As someone also working in the imaging space, ai generated data is useful solong as it's used carefully.

Specifically, we're implementing AI culled training sets which contain some generated data that then gets reviewed manually for a few specific things, then pushed into our normal training workflows. This makes for a huge speedup versus 100% manual culling and the metrics don't lie, the models continue to improve steadily.

There may be a point where they're poisoned and will collapse, but I haven't seen it yet.