say" shows up on a lot of AI grading product pages, usually in a sentence right next to a testimonial and a pricing button. It's easy to read as marketing language — the kind of line every vendor writes because every vendor knows teachers are wary of AI making decisions about their students. But underneath that sentence is a real legal question with a real answer, and knowing the answer changes what you should expect from the tool sitting between you and a student's grade.

The legal version: GDPR Article 22

GDPR Article 22 gives people the right not to be subject to a decision "based solely on automated processing" when that decision produces legal effects or similarly significantly affects them. A grade is exactly the kind of thing this applies to — it can affect a student's academic standing, so a grade produced entirely by an algorithm, with no human step in between, is the situation Article 22 exists to prevent.

The operative word is solely. Article 22 doesn't ban AI from being involved in grading. It bans AI from being the entire decision. The moment a human meaningfully reviews the output before it takes effect — actually looks at it, has the authority to change it, and sometimes does — the automated-decision-making concern falls away. The EU AI Act's Article 14 (human oversight) works alongside this: a system that materially influences an assessment outcome without genuine human oversight is treated differently than one that doesn't, which is part of what determines whether an AI grading tool counts as high-risk in the first place.

So "teacher has the final say" isn't just a reassuring phrase. It's the specific condition that keeps AI-assisted grading on the right side of Article 22. Which means it's worth checking whether a given tool actually meets that condition, or just says it does.

What "meaningful review" is not

A few patterns look like human oversight without actually being it:

• A grade that's already applied, with an "edit" option nobody uses. If the AI-suggested grade is live the moment it's generated, and the teacher's only role is to notice something's wrong and fix it, that's not review — that's an opt-out from an automated decision, which is a different thing under Article 22.
• A single "approve all" button for the whole class set. Technically a human clicked something. Practically, nothing about the output was inspected. Article 22 case guidance generally treats rubber-stamping as functionally equivalent to no review at all.
• A confidence score with no path to actually change the number. Showing a percentage next to an AI grade isn't oversight if there's no straightforward way to override it.

The pattern in all three: a human is technically in the loop, but the interface doesn't require them to do anything with that position. Oversight that can be skipped isn't oversight — it's a default that happens to have a bypass.

What a real review workflow looks like

The distinction that matters isn't "does a human exist in the process" — it's whether the system requires that human's action before anything reaches a student, or whether it merely permits it.

A workflow that actually enforces review tends to have a few concrete properties: the AI's suggested grade is visible but not editable in place — the teacher has to enter their own grade in a separate field, which means typing a number is an active decision, not a confirmation click. Nothing about the submission — grade, written feedback, corrections — reaches the student until the teacher takes an explicit "submit" action, so there's no default path where an AI-generated grade goes out on its own. And every case where a teacher's final grade differs from the AI's suggestion gets recorded, not because a single override matters in isolation, but because a record of overrides is what lets a school demonstrate, months later, that review was actually happening — not just theoretically available.

That last point matters more than it sounds like it should. A vendor can build a review step and still not be able to prove it was used. A logged override history is the difference between "teachers can review AI output" and "here's evidence that they did."

Why this is better pedagogy, not just compliance

It would be easy to frame all of this as a legal box to check, but the same design choices that satisfy Article 22 also produce better grading. An AI-drafted evaluation is a first pass, not a verdict — it catches the mechanical stuff (a subject-verb agreement error, a missing article, a repeated word) reliably, but it doesn't know that a particular student has been working on run-on sentences all semester and this essay shows real progress on that specific thing, even if the sentence structure is still imperfect. That context lives with the teacher, not the model.

Requiring the teacher to actually enter the grade — rather than accept a number — creates a moment where that context gets applied. It's a small amount of extra friction, but it's friction in the direction of better grading, not worse. The alternative — a system where AI grades flow through untouched unless something looks obviously wrong — quietly shifts judgment away from the person who has the fuller picture of the student, toward a model that only ever sees one essay in isolation.

Data minimisation is part of the same principle

"Teacher decides" is the headline, but it sits on top of a second, quieter commitment: the AI shouldn't need more of a student's data than the task requires. On Writing, no kidding, AI requests include the submission text and the rubric criteria — not the student's name, email, or any other identifier — and a request is only ever sent when a teacher takes an action that triggers it, never automatically or in the background. Less data flowing to a model is a smaller surface for something to go wrong, and it's consistent with the same instinct behind requiring teacher review: the fewer decisions the system makes on its own, the fewer things need to be gotten right by the model instead of the person actually accountable for the outcome.

What this looks like in practice on Writing, no kidding

Concretely, on this platform: the "Save rubric" action is blocked until the teacher has clicked a level for every AI-scored criterion — there's no way to save a rubric with an AI suggestion sitting unreviewed. The "Submit Review" action requires a teacher-entered grade in its own field; the AI-suggested grade is shown read-only next to it and is never itself submitted. AI-suggested rubric levels are visually marked as AI-suggested, distinct from teacher-confirmed ones, so it's clear in the interface which is which. And every time a teacher's final rubric level differs from the AI's suggestion, that override — the AI's original suggestion, the teacher's final decision, and the timestamp — is recorded. None of this is described only in a policy document; it's enforced by what the "Save" and "Submit" buttons will and won't let you do, which is the version of "teacher has the final say" that Article 22 is actually asking for.

The bottom line

"Teacher always has the final say" is a claim worth testing, not just believing. The test isn't whether a human is somewhere in the workflow — it's whether the system requires that human to act before a grade reaches a student, whether that action is more than a click, and whether there's a record showing it happened. A tool that can answer yes to all three isn't just avoiding a GDPR Article 22 problem. It's built around the same judgment that made human grading worth trusting in the first place — it's just faster to exercise now.

---

Start Free — you're always the last word on every grade. Take a Tour