STRING is a curated database (DB) as you state correctly. Now, for the following we can totally abstract from the nature of items in the DB. A DB allows for (efficient) storage and retrieval of information (e.g. entities and relations) from the modeled application domain ("real world"), not for discovery of completely new things. So people are using databases to get access to information that is already there. Of course, new hypotheses or even information can be derived from such efforts, but such are new (human or algorithmic) interpretations of already known facts.
That is the same reason for using f.e. a library or shop catalogue, and for such a DB it would be sufficient to contain all books that are accessible in the library or the shop, and not all books or items that could at sometime in the future exist or that might be written in the future.
So as there are methods to detect protein interactions, these methods are different from storing the results from such experiments, in the same way that writing a book or creating a new product is different from storing information about them in an online-shop or library catalogue. So, criticizing STRING for not containing novel interactions is like criticizing the librarian for not being an author.
In addition, in the PPI domain, using methods of detection may easily create false positives that indicate interactions that do not exist or are un-biological, compare this to a pretty badly written book or a vapor-ware item in a shop. The curation process might get rid of those.
Criticizing STRING for not containing novel interactions is like criticizing the librarian for not being an author.