The Plan Was for One Child

On May 20, NPR reported that special education teachers across the country are quietly turning to artificial intelligence to help them write Individualized Education Programs, the legally binding documents that lay out what a child with a disability needs in order to learn.¹ The story drew on a Center for Democracy and Technology survey of special educators conducted across the 2024 to 2025 school year. Fifty-seven percent reported using AI to develop an IEP or 504 plan, an eighteen-point jump from the year before.² Thirty-one percent use it to identify trends in a student's progress.² Thirty percent use it to summarize the plans themselves.² Twenty-eight percent ask it to help choose which accommodations a child should receive.²

The number worth sitting with is the one underneath. Fifteen percent of the teachers surveyed said they rely entirely on AI to develop a child's IEP.¹

If you have not worked inside special education, the weight of that fifteen percent can be hard to feel. An IEP is not a templated form. Under the Individuals with Disabilities Education Act, it is, by law, a plan written for one child. Her present level of performance. Her measurable goals. The services and accommodations she will receive. The teacher knows her by face, by handwriting, by the way she circles back to a story she has told before. The plan exists because the law decided, decades ago, that what works for most children is not enough information to figure out what will work for this one.

Personalized, or Programmed

Ariana Aboulafia, who leads the Disability Rights in Technology Policy project at the Center for Democracy and Technology, was the report's lead author.¹ She did not frame the trend as a scandal. She called the tools "a Band-Aid" for a workforce that has been asked to do too much with too little for too long.¹ Her sharper line came later. Pattern recognition, she said, is "to a certain extent, inherently incompatible with a process that legally requires individualization."¹

The title of the report says the same thing more quietly. From Personalized to Programmed.² Two words that look similar in a corporate slide deck and are, in fact, opposites. Personalized means the plan begins with one child. Programmed means the plan begins with many and gets bent toward one. The first is what IDEA was written to require. The second is what a large language model is built to do.

None of this is the teachers' fault. The shortage of qualified special education staff is real. Caseloads are heavier than they have ever been. The hours an IEP swallows are hours that could have been spent next to the child for whom the IEP was drafted. When Aboulafia uses the word Band-Aid, she is not condemning the teachers. She is naming the wound.

What the Future of Education Will Have to Build

The deeper question the report poses is not whether AI can write an IEP. It can. The question is whether the document an AI drafts can carry the thing the law was trying to protect. A child's specificity. The reading goal her last teacher tried, which did not work. The breakdown in writing she has on Mondays. The two-minute pause she takes before she answers, every time, the kind of pause that an algorithm reading a hundred IEPs at once will smooth into nothing.

The future of special education, and increasingly of education writ large, will depend on whether the tools we use can hold on to that level of detail about one student at a time. Not data exhaust. Not engagement metrics. The grain of the work. The draft she abandoned. The question she finally asked her aide in the third week of October. The graph her teacher saw shift in February.

This is part of why Koan has been quietly built the way it has. Not as a replacement for the teacher who knows the child, but as a way of preserving everything that teacher would otherwise have to remember on her own. Every revision the student made. Every place she paused. Every breakthrough that arrived three days later than expected. A teacher writing an IEP next June should be able to open the record and find, in one place, the trail of evidence that proves this plan is for this child. Not a synthesis of patterns from many. The actual record of one.

The CDT report does not call for a ban. It calls for human oversight. For legal review. For the inclusion of disabled people in the design of these tools.² All of that is right. The harder ask, the one the next five years will have to answer, is whether we can build software that earns the word individualized. The tools we have now do not. They were trained on the average of many. The children IEPs were written for were never the average of anything.

If the plan in front of you was drafted for one child, would you be able to point to the moment it learned her?