I don't know that this take is wrong, per se, but I think it's possibly a situation where the "actor with a single mind" model of thinking about corporate behav

I don't know that this take is wrong, per se, but I think it's possibly a situation where the "actor with a single mind" model of thinking about corporate behavior fails to be particularly useful. Scraping tends to run counter to the company's interests, too. It's relatively time-consuming - and therefore, assuming you pay your staff, expensive - compared to paying for an API key or data dump. So when engineers and data scientists do opt for it, it's really just individuals following the path of least resistance. Scraping doesn't require approval from anyone outside of their team, while paying for an API key or data dump tends to require going through a whole obnoxious procurement process, possibly coordinating management of said key with the security team, etc. The same can be said for people opting to use GPT to generate synthetic data instead of paying for data. The GPT-generated data tends to be specious and ill-suited to the task of building and testing production-grade models, and the cost of constant tweaking and re-generation of the data quickly adds up. Just buying a commercial license for an appropriate data set from the Linguistic Data Consortium might only be a fraction as expensive once you factor in all the costs, but before you can even get to that option you first need to get through a gauntlet of managers who'll happily pay $250 per developer per year to get on the Copilot hype train but don't have the lateral thinking skills to understand how a $6,000 lump sum for a data set could help their data scientists generate ROI.