Gemma-Based Séance Tool Reads EEG Signals, but Audit Finds Most of the Output Comes From the Model

A developer built a séance-style console that uses a closed-form ridge probe on 630-dimensional EEG features to guess which of 200 concepts a sitter viewed, then has a local Gemma 3 4B model voice the result as a spirit message. The probe identified the correct concept in 4.5% of single-second trials, roughly nine times chance, and the shuffled-label control stayed at chance, indicating the signal is real. Accuracy rose sharply as more presentations were averaged, reaching about 39.5% at 16 presentations. However, when the message was matched against a shared concept atlas, the output tracked the sitter's actual concept in only 16.7% of sittings, falling to zero when the brain signal was swapped or removed. The author's audit concludes that much of the perceived content comes from the language model itself rather than from the brain data.