This is my exact worry with things like chat gpt polluting the scrapable internet. The feedback loop might eventually ruin whatever value the models currently have by filling them with incorrect but plentiful generated nonsense.
I’m no business expert, but this approach seems like a huge gamble to me. Why spend years building a project before the usefulness and public reception is quantified? I’d rather spend a hard few months building something that fails rather than an easy 6 years.