I'm by no means an expert but I've seen some of what's being used to train AI, and I have to wonder if the hallucinations aren't coming from collisions of errors that got introduced while training. For example, if you don't have accurate text descriptions of the photos you want to ingest, you use BLIP or something similar to describe the photos, knowing that BLIP was also automatically trained with a certain percentage of inaccurate descriptions. So you're using a flawed AI image description model to give flawed descriptions of the photos you're ingesting, which then introduces even more errors into the photo models. The flawed photo models generate flawed pictures with flawed descriptions, which are then ingested into the next round of models.

At what point do all of the accumulated errors in training make it so that all models are doing is hallucinating based on all of the flawed training data? Do we need to hire a bunch of humans to accurately classify the first couple of levels of what the models are ingesting or something similar, and would anyone actually pay for that to happen?