I actually have some accidentally poisoned data out there. Many years ago I had a job that entailed doing red carpets. There were a lot of photos of me at small to mid-sized events with two of my business partners.
Then for two years in a row my partners went to the MTV VMAs with someone else. The VMAs are MUCH bigger in terms of media footprint. That guy got tagged as me in the photos. So now a Google image search for me will at least be mixed, and often the other guy comes up first.
So the idea is sound.
Narrow models trained on a task specific data set tend to be very good at their specialization. So protien folding, or material sciences have benefitted from machine learning, but we shouldn't mistake that for being the same thing as chatGPT.
One of the bigger problems we have with AI at the moment (in my very inexpert opionion) is that they seem to be trying to throw LLMs at every problem and swearing that it'll achieve AGI soon.
Meanwhile Alpha Fold is more closely related to stable diffusion than it is to ChatGPT.