Researchers have developed a model called Brain2Qwerty v2 that can accurately decode natural sentences from non-invasive brain recordings. The model uses magnetoencephalography (MEG) recordings to achieve an average word error rate of 39%. While this is still lower than the accuracy of intracranial implants, the results suggest that data scaling could partially bridge the performance gap. The study shows that AI can improve decoding accuracy through deep learning and fine-tuning large language models. --- Why it matters: This matters to researchers in AI because it demonstrates the potential for non-invasive brain-computer interfaces to achieve high levels of accuracy, which could lead to new treatments for individuals with speech or movement disorders. Source: https://arxiv.org/abs/2608.18114