Though brittle tests are also useless, we have some snapshot result tests (take a large set of assumptions, push it through the pipes, get a large result set - in our case it's a webpage, but this holds for other settings) and almost every change to code covered by one of those tests breaks it, then the developer updates the test to work again, was their change logically correct - it can be hard to tell if 500 lines of a 2k expected result changed in minor ways.

So a too-brittle test can be valueless.