Researchers from EPFL's NeuroAI Lab have successfully used a next-generation Vision-Language Model to simulate dyslexia, offering new insights into the disorder and identifying fonts that improve reading for those affected. Published as a paper at the 2026 International Conference on Learning Representations and on the arXiv preprint server, this study marks the first time an AI model has accurately mirrored the behavioral profile of a human brain impaired by dyslexia. Dyslexia, affecting an estimated 20% of the global population, is traditionally studied through behavioral tests and neuroimaging. While these methods provide valuable data, they struggle to isolate the specific underlying mechanisms of reading impairments due to ethical constraints on human experimentation. The EPFL team leveraged advanced AI capable of processing both visual and linguistic data to create a digital twin of the human brain. By identifying units within the AI model corresponding to the human visual word form area, the researchers disabled them to simulate the hypoactivation seen in dyslexic brains. The results were striking. When these specific units were ablated, the AI model struggled to read text but retained its ability to understand images and general language. This specific dissociation between reading ability and general intelligence closely mirrors the experience of humans with dyslexia. Professor Martin Schrimpf, head of the NeuroAI Lab, noted that while previous vision-only or language-only models could predict some brain activity, they lacked the capability to target dyslexia specifically. This breakthrough was only possible due to the recent exponential advances in combined Vision-Language Models. Following the simulation of the disorder, the team utilized the model to evaluate various typefaces. The AI tested both common fonts and those specifically designed to aid readers with dyslexia. The model demonstrated significantly higher accuracy when processing text in fonts created for dyslexic individuals, while performing poorly on those known to be problematic for this population. Based on these findings, the researchers are now using the AI to optimize and create the most effective font possible for people with dyslexia. Lead author Melika Honarmand highlighted the significance of the behavioral similarities. Although the AI does not share the exact biological architecture of the human brain, the fact that damaging a specific computational component produced the same functional deficit as a human neurological condition provided strong validation for the model. Beyond the immediate application to dyslexia and font optimization, the study establishes a general computational framework for investigating a wide range of brain disorders. Schrimpf explained that this approach allows researchers to perform mechanistic perturbations that would be impossible ethically in humans. The team is already applying this framework to study visual hallucinations linked to Parkinson's disease and depression. While not every clinician may immediately accept AI-based digital brains as a diagnostic tool, the researchers argue that the evidence of its utility is growing. As AI capabilities continue to evolve, such models are poised to become a crucial instrument for understanding and treating complex neurological conditions.