It is not a scandal that an EEG model performs better when it has already seen the person whose brain it is reading. It is a straightforward validation leak that, left unaddressed, quietly inflates claims and misdirects research. The work reported in *Towards Data Science*, showing how participant-aware evaluation artificially boosted performance across three machine-learning models, is the kind of methodological check that every applied AI field should conduct on itself before promising generalisable insights.
The core finding is simple: when the same individuals appear in both training and test sets, the model learns to recognise the person, not the pattern. That sounds obvious. Yet the practice persists, and the numbers in the study make the cost explicit. Performance that looked like a breakthrough under participant-unaware splits collapsed when the evaluation was done properly. This is not a failure of machine learning. It is a failure of experimental design, and it has real consequences for any field that hopes to deploy these tools in the wild.
This story connects directly to parallel problems in other domains. Consider When Rule-Based Checks Succeed at Finding but Fail at Verifying, where a tool found the right passage but could not confirm the mismatch. The same gap appears here: a model that can identify previously seen faces in EEG data is technically succeeding at a task, but it is not verifying what researchers think it is. Accuracy is not the same as understanding. Meanwhile, Waymo brings self-driving tests to Toronto streets this year, where autonomous systems must generalise to unfamiliar environments or fail. The same principle applies to EEG models: if a system cannot handle unseen participants, it is not ready for real-world deployment.
Our take is that this finding should be a standard checkpoint, not a rare critique. Any lab claiming a new performance milestone for a biophysical or behavioural model should first demonstrate participant-aware evaluation results. Without that step, the number is basically meaningless. The takeaway is specific: if you see a headline about EEG models achieving human-level recognition of faces, look for the evaluation scheme. If every participant was present in the training set, the model is likely learning individual quirks, not universal neural signatures. That distinction is the difference between a useful tool and a clean lab result that cannot survive its first encounter with a stranger.