Learning In The Void: The Students Nobody Designed For
Travis Gilly is the Executive Director of Real Safety AI Foundation, an AI safety, ethics, and literacy nonprofit. He can be reached at t.gilly@ai-literacy-labs.org
The AI-in-education conversation has two volumes: very loud and completely silent.
The loud part is familiar. Students are cheating. Teachers cannot tell the difference between human and machine-written essays. Cognitive offloading is eroding critical thinking. Trust between institutions and students is collapsing. Detection tools produce false positives that disproportionately flag non-native English speakers and neurodivergent writers. Everyone has an opinion. Everyone is arguing.
The silent part is fourteen percent of the student population.
Roughly one in seven students in American schools receives special education services under an Individualized Education Program. These students have legally documented cognitive, developmental, or physical differences that affect how they learn. Many of them already use AI-powered assistive technology written directly into their IEPs. Text-to-speech. Speech-to-text. Predictive text. Organizational tools. The technology is not new to them. What is new is that the entire education system is now panicking about the same category of tools these students have been told to rely on for years.
While the rest of the conversation argues about whether AI belongs in classrooms at all, special education students are watching from the margins, wondering if anyone remembers they exist.
The Design Problem
Most educational technology is built for the median student. A general-purpose learner, roughly on grade level, without significant processing differences or accommodation requirements. Accessibility gets addressed in a later sprint, if it gets addressed at all. A contrast ratio adjustment here. A screen reader tag there. Compliance, technically achieved. Experience, fundamentally unchanged.
AI tutoring platforms follow the same pattern. The architecture assumes a student who can read standard text, process linear instructions, maintain attention through uniform interaction patterns, and demonstrate understanding through conventional written output. When a student with an IEP arrives, the platform offers the same experience with minor surface-level adjustments. The accommodation is cosmetic. The architecture remains hostile.
This matters because the accommodation is the architecture. For a student with ADHD, the difference between productive engagement and complete shutdown can depend on whether the interaction pattern provides enough novelty to sustain attention. For a student with ASD, the opposite may be true; predictability and consistency in how material is presented can be the difference between comprehension and cognitive overload. These are not preferences. They are neurological requirements, documented in legal instruments that schools are federally mandated to follow.
No mainstream AI tutoring platform treats this as a first-order design constraint. The student with an IEP is, architecturally, an afterthought.
What "Designing For" Actually Requires
I run Real Safety AI Foundation, an AI safety, ethics, and literacy nonprofit. Among other projects, we are developing an AI tutoring platform called Teacher in the Loop. I want to walk through some of its design decisions, not as a sales pitch (the platform is not yet publicly available, and this article will not change that), but because the decisions themselves illustrate what it looks like when special education students are the primary use case rather than an edge case the engineering team handles in week twelve, if ever.
I also want to be honest about which parts I am confident in, which parts are genuinely uncertain, and where the tradeoffs get uncomfortable. Anybody selling certainty about AI in education right now is lying.
Invisible Accommodations
When a student with an IEP uses Teacher in the Loop, the AI adapts its instructional approach based on the accommodations documented in that student's plan. A student whose IEP indicates ASD receives interaction patterns that prioritize consistency, predictability, and clear structure. A student whose IEP indicates ADHD receives interaction patterns that introduce more variability, novelty, and shorter feedback cycles.
The student does not see a label. Their peers do not see a label. There is no "accommodated mode" toggle. The experience simply works differently for different students, invisibly, the same way a well-designed building does not announce which features exist for wheelchair users and which exist for everyone else. The ramp is just how you get in.
This matters because stigma is one of the primary reasons students refuse accommodations they are legally entitled to. A student who has to publicly activate "accessibility mode" in front of classmates is being asked to trade dignity for support. That is not a tradeoff any student should face, and it is a tradeoff that exists only because the designers did not consider it.
What I am less certain about: whether the differentiation engine (the system that determines which interaction pattern to use) will be accurate enough in early deployment to get this right consistently. ASD presentations vary enormously. ADHD presentations vary enormously. An IEP provides a starting point, not a complete map of how a specific child's brain processes information on a specific day. There will be cases where the system gets it wrong, and the question is whether the feedback mechanisms are fast enough to correct course before the student disengages. I believe they will be. I do not yet have the data to prove it.
The Anti-Displacement Clause
Teacher in the Loop includes a contractual provision: if a school adopts the platform and subsequently reduces teaching or aide staff, they lose access. This is not a suggestion, not a best practice recommendation, not a line in a terms-of-service document nobody reads. It is a binding contractual obligation.
This exists because the pattern in educational technology is well-documented and consistent. A tool is introduced as "teacher support." Administrators observe efficiency gains. Budget pressures create incentives to reduce headcount. The tool that was supposed to augment human educators replaces them. The students who depended most on human relationships (which, disproportionately, are special education students) lose the people who understood them.
For students with IEPs, the teacher is not interchangeable with an algorithm. The teacher is the person who notices that a student's behavior changed this week. Who knows that a particular student processes verbal instructions differently on days following disrupted sleep. Who recognizes that a meltdown is not defiance but sensory overload. No AI system can replicate that contextual, relational knowledge, and any platform that implicitly encourages schools to believe otherwise is causing harm.
