Fair use covers research, but creating a training database for your commercial product is distinctly different from research. They're not publishing scientific papers, along with their data, which others can verify; they are developing a commercial product for profit. Even compared to traditional R&D this is markedly different, as they aren't building a prototype - the test version will eventually become the finished product.
The way fair use works is that a judge first decides whether it fits into one of the categories - news, education, research, criticism, or comment. This does not really fit into the category of "research", because it isn't research, it's the final product in an interim stage. However, even if it were considered research, the next step in fair use is the nature, in particular whether it is commercial. AI is highly commercial.
AI should not even be classified in a fair use category, but even if it were, it should not be granted any exemption because of how commercial it is.
They use other peoples' work to profit. They should pay for it.
Facebook steals the data of individuals. They should pay for that, too. We don't exchange our data for access to their website (or for access to some 3rd party Facebook pays to put a pixel on), the website is provided free of charge, and they try and shoehorn another transaction into the fine print of the terms and conditions where the user gives up their data free of charge. It is not proportionate, and the user's data is taken without proper consideration (ie payment, in terms of the core principles of contract law).
Frankly, it is unsurprising that an entity like Facebook, which so egregiously breaks the law and abuses the rights of every human being who uses the interent, would try to abuse content creators in such a fashion. Their abuse needs to be stopped, in all forms, and they should be made to pay for all of it.