Survey connects brain signals to language generation via modern neural networks

The survey reveals that brain-to-language systems are now constrained not by neuroscience instrumentation but by the same representation bottlenecks that limit LLMs. This inversion matters: the brain becomes a validation testbed for whether AI architectures are learning language the way humans do, not just whether they produce correct outputs.

This connects directly to the evaluation robustness problem documented in the legal AI study from today (Same Scores, Different Decisions). Both papers expose a shared vulnerability: models can achieve high aggregate scores while failing on consistency and transfer. Brain decoding adds a neuroscience lens to that gap. If a brain-to-language system trained on one patient's electrode array fails to generalize to another, that's the same distribution-shift fragility we see in legal document understanding and AI-generated educational content assessment. The difference is that neuroscience offers a ground truth (actual neural activity) that educational and legal benchmarks lack, making it a rare opportunity to debug whether the architecture itself is learning robust representations or just memorizing task-specific patterns.

If published brain-to-language models trained on invasive recordings show transfer to non-invasive neuroimaging (fMRI, EEG) without retraining, that confirms the learned representations are truly language-agnostic. If they don't transfer, watch whether the authors attribute the failure to signal quality or to fundamental architectural mismatch. That distinction will tell us whether neuroscience is actually informing LLM design or just validating existing choices.