What I am uncertain about: enforcement. If a district reduces staff and claims the reduction was unrelated to the platform, the contractual mechanism depends on the ability to establish causation. I do not have a clean answer for this. The clause exists because structural protections are better than voluntary ethics, even imperfect structural protections. But I would be dishonest if I said enforcement will be simple.
The Parent Portal and IDEA 2004
The Individuals with Disabilities Education Act was amended in 2004 to allow IEP teams to make changes without convening a full meeting, provided the parent agrees. This was intended to reduce bureaucratic burden. In practice, it created a communication gap: changes could happen faster, but parents often received less information about what was changing and why.
Teacher in the Loop includes a parent-facing portal where parents can see, in plain language, what their child is working on, how the AI is adapting to their child's needs, and what progress looks like. Parents choose the frequency of updates. Every parent we have spoken with during development has asked for daily updates. Every single one.
That statistic is not a product feature. It is an indictment of how underserved parents of special education students feel by current communication structures. When given the option, one hundred percent of parents want to know what is happening with their child's education every single day. The fact that this is remarkable tells you everything about the baseline.
Data Privacy: Pre-Opt-In and Opt-Out by Design
The standard complaint about AI in education, from parents, educators, and advocacy organizations, is that features are activated by default with no meaningful ability to decline them. Data collection runs silently. Behavioral analytics profile students without parental awareness. The complaint is not that these systems lack features. The complaint is that families cannot escape features they never chose.
Teacher in the Loop's feature set is divided into tiers. The foundational layer (Socratic-only instruction), where the AI never gives direct answers; the hybrid privacy architecture that separates personally identifiable information from cloud-processed instructional data; and core IEP accommodation delivery) is non-negotiable, as these form the backbone of the other functions. These protect the student's rights regardless of any other configuration choice.
Everything above that baseline requires informed parental consent. During onboarding, parents receive a plain-language explanation of what each optional feature does for their specific child, within the context of that child's IEP. Cognitive profile modeling: opt-in. Process capture for formative assessment: opt-in. Behavioral pattern analysis: opt-in. If a parent has privacy concerns about any capability, they decline it. The core tutoring platform still functions. Features that parents do opt into remain revocable at any time.
This is more conservative than it needs to be from a legal compliance standpoint. COPPA is satisfied by the school's adoption process and standard parental consent mechanisms. We could, legally, run most of these features by default. We chose not to because "legally permissible" and "right" are not the same question, and the parents of special education students have been on the losing end of that distinction for long enough.
What makes me uneasy: the tradeoff between data collection and service quality. Some of the most powerful features for special education students (adaptive difficulty, learning pattern recognition, early intervention flagging) require behavioral data to function well. A parent who opts out of all data collection beyond the minimum will receive a less personalized experience for their child. That is a real cost, and I do not want to pretend it is costless. The honest answer is that some parents will make that tradeoff willingly and others will not, and both choices deserve to be respected rather than engineered around.
The Argument Nobody Is Making
The AI-in-education debate is stuck in a binary: ban it or deploy it. Restrict access or expand access. Catch cheaters or embrace transformation.
Neither side is talking about the students who need AI to function. Not as a convenience. Not as a shortcut. As an accommodation that allows them to demonstrate what they actually know through pathways their neurology can support.
A student with dysgraphia whose handwriting is physiologically painful is not "cheating" by using speech-to-text. A student with ADHD whose executive function collapses during long-form writing is not "offloading cognition" by using an organizational tool. A student with ASD who cannot parse ambiguous instructions is not "dependent on AI" for requesting that prompts be structured clearly.
These students were promised, by federal law since 1975, that their education would be individualized to their needs. Fifty-one years later, the technology to actually deliver on that promise exists. And the loudest voices in the room are arguing about whether to let anyone use it.
The question is not whether AI belongs in education. The question is whether anyone is going to build it for the students who need it most, with the care, the honesty about limitations, and the structural protections they deserve. That requires designing with those students as the starting point, not the afterthought.
This is not a pitch. Nothing in this article will make anyone money. This is an argument that special education students deserve to be designed for, and a description of what that looks like when someone actually tries.
References
Education for All Handicapped Children Act of 1975, Pub. L. No. 94-142, 89 Stat. 773 (1975). https://www.govinfo.gov/content/pkg/STATUTE-89/pdf/STATUTE-89-Pg773.pdf
Federal Trade Commission. (2024). Complying with COPPA: Frequently asked questions. U.S. Federal Trade Commission. https://www.ftc.gov/business-guidance/resources/complying-coppa-frequently-asked-questions
Individuals with Disabilities Education Improvement Act of 2004, Pub. L. No. 108-446, 118 Stat. 2647 (2004). https://www.congress.gov/108/plaws/publ446/PLAW-108publ446.pdf
National Center for Education Statistics. (2024). Students with disabilities. U.S. Department of Education, Institute of Education Sciences. https://nces.ed.gov/programs/coe/indicator/cgg