Abstract
Introduction: Generative models have moved non-invasive brain decoding from closed-set classification toward the reconstruction of continuous language, natural scenes and music, reviving long-standing fears about mental privacy. The capabilities and the safeguards reported by these studies have not been synthesized together. Methods: A systematic review searched four bibliographic sources in five languages for empirical studies that used a generative model to decode mental content from functional magnetic resonance imaging, magnetoencephalography or electroencephalography. Eligibility, data extraction and synthesis were tied to three questions covering content domains, accuracy and cooperation, and privacy safeguards. Results: From 648 records, 15 studies were included. They reconstructed continuous language, perceived speech, natural scenes, faces, music and illusory percepts across all three modalities. Best reported identification or top-1 accuracy ranged from 34.4% with 8-channel EEG to 96.2% with fMRI, and language decoding failed without participant cooperation. Only four studies discussed mental privacy; the remaining eleven reported informed consent and ethics approval only. Conclusions: Non-invasive generative decoders reconstruct a broad set of mental content, but their accuracy is cooperation-gated and their privacy safeguards remain largely aspirational rather than implemented.

This publication has 51 references indexed in Scilit: