I actually had a similar idea that I cam across in my own research in brain-to-video applications. You could theoretically train the model in real-time by using brain data as a reward signal, such that the model would find the image or video that most strongly produces the desired response (e.g. an increase in certain brainwave bands associated with pleasure, or alertness, or fear).

I can imagine this has some very undesirable usecases if used in the wrong way.