AI and IEP Writing: IDEA and FERPA Boundaries

Current status: Last reviewed September 7, 2026. This resource summarizes IDEA and FERPA regulations, published research, and state model policy as read on that date. It is not legal advice and does not determine whether any specific practice complies with IDEA. Special education decisions belong with the team and district counsel.

A majority of special education teachers are already doing this, and almost none of them were trained for it. Research published in October 2025 found that 57 percent of special education teachers reported using AI to help develop an IEP or 504 plan during the 2024-2025 school year, an 18 point jump in a single year, while only about 22 percent of surveyed teachers said they had received any training on risks like inaccuracy or bias.

Districts asking whether to allow this are asking a question that has already been answered in their buildings. The useful question is narrower: what does a legally sufficient plan require, what may enter a prompt, and what has to be reviewed before a draft reaches a meeting.

There is no federal guidance on the practice, which is why this guide works from the regulations that constrain it instead. The district templates that support this work are in the free district AI governance starter kit.

What IDEA requires, and what that means for a draft

Requirement	What the regulation says	Implication for AI-assisted drafting
34 CFR 300.320(a)	An IEP is a written statement for each child with a disability that is developed, reviewed, and revised in a meeting	The operative act is the meeting, not the document. AI output is pre-meeting input at most; a plan settled outside the team process is procedurally defective
34 CFR 300.320(a)(2)(i)	The IEP must contain a statement of measurable annual goals, both academic and functional	A generated goal without a baseline, a condition, and a criterion is a substantive defect, not a wording preference
34 CFR 300.320(a)(4)	Services must be based on peer-reviewed research to the extent practicable	A language model cannot attest to a research basis, and fabricated citations are a documented failure mode
34 CFR 300.321(a)	The team includes the parents, a regular education teacher, a special education teacher, an agency representative, an individual who can interpret the instructional implications of evaluation results, others with knowledge, and the child when appropriate	Membership is a list of people. A tool is not a member and cannot supply the interpretive judgment the regulation assigns to a person
34 CFR 300.324(a)(1)	In developing the IEP, the team must consider the child’s strengths, the parents’ concerns, the results of the most recent evaluation, and the child’s academic, developmental, and functional needs	Individualization is a required deliberative step. A goal produced from a disability label alone has skipped all four inputs
34 CFR 300.622(a)	Parental consent must be obtained before personally identifiable information is disclosed to parties other than officials of participating agencies	IDEA adds a consent rule on top of FERPA for special education information
34 CFR 300.623	The agency must protect confidentiality at collection, storage, disclosure, and destruction, and all persons using personally identifiable information must receive training or instruction	Untrained staff using these tools is itself a compliance gap, independent of any vendor question
34 CFR 99.31(a)(1)(i)(B)	A vendor is a school official only if it performs an institutional service the district would otherwise use employees for, is under the district’s direct control over the use and maintenance of education records, and is subject to redisclosure limits	A personal consumer account satisfies none of the three

Read together, these say something more specific than “keep a human in the loop.” They say that the determinations are the team’s, that measurability is a legal standard, and that the confidentiality rules for special education information are stricter than the general student-records rules most district AI policies were written against.

What federal guidance exists

No federal guidance specific to AI-assisted IEP drafting was located.

The July 2025 Dear Colleague Letter on artificial intelligence addresses the use of grant funds. Its privacy content is general, and it does not mention IEPs or IDEA drafting. The 2023 education technology report on AI mentions IEPs only in passing, as context a system might use or fail to consider, and its central recommendation is to establish a human in the loop as a requirement in educational applications.

The one on-point federal document has been withdrawn. An Office for Civil Rights resource on avoiding the discriminatory use of artificial intelligence included a worked example of a district that allowed schools to use generative AI to write Section 504 plans without a policy or review, where parents reported that plans looked almost identical and did not match their children’s specific needs. That document has been formally rescinded and remains online for historical purposes only.

For a district, the practical consequence is that the standard is not “follow the federal AI guidance.” It is “produce a plan that satisfies IDEA,” and to be able to show how.

