This page records the public comment Basil C. Puglisi submitted to the U.S. Department of Education on docket ED-2026-IES-3268. The comment asks whether general-purpose AI reaches children with disabilities as assistive technology under IDEA, and whether school and state AI policies block that use where an IEP calls for it. Puglisi submitted it on November 14, 2026, and the Department posted it two days later as comment ED-2026-IES-3268-0004.
IDEA is the Individuals with Disabilities Education Act, and the Institute of Education Sciences had asked the public which questions should guide its next evaluations of how states and schools carry it out. Its request for information ran in the Federal Register on November 13, 2026, from the National Center for Education Evaluation and Regional Assistance (NCEE). The notice named six areas of IDEA implementation, put eight questions to commenters, and set December 11, 2026 as the date by which the Department must receive comments.
Comment as posted on Regulations.gov | Filed comment (PDF, on the docket)
The statutory definition and the functional test
As defined in IDEA, 20 U.S.C. § 1401(1)(A)
“any item, piece of equipment, or product system, whether acquired commercially off the shelf, modified, or customized, that is used to increase, maintain, or improve functional capabilities of a child with a disability”
As tested in the comment
- A disability creates an identifiable functional barrier.
- A particular AI capability directly reduces or bypasses that barrier.
- The intervention increases, maintains, or improves the person’s functional capability.
The IDEA definition asks what an item does for the child, not who designed it or how it was marketed. The first of the three working papers behind the comment, When AI Becomes Assistive Technology, names a functional category the author calls Artificially Intelligent Disability Assistance (AIDA). It sets a three-prong test for when a particular use of general-purpose AI qualifies, which the author calls the AIDA Functional Test. The third prong borrows the statutory verbs directly.

The evaluation gap
IDEA requires each IEP team to “consider whether the child needs assistive technology devices and services” (20 U.S.C. § 1414(d)(3)(B)(v)). The first evaluation question is whether that consideration now reaches general-purpose AI, and on what basis teams accept or reject it.
School AI policies written to protect assessment can reach the same functions a child with a disability needs for access. The second paper, When AI Policy Becomes an Accessibility Barrier, names six ways a policy reaches an assistive user. They are use prohibition, process restriction, disclosure requirement, distribution or opportunity penalty, discipline or exclusion, and perception enforcement. Only the first stops the use, while the other five regulate how AI use is carried out, disclosed, received, or punished.
Perception enforcement can reach a student whether or not AI was used, because detection tools act on whether work seems AI-made. Research on GPT detectors found that they “consistently misclassify non-native English writing samples as AI-generated” (Liang et al., 2023). The open question for IDEA is what that mechanism means for a child whose writing is formulaic because of a disability, or restructured by an assistive tool.
Payment is the third gap, because IDEA counts “purchasing, leasing, or otherwise providing for the acquisition of assistive technology devices” as an assistive technology service (20 U.S.C. § 1401(2)(B)). The first paper’s review identified no single general rule for who pays for a general-purpose AI subscription used as assistive technology. It also names vendor dependency, since a model retirement, a policy change, or a change in price or capability can remove an accommodation overnight.
Research questions by IDEA area
Ten research questions go to NCEE for the next five years, sorted under five of the six areas the notice names.
| IDEA area | Research questions in the comment |
|---|---|
| Access | 1. When IEP teams consider assistive technology, how often do they consider general-purpose AI tools, and what reasons do they record when they accept or reject one? 2. How consistently can different evaluators apply a functional test, such as the AIDA Functional Test, to the same use by the same child? 3. How many district and state AI policies carry an accommodation exception, or a cross-reference to the IEP and Section 504 processes? |
| Appropriate services and educational outcomes | 4. Which functional barriers show benefit from general-purpose AI, and which show harm? 5. Where does dependence emerge, and which tasks should stay with the child? 6. Do functional gains measured in the child persist across grades and transitions? |
| Procedural safeguards | 7. How do schools treat false-positive AI detection involving students with disabilities, and does a human with the time, the evidence, and the authority to overturn the flag review it before any discipline? 8. Does a requirement to disclose AI use produce different consequences for students with disabilities than for their peers? |
| Statewide systems, accountability, and flexibility | 9. When a state or local agency uses AI to triage evaluations, eligibility, or service requests, who reviews its output, and what record shows individualized review? |
| Fiscal and resource effectiveness | 10. Who pays for a general-purpose AI subscription used as assistive technology under an IEP, and what happens to the child’s services when the vendor changes its price, its policy, or its model? |
Data for future SI uses
The notice uses Super Intelligence, or SI, for what it formerly called Artificial Intelligence, or AI. Its sixth question asks how data collection could serve those uses, and the comment answers with three practices.
