Multiple choice can be graded by a string comparison. A sentence you wrote cannot, and that is exactly the work worth doing.The AI Validator exists so that writing an answer stays possible without a teacher in the room.
Producing language and recognising it are different skills, and the corrective-feedback research finds the clearest effects on freely constructed responses rather than on picked options. Feedback that pushes you to repair your own sentence also outperformed feedback that simply restated the correct form in the classroom studies.That is the whole argument for a written item: it is the only kind of item that can show you can produce a form rather than spot it.
Exact answers never reach a model
Every written item carries a reference answer and a list of accepted alternatives. When what you typed matches one of them, the check is a plain comparison on our own server. No AI call happens, nothing is charged, and the result is the same every time.The Validator is for the rest: a paraphrase we did not list, a right meaning in an unexpected form, a small grammar slip inside an otherwise correct sentence. It receives the item, the reference answer, the accepted alternatives and the rule context. It never receives the answer key from your browser, because your browser never has it.
One repair at a time
Useful formative feedback is timely, specific and small enough to act on. So a verdict names one thing to fix and offers you the revision, rather than annotating every difference between your sentence and the reference.
Your first answer is kept. The revision is recorded as supported work beside it, not on top of it. Independent evidence and repaired evidence answer different questions, and merging them would quietly make your profile look better than your grammar is.
What it will not do
A model verdict can be wrong. It can name the wrong rule, miss a real error, or mark a correct sentence as broken. Every verdict carries a report action, and a reported one is read by a person.
Feedback aimed only at surface form has a measured cost as well as a benefit: across roughly 200 comparisons it improved surface accuracy and moved deep-level writing outcomes the other way. Treat a clean verdict as evidence about one sentence, not about your writing.
The effect sizes above belong to the studies they come from. Nothing here has been measured on Infinite Story. Whether our version of this helps a learner is an open question that the learner evaluation is meant to answer, and until then this page makes no effectiveness claim about the product.
How long your writing is kept
Written answers and the item-level feedback on them are kept for 90 days, then removed. A report you file, including any note you wrote on it, follows the same 90 days and its link back to the item disappears with it.What survives is the summary evidence: which rule, which outcome, whether the attempt was independent. You can download everything still held from your Grammar profile before it expires.
Getting the most out of it
Write the answer you would say out loud, not the one you think the matcher wants. An unusual phrasing that is correct is exactly the case the Validator is there for.
Do the repair while the item is open. That is the part with the evidence behind it.
Report a verdict you disagree with. Rule-linked feedback should fail loudly, and a report is how it does.
Read a clean verdict as one sentence graded well, not as a level.
Under the hood
The Validator runs once per submitted written response against a structured prompt containing the server-owned item, its reference answer, its accepted alternatives and the rule description. Conversation history is never included.The first independent verdict is immutable. A revision writes a separate supported record that carries its own outcome and never overwrites the original.Exact and accepted-alternative matches resolve server-side without a provider call, which is why they cost nothing.Historical release notes describe this capability as the AI tutor review, because that was its name before v0.17. Those entries stay as written and refer to what is now the AI Validator.
Sources
Lyster & Saito (2010) Oral feedback in classroom SLA: a meta-analysis. Studies in Second Language Acquisition.
Li (2010) The effectiveness of corrective feedback in SLA: a meta-analysis. Language Learning.
Shute (2008) Focus on formative feedback. Review of Educational Research.