Responsible AI

Smart assistance. Human judgment.

More than half of US special-education teachers already use AI somewhere in writing IEPs — up from under 40% a year earlier, per the Center for Democracy & Technology. The question isn't whether AI touches this paperwork. It's whether the product using it will show you exactly what it did, and whether a person still makes the call.

What the model sees

Your screen on one side. The text that reaches Claude on the other. Drag the handle.

Your screen

Present levels · Maya R. · case manager D. Alvarez

Maya reads grade-level text with 68% accuracy across three probes this quarter, up from 54% at the last review. She responds well to a short, chunked passage with a preview of the key words, and reads with Jordan in a small group on Tuesdays.

Guardian: Ms. Reyes, (904) 555-0142. Reports go home in Spanish.

What reaches Claude

Present levels · STUDENT_1 · case manager STAFF_1

STUDENT_1 reads grade-level text with 68% accuracy across three probes this quarter, up from 54% at the last review. She responds well to a short, chunked passage with a preview of the key words, and reads with a peer in a small group on Tuesdays.

Guardian: GUARDIAN_1, [redacted]. Reports go home in Spanish.

The map from token to name exists only in memory, for the length of one request. Sample content.

The model may. The model may never.

May with a person reviewing

Draft a present-levels section, a goal, an FBA, a BIP or a transition plan from the record.
Summarize a caseload or a student for the morning, labelled advisory.
Say, deterministically, when a selection does not match what the data supports.
Phrase an answer to a parent from facts already in the child's record, where a district turns that on.
Answer a visitor's question on this website, phrased only from the published Help pages.

Never by design, not by policy

Save anything to the record without an affirmative "I have reviewed this".
Build an eligibility category, a procedural deadline or any list with a legal consequence.
See a student's name, a peer's name or a contact identifier.
State a new fact, predict, characterize, diagnose or compare in front of a family.

The record keeps the before and after

For any section AI touched, the transparency record shows what was drafted, what the reviewer changed and when it was approved.

ai_draft09:41
Drafted from the record. The model saw STUDENT_1.
Responds well to a chunked passage with a preview of the key words.
ai_edit09:44
One sentence changed by the case manager, stored as a diff.
Responds well to a ~~chunked passage~~ short, chunked passage with a preview of the key words.
ai_approve09:45
Exactly what was on screen, signed by the reviewer. The section can now be saved.

Sample record.

Two narrow, disclosed exceptions

Family Answers.A parent cannot review AI output about their own child, so the model may not state a fact, only phrase one it was handed. Every sentence names its source in the record; a sentence it cannot ground is dropped before the parent sees it. Answers are not stored. Off by default until a district turns it on.

The Ask button on this website.It phrases answers only from our published Help pages, and every sentence is checked against its section before you see it. Its one unsourced line is a short clarifying question with no facts in it. A question carrying a student's details is refused before the AI runs, and nothing you type is kept except a question the Help pages could not answer, held 30 days with contact details removed.

Why this matters more than it sounds like it should

Independent research has found bias in AI-generated IEP content from named tools in this category, and only a small fraction of tracked special-education AI products carry an independent security assessment at all. A review gate and a visible audit trail aren't a compliance checkbox — they're the difference between a tool a teacher can trust and one they have to second-guess line by line.