New York’s Dyslexia Bill: How A.I. Can Transform Special Education and Save Billions

Lack of literacy is one of the biggest and costliest problems in our technologically advanced society. In 2024 alone, the U.S. spent over $850 billion on general education and $120 billion on special education. As the most prevalent of all special education subcategories, dyslexia impacts an estimated 20 percent of the American population. The bottom third of readers—or the “30th Percentile” —is the threshold generally used by reading researchers for dyslexia. Due to cost constraints, only about half of those 30th percentile readers can access special services. Most students with dyslexia don’t get the service that they need. Again: Most students with dyslexia don’t get the service that they need. How is that possible?

Given the size of the problem, over 40 U.S. states have passed laws mandating dyslexia screening, and over 30 have passed laws mandating dyslexia intervention. In joining this movement belatedly, New York has the opportunity to lead with a forward-looking law. Unfortunately, the New York Individuals with Dyslexia Education Act (Bill A7101A), which is currently in committee, contains the same outdated stipulations on dyslexia screening and intervention. But it isn’t too late to course-correct, adjust and improve our approach. 

The problem with current dyslexia screening and intervention

Catching at-risk learners early is a great idea. However, the New York Individuals with Dyslexia Education bill need not require periodic testing with dyslexia screeners, as schools already administer state and standardized reading assessments on a quarterly and annual basis. A better use of resources would be for schools to provide special services to all students performing in the bottom third of their regular assessments. Current dyslexia screeners were developed in the pre-A.I. era. With artificial intelligence, there is no need to disrupt teachers’ and students’ schedules to administer, score and interpret tests. With A.I., every session is an assessment, as the technology can automatically analyze a student’s responses in both the context of their previous answers and those of their same-age peers. A responsively intelligent system can dynamically assess and train—or retrain—throughout the school year. 

Even before factoring in the over-testing and schedule disruptions that the bill current dyslexia screeners would introduce, New York schools are hard-pressed to pay for teacher training to use them, as well as staffing to translate screening data into intervention. At a conservative estimate, a translation taking just 15 minutes per pupil means a small elementary school with 500 students will have to pay for 125 hours of teacher time for post-screening/pre-intervention work alone. Exacerbating the issue, the NY Dyslexia Bill specifies that schools must use a human-specialist intervention method. How many teachers—let alone those working in special education—aren’t already burdened by the heavy workload and emotional toll of the profession? As computing technology can provide needed relief, the language of legislation should accommodate, not exclude, it.

A.I.’s role in diagnosing and managing dyslexia

Can A.I. replace human workers in dealing with language, the faculty that makes us human? Before answering this, it is essential to clarify that A.I. is not replacing anyone; it is fulfilling a task that no human can perform adequately. Dyslexia presents due to inefficient brain processing of language. In linguistics, natural language is modeled as a conglomeration of complex, interlocking systems with different types of codes—systems so complicated that much of them have not yet been fully deciphered. In research, the population with dyslexia collectively has shown deficits across all major linguistic systems. Yet the study and treatment of dyslexia primarily focuses on the phonological (sound) system, which has the smallest (and hence most manageable) basic units. 

The phonological system, which comprises sounds and sound patterns, is more than a human specialist can handle. There are over 40 consonants, vowels and diphthongs in contemporary American English. For example: Let’s just take one sound, /p/. There are over 10 million possible words containing /p/, as in the permutation /p/ + V + CC (e.g., park, where V = Vowel, C = Consonant). Next, add the possible permutations for all 40+ consonants, vowels and diphthongs. Critics might say that an adult can function with just 10 million words in their vocabulary. But linguistic ability is not about remembering specific words; it is about the brain’s ability to recognize the patterns of permutations possible in one’s native language, including future words. Natural language comprises endless combinations of sounds into syllables, syllables into words, and words into sentences and texts.

Now, consider what the student must do with these permutations: listen to a spoken word and write it down (one modality) or see a word in print and read it (another modality). The figure quickly surpasses billions when we multiply possible permutations of sounds with possible modalities and other language functions. Since any part of this linguistic system can have inefficient processes, locating them becomes a computational problem. Hence, computing technology is the solution to dyslexia.

That is why, over the past 100 years of trying to do so, no human-led intervention has successfully corrected dyslexia. The DOE’s What Works Clearinghouse does not have a single program with a significant positive effect of intervention on reading comprehension for struggling readers in the bottom third, after third grade. Published meta-analytic reviews of rigorously designed replicable clinical trials reached a similar conclusion. 

Restricting any intervention to a single approach locks us into the current state of science. New technologies may help overcome traditional obstacles, such as speed and capacity, as with dyslexia. Natural language is processed in milliseconds, which is impossible for humans to track and correct, but it is easy for computers to do. The complexity of the linguistic problem of dyslexia demands vast computing capacity beyond human ability. Yet using computers to serve millions does not cost considerably more than serving ten people.

Automation can reduce New York State’s total spending on each special ed student by 90 percent. In my local district of Hyde Park, residents pay $14 million a year to support 500 students in special ed, which amounts to ~$28,000 per pupil (on top of another $20,000+ for general ed). Still, Hyde Park special ed fails to meet state reading and math standards yearly. 

Take Texas, a state which is inarguably well ahead in dyslexia mandates. As the number of students served doubled in the past six years, Texas’ special ed budget deficit now exceeds $2.3 billion annually. The problem is evident when we look at the Lone Star State’s allocation of federal funds: about half of the money goes to staff training and technical assistance and less than half to intervention. Texas is pouring a lot of money into getting teachers equipped to deal with dyslexia, but students continue to read below grade level—because humans are not equipped for this task. And Texas is struggling to recruit enough special ed teachers to do the job. I hope the New York Dyslexia Task Force considers state-of-the-art technology that can revolutionize life for the Empire State’s residents who have dyslexia.