A lot of systems would benefit from formal (or more formal) methods. I worked in aviation software, formal methods would have saved us so much trouble, especially since the code size is often relatively small (10k-200k LOC) and because we're dealing with other systems the IO spec was reasonably complete. Sure there'd be issues, but feature creep was not one of them (except on one project). Feature creep makes formalisms difficult, but only when you're looking at whole program formalisms. So certainly safety critical systems benefit.

What else might? How about medical systems (see Therac-25 for faulty UI leading to deaths). Network code, especially fundamentals like routers, switches and core services like DNS that so much depends on. Cryptographic and other security systems. Compilers, obviously, similar to network code they're too ubiquitous to be left to error prone ad hoc approaches. Anything handling money. Anything handling privacy (ensuring HIPAA and whatever European privacy rights). Software handling production automation or managing warehouses/inventory, failures there mean lost production time, wasted inventory, lots of money and productivity gains to be had by having software that works.

If you accept formal methods on the small scale (essentially the promise of functional programming a la Haskell and the ML family), you can be confident in composing large systems out of smaller, formally verified systems.

That is plain FUD right there.