AI Models Mimic Dyslexia
- Researchers at EPFL built an AI system designed to replicate dyslexia, aiming to study learning disorders. - The Spanish-language report emphasised the system's use as a research instrument rather than a classroom intervention. - The work suggests AI's role may be strongest in modelling learner difficulty, not directly teaching in classrooms. (eldiario.com)
Dyslexia is a reading disorder, not a vision problem, and researchers at the Swiss Federal Institute of Technology Lausanne built an artificial intelligence model to mimic its reading failures so they could test ideas about how the disorder works. (actu.epfl.ch) The team at the NeuroAI Lab at École polytechnique fédérale de Lausanne, or EPFL, used a vision-language model — software that reads both images and text — and presented the work this week at the International Conference on Learning Representations 2026 in Rio de Janeiro. (actu.epfl.ch) (iclr.cc) Their paper, “Inducing Dyslexia in Vision Language Models,” says the researchers identified artificial units that act like the brain’s word-recognition circuitry and then disrupted those units to reproduce dyslexia-like reading problems. (arxiv.org) (openreview.net) In plain terms, the model could still understand images and language overall, but it struggled with reading after those word-processing units were altered, which the EPFL team said mirrors a core pattern seen in people with dyslexia. (actu.epfl.ch) (eldiario.com) The researchers framed the system as a lab instrument, not a classroom tutor. EPFL said behavioral tests and brain scans have helped dyslexia research, but they do not let scientists directly test cause-and-effect mechanisms in the way a computational model can. (actu.epfl.ch) (arxiv.org) That matters because dyslexia affects reading, spelling and writing and is estimated in the EPFL report to affect up to 20% of the global population, while its underlying mechanisms remain disputed enough that invasive experiments in humans are off limits. (actu.epfl.ch) The model also read better with fonts designed for dyslexic readers, according to EPFL and follow-up coverage, and lead author Melika Honarmand said the group is now studying whether it can design an even better typeface from the model itself. (eldiario.com) (medicalxpress.com) Martin Schrimpf, who leads the NeuroAI Lab, said the same framework could be extended beyond dyslexia to disorders including Parkinson’s-related symptoms and depression, a sign that the group sees these models as test beds for brain disorders more broadly. (eldiario.com) (epfl.ch) The work is also tied to a recent shift in artificial intelligence itself: Schrimpf said earlier vision-only or language-only systems were not capable enough for this project, and that newer combined models made the dyslexia simulation possible only in the past few months. (actu.epfl.ch) For now, the claim is narrower than “AI can teach dyslexic students better.” What EPFL has shown is that a model built to fail in a human-like way may give researchers a new way to probe why reading breaks down in the first place. (actu.epfl.ch) (arxiv.org)
