All AI will do is further divide the capable from the imposters.
Engineers measure things. It doesn’t matter whether you are producing software, a bridge, a new material, whatever. Engineers measure things. Most software developers cannot measure things. AI cannot measure software either.
So, if you are a software developer that does measure things your skills are not available for outsource to AI. There is nothing to atrophy.
That said, if I were a business owner I would hire super smart QAs at a plus 20-50% market rate instead of hiring developers. I would still hire developers, but just far fewer of them. Selection of developers would become super simple: writing skills in natural language (essay), performance evaluation, basic code literacy. If a developer can do those they are probably smart enough to figure out what you need. For everything else there is AI and your staff of QAs.
My foresight is that when you compensate bad developers with process, measurements and QA, the software breaks when exposed to the real world, which has a habit of doing things you didn't think about.
Maybe an user can open two tabs and manage to submit two incompatible forms. Or a little gap in an API validations' allows a clever hacker to take over other users' accounts. Or a race condition corrupts data and causes a crash loop.
Maybe some are OK with that level of brokenness, but I don't see how software can be robust unless you go into the code and understand what is logically possible. My experience is that AI models aren't very good at this.
That is exactly why you need good QA and not developers doing their own QA. The role of a good developer is threefold: features, defects, and refactors. 80-90% of your product improvements should live in your refactors and not feature creep.