Measure the outcome in the child. Data built for future SI uses should record what the tool did for the child, not only which tool was used. The first paper lists candidate measures that fit this purpose:
- independent task completion
- task initiation time
- error reduction
- successful communication
- reduced dependence on another person
- completion across interrupted sessions
- user-reported autonomy
- retained decision authority
Record the human decision beside the machine output. Under Checkpoint-Based Governance, “There is no AI Governance without human authority and accountability.” At each checkpoint, the human governor accepts, modifies, or rejects what the system produced. An AI-assisted decision about a child’s evaluation, eligibility, services, or discipline is not a governed decision until a named human completes it. The record should show who decided, on what evidence, and why.
Protect what the data reveals. Adaptive systems learn sensitive facts about a person’s cognitive, psychological, and physical limits, and those facts need protection. Data collections built for SI uses face a choice about where a child’s diagnosis travels.
Products for practitioners and policymakers
Schools and agencies can take three products from the papers today. The first is a model accommodation clause for school AI policies, which the second paper offers in this text:
Nothing in this AI policy is intended to restrict the use of AI or AI-enabled functionality approved as a reasonable accommodation or other disability-related assistive technology. Requests for such use will be evaluated through the organization’s applicable disability accommodation process, subject to legitimate requirements concerning essential functions, security, confidentiality, academic integrity, and fundamental alteration.
No request is guaranteed by the clause, only that a request is treated as an accessibility question from the start. In a school that process is the IEP or the Section 504 plan.
A minimum accommodation layer is the second product, and the same paper lists the eight elements a disability-aware AI policy carries:
- an accommodation exception, or a cross-reference to the accommodation process
- individualized review
- documentation proportional to need
- approved secure alternatives
- confidential handling of accommodation information
- an appeal path
- separation of accommodation from ordinary AI enforcement
- human review before any adverse action based on detection or reports
For IEP teams, the third product is a three-prong screen. The AIDA Functional Test gives teams a plain way to separate assistance from substitution, and the first paper works a case the test excludes. A person asks AI to write a report on a subject the person never studied, then submits it unread. The use fails because the AI answered the absence of subject knowledge rather than a barrier the disability created. The boundary lets a school grant the tool without granting every use of it.
The Open University already works this way, stating that students who have shared a disability “may use Generative AI as a reasonable adjustment” (The Open University, 2025). Its condition is that the use fits the assessment’s intended learning outcomes and does not compromise academic integrity.
Beyond the six areas
Disability law reaches only people who meet its definitions, and many people with partial or episodic limitations fall short of them. A child outside IDEA eligibility may still face a real barrier that general-purpose AI reduces. The second paper states the governing principle as “Same standard, different means,” which applies as a matter of fairness before it applies as a matter of law.
A dated notice is the proposal of the third paper, A Call for an AI Disability Standard. It names the protection, names the policy or action that conflicts with it, states that AI is the person’s disability tool, and explains why. The notice has force today against covered employers, funded programs, and public entities, where the sender is a person the statute protects and the need is communicated. Force there means a duty to consider and respond under the statute’s terms, never a duty to grant the means named. Schools and state agencies sit in that first group, so an evaluation of how they respond to such notices would show whether existing protections reach AI use in practice.
Dissent from inside the disability community
The first paper carries the strongest opposition to its own argument in its own pages, and the comment brings that opposition to NCEE. The AI Now Institute report Disability, Bias, and AI argues for moving away from a narrow focus on technology’s potential to solve and to assist. It calls instead for approaches that account for social context, history, and power (Whittaker et al., 2019). The Autistic Self Advocacy Network asks organizations not to use generative AI for plain language, on tested evidence of errors that change meaning (Autistic Self Advocacy Network, 2025). Research question 4 carries that dissent into the evaluation by asking which functional barriers show harm from general-purpose AI.
Development record
Three working papers published between September and November 2026 stand behind the comment, and the record below runs from the first paper to the posted comment.
- September 23, 2026. When AI Becomes Assistive Technology named the functional category and set the three-prong test.
