AI training clauses: is your students' writing training someone's model?
"Processed" means the AI reads an essay, returns feedback, and retains nothing. "Trained on" means a copy of the essay becomes raw material folded into the model itself. A policy can permit the second while the marketing describes only the first, so the clause to find is the one about improving or training models.
A platform's marketing page says its AI feature reads a student's essay and hands back feedback in seconds. That sentence is true, and it is also incomplete. It describes one thing the AI can do with the essay. It says nothing about whether a copy of that essay also becomes part of what the AI itself permanently learns from.
Those are two different processes wearing the same name, "using AI." The gap between them is worth thirty seconds of your attention before you upload a student's writing anywhere.
Processed versus trained on
"Processed" means the essay goes in, the AI reads it, feedback comes out, and the essay itself leaves with the student when the session ends. Nothing about that essay is retained inside the AI's own knowledge. It is closer to a very fast, very literate assistant reading a page and handing it back.
"Trained on" means something else. A copy of the essay is added to a pool of examples used to adjust the AI model itself, so that patterns from that essay, its phrasing, its structure, its content, can shape how the model responds to other people's essays later. The student's writing does not leave when the session ends. It becomes raw material folded into the model.
A privacy policy can permit the second process while a marketing page describes only the first. Neither statement has to be false for that gap to exist. "Our AI gives instant essay feedback" and "we may use your content to improve our models" can both appear in the same platform's documents, answering different questions.
The clause to look for
Here is a pattern worth watching for, a composite built from language common across privacy policies rather than a quote from any one vendor: "we may use your content to develop and improve our services, including our machine learning models." That sentence, read carefully, is a training clause. It says content, plausibly including essays, can become training material.
Compare that to a different pattern, a real commitment rather than a permission: "we do not use customer content to train our models." That sentence closes the gap outright. The difference between the two patterns is not tone. It is whether the door to training is open or shut.
Why this matters more for essays
An application essay is among the most personal documents a teenager will produce before adulthood. It often includes family history, health struggles, identity, or the kind of failure a student would not otherwise put in writing. In my view, families bringing that material to a platform expect a grader, not a contributor to someone else's training pool, though that expectation is not universal and reasonable people can disagree about it.
That expectation is exactly what a processed-only promise satisfies and what a training clause does not. Nobody is accusing a platform of hiding this in their policy; when a policy does permit training, the permission tends to sit in service-improvement language rather than a sentence like "training AI on your child's essay."
What to ask
Three questions get past the marketing language to the actual practice.
Is essay or record content used to train AI models, in plain terms, not a page reference? If the answer involves "improve our services," ask directly whether that includes model training.
Is there an opt-out, or is training default-on for every account unless someone actively turns it off? A checkbox that exists but defaults to enrolled is a different practice than one that defaults to excluded.
Does the training promise, whichever way it goes, cover the AI subprocessors the platform itself relies on? A platform can promise not to train its own systems on student content while still sending that content to a separate AI vendor whose own terms permit exactly that.
The practical takeaway
"Our AI reads essays and gives feedback" and "your student's essay may train our model" can both be true of the same platform at the same time. The first describes what happens on screen. The second describes what happens to the copy that stays behind.
Before you treat a platform's AI feature as safely self-contained, ask in plain language whether content trains the model, whether training is opt-in or opt-out by default, and whether any AI subprocessor the platform uses is covered by the same promise. The checklist walks through this and the related questions in order. Two related reads: what a platform means when it says we never sell your data, and what a policy's use of de-identified actually means for a record that still exists somewhere.
Before the next vendor demo, print the nine-question checklist: the checklist.