# Harnessing AI to Enhance Reading Comprehension & Learning Outcomes for Students with Disabilities Source: https://www.youtube.com/watch?v=8oJfejvsxfA Recap page: https://rapidrecap.app/video/8oJfejvsxfA Generated: 2025-12-15T17:42:24.604+00:00 --- ## Quick Overview The AI-based reading comprehension tool, Kai, demonstrated statistically significant improvements in students' ability to answer purpose questions compared to baseline, independent work, and adult support conditions in a randomized control trial (RCT) focused on middle and high school students with intellectual and developmental disabilities. **Key Points:** - The RCT compared student performance on reading comprehension tasks under four conditions: baseline (paper/pencil), independent work, adult support replicating typical school support, and using the Kai AI tool (with and without hints). - For the answer to the purpose question, both the Kai and Kai plus hints conditions were statistically significantly better than all three other conditions, indicating enhanced comprehension. - The core strategy embedded in Kai is 'Get the Gist,' which guides students through identifying 'who or what is the most important' and synthesizing this into a 10-15 word gist statement. - The development philosophy prioritized pedagogy first, involving educators and students with lived experience of disability from the design stage, ensuring alignment with evidence-based practices like explicit instruction and the I Do, We Do, You Do model. - The engineering team found that smaller, fine-tuned models (like a 1.7 billion parameter model) showed promise for deployment on low-power devices, achieving nearly 90% coverage compared to GPT-4 baseline after fine-tuning on anonymized student data. - One qualitative finding was that students conditioned against pencil/paper tasks (e.g., ripping up a worksheet) aced the task when using Kai, suggesting increased engagement with the digital tool. - The adult support condition surprisingly scored lower than independent work on the purpose question in the initial analysis, suggesting that unguided adult prompting can sometimes provide incorrect support. **Context:** The presentation introduces Kai, an AI co-pilot agent designed by researchers including Lakshmi Balasubramanian and Chris Lemons from Stanford's Graduate School of Education (GSE), along with engineer Utkarsh, to enhance literacy skills, specifically reading comprehension and writing, for middle and high school students with intellectual and developmental disabilities (IDD). The project received seed funding from HAI and is grounded in extensive pedagogical research emphasizing inclusive education, UDL principles, and evidence-based strategies like explicit systematic instruction and the 'Get the Gist' methodology. ## Detailed Analysis The team developed Kai as an AI tool to address the writing proficiency gap and limited employment outcomes often faced by students with IDD, aiming to allow them to be educated alongside typical peers. Kai integrates research-backed pedagogy, including explicit instruction, modeling ('I Do'), guided practice ('We Do'), and independent work ('You Do'), using the 'Get the Gist' strategy which requires students to identify the 'who or what' and the 'most important information' in a paragraph to create a summary statement. The tool incorporates Universal Design for Learning (UDL) principles, ensuring variability is the norm and providing scaffolding such as hints, sentence starters, and multiple-choice options to manage cognitive load and prevent frustration during productive struggle. Utkarsh detailed the engineering challenges, emphasizing that a pedagogy-first approach required extensive collaboration with educators to constrain the LLM's unconstrained responses, focusing on data privacy (no personal student information stored centrally) and reliability. The RCT data showed that while initial gist statements (paragraphs 3 and 4) showed only marginal improvement over baseline/adult support, the crucial measure—answering the overall purpose question—was statistically significantly better when using Kai (with or without hints) compared to baseline, independent work, or adult prompting alone. The researchers noted that future work involves fine-tuning smaller, open-source models for better deployment on low-resource devices and expanding multimodality and multilingual support. ### KAI Tool Philosophy and Pedagogy - KAI is an AI co-pilot enhancing literacy for students with learning differences - Philosophy centers on disability existing at the intersection of the individual and environment, requiring context scaffolding - It integrates explicit instruction, clear language, modeling, and corrective feedback, aligning with UDL principles. ### The 'Get the Gist' Strategy - Strategy involves four pre-reading steps: set purpose, select text, provide background knowledge ('instructional velcro'), and pre-teach vocabulary - During reading, students identify 'who or what' and 'most important information' to write a gist statement (10-15 words). ### AI Implementation and Engineering - Initial reliance on large models (GPT-4) provided stability but incurred high cost and latency; smaller, open-source models like Quen showed promise when fine-tuned with anonymized student data - Development mandated a pedagogy-first design, extensive prompt engineering to coax desired pedagogical responses, and rigorous focus group testing. ### Randomized Control Trial (RCT) Results - Kai significantly improved performance on the final purpose question compared to baseline, independent work, and adult support conditions - Adult support condition scored surprisingly lower than independent work on the purpose question, suggesting potential issues with on-the-fly prompting accuracy. ### Scaffolding and User Experience - Kai utilizes an I Do (model), We Do (guided practice/hints), You Do (independent attempt) structure, providing escalating levels of support, including sentence starters and multiple-choice options if initial attempts fail - Gamification, including stars and an octopus sidekick, was incorporated after focus groups confirmed student preference.