It's not a foregone conclusion. I've had great success writing BDD-style literate tests (not using a framework, just comments that follow a particular grammar). They've allowed verbally negotiating the approximate expected specification, and dealing with the realities of what comes up once you start writing code. But you end up with a spec, expressed as a number of BDD scenarios.

Obviously this does't work for all software. But it suits systems with end-to-end behaviour and sequences of actions.

So it's possible to leave it to developers, but expect a high quality spec along with the code.