What may and may not enter a prompt

The controlling question is not whether a tool is good. It is whether the vendor is a school official under FERPA and whether IDEA’s consent requirement has been handled.

May be entered, where a district agreement satisfies the school official test and forbids secondary use:

de-identified present-levels narrative;
a goal the team member wrote, submitted for tightening into measurable form;
general accommodation menus associated with a disability category, without a student attached;
plain-language translation or reading-level adjustment of text the team authored;
format, structure, and rubric checks.

May not be entered, absent an executed agreement covering it:

the student’s name, identification number, school, date of birth, or a grade and disability combination that identifies them;
evaluation scores, subtest data, or diagnostic conclusions;
disability classification tied to any identifier;
behavioral incident narratives;
parent statements from the meeting or intake;
text carried over from a prior IEP.

One asymmetry deserves attention because it defeats casual de-identification. Disability status is often inferable from the request itself. A prompt asking for a fourth-grade reading goal for a student with autism at a named elementary school discloses a disability even with the student’s name removed, because the population it describes may be one child. The data boundaries framework covers the difference between redaction, pseudonymization, and true de-identification, and the accommodations and student privacy guide works through the minimization ladder for plan-derived information.

Even with a valid agreement, IDEA requires that the people using personally identifiable information receive training. Given that roughly one in five surveyed teachers reported any training on AI risks, most districts have that gap open right now.

Documented failure modes

These are the patterns actually reported in research and reporting, rather than hypotheticals:

Near-identical plans across different students. The rescinded federal example described exactly this, with parents reporting that provisions looked almost identical and did not match their children’s needs.
Full plans generated from very little student-specific information, which is the mechanism that produces the pattern above.
Fabricated or misrepresented research. Tools sometimes invent studies or misstate findings, which collides with the peer-reviewed research provision.
Disability bias inherited from training data, where models trained on literature that poorly reflects the experiences of people with disabilities reproduce stereotypes in goals and accommodations.
Personally identifiable information entered into tools with no agreement, including evaluation results.
No training for the staff using the tools, which IDEA’s confidentiality provision does not treat as optional.

What states have said

District of Columbia draws the clearest line in its September 2026 model policy for local education agencies. Drafting IEP language sits in the limited-use tier with safeguards. Determining eligibility for an IEP or a Section 504 plan is prohibited outright. That split is the right one to borrow: expression assisted, determination reserved.

Oklahoma SB 1734, effective July 1, 2026, provides that AI may not be the primary basis for placement, promotion, retention, discipline, or grading. Placement is an IEP team determination, so the statute reaches special education directly.

California guidance directs that personally identifiable information should only ever be entered into closed AI systems, because open systems do not contain the protections required, and treats resharing as a privacy violation.

Virginia requires state guidance and local policies for AI in instructional settings, including prohibitions on relying solely on an AI system for certain high-stakes decisions, with the definitions left to the department.

The state laws and guidance tracker keeps these current.

District policy language to adopt

Counsel writes the final text. This is the structure that follows from the regulations above.

Permitted with review. Approved AI tools may assist with expressing team determinations: tightening a team-written goal into measurable form, adjusting reading level for a parent-facing summary, translating team-authored text, checking format and completeness. The staff member must verify every statement against the student’s actual record before the draft is used.

Required before use. The tool must be district approved for special education content, covered by an agreement that establishes school official status and forbids use of inputs for model training or product improvement, and the staff member must have completed the district’s confidentiality training.

Prohibited. No AI system may determine eligibility, disability classification, placement, services, or service minutes, and no AI output may serve as the primary basis for those decisions. No personally identifiable information from an education record, and no evaluation data, may be entered into any tool not covered by an executed agreement for that purpose. AI-generated citations may not be included in a plan without independent verification.

Quality controls. Before any AI-assisted draft goes to a meeting, a team member confirms the goals are measurable, the present levels match the actual evaluation and progress data, and the language reflects this student rather than a template. The district periodically samples AI-assisted plans across students to check for boilerplate repetition, which is the failure mode most likely to appear in a complaint.

