One of my first thoughts when GPT-3 came out was that the value of "curated gardens" of quality data (wikipedia, SO) is going to become immensely more valuable because of this problem. If you pollute the source of training data it eventually becomes worthless for training better models in the future