AI Can Draft an IEP Section. It Shouldn't Be Able to Save One.

The Center for Democracy & Technology’s most recent survey found that a majority of special-education teachers in the US now use AI somewhere in writing IEPs — up sharply from the year before. That number is going to keep climbing, whether or not the software a district has adopted was built with that in mind.

Which makes the actual design question narrower than “should AI touch an IEP.” It’s already touching one, in a browser tab a teacher opened separately from whatever their district paid for. The real question is what stands between an AI-generated paragraph and a document that legally governs a child’s education.

A draft is not a document

The distinction sounds obvious stated plainly, and it disappears fast in practice. A tool that lets AI output land directly in a saved record — even with a light “edit if you want to” affordance — is asking a busy person to opt into scrutiny rather than requiring it. Most people, most of the time, won’t.

A review gate flips that default. The AI’s draft sits in front of the person responsible for it — not saved, not final — until they affirmatively approve it. That’s a small UI difference with a real consequence: it makes the default outcome “a person looked at this” instead of “nobody had to.”

What “reviewed” should actually mean

A checkbox that says “I reviewed this” is honest only if there’s something to check it against. Two things make that real:

1. The diff is visible. What did the model write, and what did the reviewer change? If a reviewer can’t see their own edit against the original, “review” is a formality.
2. Legally-sensitive lists are never generated from scratch. An eligibility category, a procedural deadline, a placement code — these should be computed deterministically from the record and the applicable rule, with AI never in a position to invent one. Independent research has already documented bias in AI-generated IEP content from named tools in this category; the fix isn’t a better prompt, it’s not letting the model originate that class of decision at all.

The audit trail is the part districts actually ask for

Ask a district what they want to know about an AI tool touching student records, and the real answer is rarely “how good is the model.” It’s “show me what it did.” A visible record — this section was AI-drafted, this is what the teacher changed, this is when they approved it — answers that question directly instead of asking a district to trust a policy document.

That’s the shape IEPeasyhub builds every AI surface around: draft, review, approve, audit — in that order, every time. More on how it works on our Responsible AI page.