- September 24, 2026. When AI Policy Becomes an Accessibility Barrier classified six types of AI policy risk and closed with the model accommodation clause.
- November 8, 2026. A Call for an AI Disability Standard proposed the dated notice, to be served under protections that already exist.
- November 13, 2026. The Institute of Education Sciences published the request for information at 91 FR 73218.
- November 14, 2026. Puglisi submitted the comment through Regulations.gov as an individual.
- November 16, 2026. The Department of Education posted it as comment ED-2026-IES-3268-0004.
The same three papers are the basis for The AIDA Act: A Congressional Package, published November 19, 2026. Other policy work on this site includes the public technical comment on ISO/DIS 21520 and AI Provider Plurality: A Congressional Package.
What the comment asks
NCEE is asked to evaluate whether general-purpose AI reaches children with disabilities as assistive technology under IDEA. The comment also asks whether school and state AI policies block that use where an IEP calls for it, and it answers Questions 1, 2, 3, 6, 7, and 8 of the notice. Its research questions, outcome measures, and model clause are offered as starting material for that evaluation. The notice states that IES will not respond directly to comments.
Questions about this comment can go to me@basilpuglisi.com.
Standpoint of the comment
This comment presents governance and policy analysis informed by statutes, regulations, and institutional practice. The legal reading is that of a retired police officer with a Master of Public Administration whose studies included constitutional law, business law, procedural law, courts, and judicial case studies. It is not legal advice, it does not predict the outcome of any dispute, and no attorney reviewed it.
Sources
Autistic Self Advocacy Network. (2025, July). ASAN says no generative AI in plain language. https://autisticadvocacy.org/2025/07/asan-says-no-generative-ai-in-plain-language
Department of Education, Institute of Education Sciences. (2026, November 13). Request for information (Docket ED-2026-IES-3268, FR Doc. 2026-23155). Federal Register, 91(218), 73218 to 73219. https://www.govinfo.gov/content/pkg/FR-2026-11-13/html/2026-23155.htm
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 100779. https://doi.org/10.1016/j.patter.2023.100779
Puglisi, B. C. (n.d.). Checkpoint-Based Governance. https://basilpuglisi.com/cbg/
Puglisi, B. C. (2026a). When AI becomes assistive technology: Artificially Intelligent Disability Assistance (AIDA) [Working paper]. Zenodo. https://doi.org/10.5281/zenodo.22908271
Version 2, November 8, 2026: https://doi.org/10.5281/zenodo.24056345
Puglisi, B. C. (2026b). When AI policy becomes an accessibility barrier: ADA exposure from restrictions on Artificially Intelligent Disability Assistance (AIDA) [Working paper]. Zenodo. https://doi.org/10.5281/zenodo.22940224
Puglisi, B. C. (2026c). A call for an AI disability standard: Notice and validation for people who use AI as a disability tool, under protections that already exist [Working paper]. Zenodo. https://doi.org/10.5281/zenodo.24056037
Puglisi, B. C. (2026, November 14). Public comment on priority research questions for evaluating IDEA implementation (Comment ED-2026-IES-3268-0004). Regulations.gov. https://www.regulations.gov/comment/ED-2026-IES-3268-0004
The Open University. (2025, July 15). Generative AI for students. https://about.open.ac.uk/policies-and-reports/policies-and-statements/generative-ai-learning-teaching-and-assessment-ou-0
Whittaker, M., Alper, M., Bennett, C. L., Hendren, S., Kaziunas, L., Mills, M., Morris, M. R., Rankin, J., Rogers, E., Salas, M., & West, S. M. (2019, November). Disability, bias, and AI. AI Now Institute. https://ainowinstitute.org/wp-content/uploads/2023/04/disabilitybiasai-2019.pdf
20 U.S.C. § 1401. Definitions. https://www.law.cornell.edu/uscode/text/20/1401
20 U.S.C. § 1414. Evaluations, eligibility determinations, individualized education programs, and educational placements. https://www.law.cornell.edu/uscode/text/20/1414
#AIassisted using the HAIA Ecosystem | CC BY-NC-SA 4.0
Free for personal, educational, and noncommercial research use with attribution. Commercial exploitation, paid productization, and enterprise commercialization require separate permission and licensing.
Suggested citation: Puglisi, B. C. (2026). Public Comment on Docket ED-2026-IES-3268: General-Purpose AI as Assistive Technology Under IDEA. basilpuglisi.com.