Where the technical control belongs

Policy tells staff what not to type. Something has to be true when a case manager is finishing an IEP at eleven at night with a hearing next week.

Tenet by TrueMadeAI is K-12 AI governance software. Tenet Edge applies district and classroom policy on supported direct-use AI surfaces on managed Chrome. For special education work specifically, a district implementation could use Tenet to help:

run supported on-device data-loss-prevention checks so identifiers and sensitive context are caught before the text leaves the device, which is the exact exposure the regulations above create;
present the district’s rule on supported AI surfaces at the moment a staff member opens one, including a role-aware reminder for staff who handle special education records;
block supported unapproved AI chat and writing interfaces when the district enables the control, so a policy is not defeated by a personal account; and
retain bounded evidence that controls were applied, without building an archive of prompts containing student information.

Tenet does not draft plans, does not make eligibility or placement determinations, and does not replace the team, counsel, or the district’s training obligation. Coverage depends on the specific product surface; review the dated capability matrix before relying on a particular behavior. TrueMadeAI does not describe learner-specific accommodation controls as a shipped capability.

Tenet Basic is free and applies one district-wide baseline. Try Tenet Basic, or request a Tenet District conversation to map role-aware rules for special education staff.

What this guide does not establish

This guide does not determine:

whether any specific AI-assisted plan complies with IDEA;
that a vendor agreement in fact establishes school official status;
that de-identification is achievable for a given student’s information;
that a state’s guidance permits or prohibits a specific practice; or
that a technical control substitutes for team deliberation or staff training.

Those conclusions require the current official sources, the student’s facts, and qualified review.

Sources

Frequently asked questions

Is it legal to use AI to help write an IEP?

No federal law prohibits using a tool to help draft language, and no federal guidance specifically authorizes it either. What IDEA requires is that the plan be developed by the team in a meeting, that goals be measurable, and that the team consider the child’s strengths, the parents’ concerns, evaluation results, and the child’s needs. A tool may help express those determinations; it cannot make them.

Can a teacher put a student’s name and disability into an AI prompt?

Not unless the vendor is under a district agreement that makes it a school official under FERPA and IDEA consent requirements are satisfied. IDEA requires parental consent before personally identifiable information is disclosed to parties other than officials of participating agencies. Disability status is often inferable from the request itself, so removing the name alone may not be enough.

How many special education teachers are using AI on IEPs?

Research published in October 2025 by the Center for Democracy and Technology found that 57 percent of special education teachers reported using AI to develop an IEP or 504 plan during the 2024-2025 school year, an 18 point increase over the prior year, while only about 22 percent of surveyed teachers said they had received any training on risks such as inaccuracy or bias.

Has the US Department of Education issued guidance on AI and IEPs?

No guidance specific to AI-assisted IEP drafting was located. The 2025 letter on AI in schools addresses grant funds and mentions privacy generally. The 2023 education technology report discusses AI supporting instruction and recommends keeping a human in the loop. An Office for Civil Rights document that included an AI-written Section 504 plan example has been formally rescinded and remains online only for historical purposes.

What are the documented risks of AI-drafted IEPs?

The recurring findings are plans that look nearly identical across different students and do not match individual needs, goals generated from very little student-specific information, fabricated or misrepresented research citations, potential disability bias inherited from training data, and personally identifiable information entered into tools with no district agreement.

What should a district policy on AI and IEPs say?

At minimum: which approved tools may be used and on which accounts; that no personally identifiable information enters a tool absent an executed agreement; that the team, not a tool, determines eligibility, goals, services, and placement; that every AI-assisted draft is reviewed against the student’s actual data before the meeting; that staff using the tools are trained, as IDEA confidentiality rules require; and that eligibility and placement decisions are off limits entirely.

AI in IEP and 504 Writing: What IDEA Requires, What May Enter a Prompt, and What Districts Should Adopt