Scientists just published a stocktake of every serious attempt to turn brainwaves into pictures, sentences, and sound - and the technology is still nowhere close to mind reading.
The survey combed through research published between 2017 and 2025 on systems that convert electroencephalography (EEG) signals, the electrical activity picked up by scalp electrodes, into images, open-vocabulary text, or audio. Most image-generating systems pair an encoder-decoder setup with generative adversarial networks, variational autoencoders, or diffusion models. Text systems increasingly lean on transformer-based language models, the same architecture behind modern chatbots, to produce free-form output instead of picking from a fixed word list. Audio systems typically convert EEG signals into mel-spectrograms, a visual map of sound frequencies, which a neural vocoder then turns into actual sound.
The real constraint isn't the model architecture, it's the data. EEG datasets stay small and inconsistent between studies, and models trained on one person's brain signals often fail to generalize to anyone else's - a serious problem for anything meant to leave the lab. There's also no standardized benchmark, so claims of progress from one paper to the next are hard to compare directly.
This is a literature review, not a new capability - useful mainly as a map of what's actually been built versus what press releases about 'reading minds' tend to imply.