I study how the brain’s language and reading systems develop and how they differ in dyslexia, using precision neuroimaging and longitudinal research. I also bring expertise in child development and literacy to the study of AI-enabled education: examining how children interact with AI tutors and designing independent trials to determine whether these tools improve learning, for whom, and under what conditions.
Through EVAL, I partner with researchers, schools, and technology developers to produce independent evidence on AI-enabled learning tools. I received the International Dyslexia Association Early Career Award and serve as President of New England Research on Dyslexia and Associate Editor of Scientific Studies of Reading.
I lead the Evidence-Based AI in Learning (EVAL) Collaborative, where we are developing rapid-cycle methods for rigorous causal research on AI-enabled learning tools. We match each study to the tool’s instructional mechanism and stage of development, from randomized tests of specific components to school-based trials of complete products. Standing school partnerships, reusable protocols, and shared measures help us produce credible findings on timelines that matter for educational decisions. Each evaluation addresses a question about one tool while building cumulative evidence about what improves learning, for whom, and under what conditions. Through the EVAL AI in Learning Challenge, we select K–12 literacy tools for independent evaluation and publish results regardless of outcome.
KIVA is an open-source, web-based, avatar-driven, multimodal AI reading companion for students with language-based disorders (e.g., dyslexia), multilingual learners, and those from disadvantaged backgrounds. It uses child-optimized speech recognition and an empirically evaluated agentic tutoring system to deliver interactive read-alouds with explicit vocabulary instruction and adaptive scaffolding. Grounded in a prior randomized controlled trial of human tutoring, KIVA's instructional design was validated through preregistered benchmarking in which AI-generated tutoring responses were rated higher than human responses on instructional quality, engagement, and responsiveness under blinded conditions. The platform logs comprehensive learning and engagement data and generates actionable reports for educators. KIVA was named a 2026 Tools Competition winner in the Accelerating K-12 Learning track. Try KIVA | Project Page | GitHub | Demo Video
Investigating the neurocognitive mechanisms underlying oral language development and their role in reading and comprehension outcomes in children with developmental dyslexia. Using precision fMRI, we demonstrated that the language network exhibits adult-like left-hemispheric lateralization by age 4, challenging the long-held progressive lateralization hypothesis. This work integrates behavioral and neuroimaging approaches to characterize how domain-specific speech perception deficits and statistical learning mechanisms contribute to literacy development.
Ozernov-Palchik, O.*, O'Brien, A. M.* et al. (2026). Precision fMRI reveals that the language network exhibits adult-like left-hemispheric lateralization by 4 years of age. Nature Communications, 17, 6505. https://doi.org/10.1038/s41467-026-72916-5
Ozernov-Palchik, O. et al. (2022). Speech-specific perceptual adaptation deficits in children and adults with dyslexia. Journal of Experimental Psychology: General, 151(7), 1556.
Ozernov-Palchik, O. et al. (2021). Distinct neural substrates of individual differences in components of reading comprehension in adults with or without dyslexia. NeuroImage, 226, 117570.
Ozernov-Palchik, O.*, Qi, Z. et al. (2023). Procedural and statistical learning in individuals with dyslexia. Neuropsychologia, 188, 108638.
Ozernov-Palchik, O., Sury, D. et al. (2023). Longitudinal changes in brain activation underlying reading fluency. Human Brain Mapping, 44(1), 18-34.
This work develops methods for answering specific causal questions about learning quickly and credibly. Different questions call for different randomized designs: whether a single instructional component changes learning is best tested with a focused experiment that varies only that component, whereas whether a complete tool improves outcomes in schools requires a trial against a meaningful comparison condition. We use Bayesian design simulations to choose sample sizes, study arms, and decision rules before data collection, and adaptive methods that allow a study to respond to evidence as it accumulates. Analyses examine how effects vary across learners and settings, so that findings speak to who benefits and under what conditions rather than only to average impact. This work involves collaboration with faculty in computer science, public health, and public policy.
Comparing highly scalable approaches to reading intervention, including fully online cross-age peer tutoring and AI-based tutoring systems. This research examines how different models can expand access to effective literacy support while maintaining instructional quality and measurable impact, including a Bayesian-designed 3-arm RCT evaluating the Amira AI reading tutor against human tutoring for 1st-2nd graders.
Olson, H. A.*, Ozernov-Palchik, O.*, Arechiga, X. M., & Gabrieli, J. D. E. (2026). Remote text-supplemented audiobook intervention supports children's explicit and incidental vocabulary learning. Developmental Science, 29(3), e70159. https://doi.org/10.1111/desc.70159
Developing AI methods to automatically quantify and evaluate pedagogical moves in naturalistic tutoring interactions. Conducted in collaboration with the National Tutoring Observatory, this work aims to characterize how instructional practices influence learning and how effective teaching strategies can be identified and scaled using multidimensional modeling and feature analysis of human and AI tutoring sessions.
Ozernov-Palchik, O., Catania, F., Gabrieli, J. D. E., & Ghosh, S. (In Preparation). Large language models deliver superior pedagogical feedback yet are penalized by source-based evaluation bias. OSF
Investigating whether advances in machine learning and automatic speech recognition can enable more precise prediction of student learning outcomes and risk trajectories. This work explores how child speech and language signals can support earlier identification of learning needs and more personalized educational support. Conducted in collaboration with MIT and Florida State University as part of a Chan Zuckerberg Initiative-supported effort. DrivenData Competition
Shangguan, Z., Huang, Z., Ozernov-Palchik, O., Ohn-Bar, E., Kosty, D., Stoolmiller, M., & Fien, H. (2024). Scalable early childhood reading performance prediction: Feasibility and challenges. NeurIPS. arXiv
Ozernov-Palchik, O.*, Sideridis, G. D., Norton, E. S., Beach, S. D., Wolf, M., Gabrieli, J. D., & Gaab, N. (2022). On the cusp of predictability: Disruption in the typical association between letter and word identification at critical thresholds of RAN and phonological skills. Learning and Individual Differences, 97, 102166.
Ozernov-Palchik, O., Norton, E. S., Sideridis, G., Beach, S., Wolf, M., Gabrieli, J. D. E., & Gaab, N. (2016). Longitudinal stability of pre-reading skill profiles of kindergarten children: Implications for early screening and theories of reading. Developmental Science, 20(5), e12471.
Examining how students who are often underrepresented in education technology research, including unhoused youth, incarcerated students, and immigrant learners, perceive and use generative AI tools. The goal is to understand how AI technologies are adopted, adapted, and experienced by students who are frequently excluded from technology design and research. In development with Boston Public Schools and Washington University.
I am committed to translating research into practice and policy. I contribute to literacy policy at the local, state, and national levels, advocating for evidence-based instruction in schools and working to ensure that educational technologies are held to the same rigorous standards we expect of other interventions that affect children's lives.
I partner with edtech and AI-enabled learning organizations, foundations, and funders to design, evaluate, and improve literacy and learning tools — turning innovation into evidence-based learning impact. Learn more about working with me →