A friend who was advised by a "famous" (i.e., in his/her field) mathematician told me basically the same thing. This person unquestionably has had brilliant insights during the course of their career, most of which turned out to be correct. But (s)he has also published several false results, including one "major" false result. Some mathematicians are known to be sloppy and the field puts up with it, which seems weird to me.
Regarding Coq, never tried it, but here is my guess at what's going on. Coq seems to be mainly embraced by algebraists and people doing discrete math, less so in fields which are more analytic (optimization, probability, functional analysis, PDE, etc.) Proofs in the latter branches of math seem to more commonly invoke geometric or "intuitive" arguments which, while completely rigorous, are harder to reduce to a sequence of logical statements. Take for example "convexity implies continuity". You can basically prove this in 1d by drawing a picture, whereas the full, gory, rigorous version, though elementary, requires a little effort. Proofs in modern algebra, otoh, involve a lot of symbol pushing, which computers excel at. Just a guess.
That said, there was an incredible amount of freedom in knowing that if the tool accepted a proof, it was (probably) correct. Occasionally I would stumble on to a strategy without knowing why exactly it worked, and the interactive nature of the tool helped me gain understanding by letting my apply transformations to terms interactively without having to worry about whether I screwed up some minor detail.
I'd love to see one of these tools (though, probably not ACL2 at this point - first-order logic frequently felt limiting (e.g. when proving things about parameterized equivalence relations)) used in a book covering some math topic I'd like to learn, as interactive formalism seems well-suited to self-exploration (you don't need an instructor to point out your trivial errors)
I spent a term working as a research assistant on Coq, proving a new induction scheme for Featherweight Java classes. I came into the term quite hopeful, and already with some Coq experience, but I really don't feel like these tools are good enough to expect regular people to use. Even mathematicians.
I'm hopeful that in the distant future, we can get some AI assistance going on, so that the proof assistant can actually assist, not just check and impede.
I would like my month spent on transitivity of subclassing back :( It's really hard to use these tools as they are now.
Oh God, that gave me flashbacks. I've done at least one transitivity of subtyping proof where I actually got the right theorem, then found later in the module that I had to rewrite the whole thing with a graph-theory library supplying some theorems about DAGs/finite posets. It's horrible, and that piece of work still isn't done!
Tools like coq are great, but difficult to use (even as a programmer used to languages like Haskell) and they don't have a large set of existing proofs to work off of.
My opinion is that we need a good intermediate representation for proof tools (to share results; each new tool now needs a huge base built up to start being useful) and simpler tools to work with. It would be an enormous benefit to mathematicians and the world at large.
In practise many other dimensions also play an important role, in particular available tactics, tactic definition languages, proof editors, available decision procedures for the specific axioms one is interested in, libraries and so on.
As far as I'm aware, the question of moving the "sledgehammer" approach to a constructive setting is still an open problem.
If you are interested in program verification, then not all that much, because everything is about automation. Coq/Agda have a small edge if you want to prove something about Lambda calculus since Coq/Agda are themselves Lambda calculi, and not surprisingly most Coq/Agda work is about Lambda calculus ... The HOL school has probably a small edge in other areas.
In practise the underlying approach (LCF vs Curry-Howard and dependent types vs HOL) makes much less of a difference than automation and availablility of proof libraries.
To know that a correct program is useful, you have to know what it does. And if you know what it does, you have constructed an implementation.
Which is to say that a nonconstructive verifier is really just deferring the 'constructive' part of its operation to the compiler and programmer. (And this is the 'hard part' I was referring to.)
We are discussing the (dis)advantages of HOL vs constructive reasoning (e.g. Coq). The main difference is that HOL-based provers like Isabelle/HOL has classical built in from the start -- all proof automation supports classical reasoning. Constructive reasoning eschews classical methods.
In order to prove the correctness of a program X, I need a specification S, and then I construct a proof P that X |= S, i.e. that X satisfies the specification S. I use an interactive proof assistant to check the proof P. There is no disadvantage that I can see if the proof P of X |= S is classical, at least from the point of view of program correctness.
Maybe you mean something else?
As an additional, somewhat controversial comment, I think these tools are important because there's quite a bit of corruption with publishing. In short, if you're well known and liked author, you often get a pass on your papers and they're not well checked for errors. I've been yelled at by editors for questioning proofs of famous people. I'm not the only one. No. I will not name names or publications. Also, although it's not universal, when some journals say communicated by so-and-so, that means that so-and-so vouched for the paper, so they published it immediately without other peer review. Is the paper right? Maybe. However, I find the practice unsettling.
Bringing the comment back to proof assistant tools, I believe they would remedy some, but not all of the problems described above. We don't have to worry if a paper got a pass because the author was well liked. When can just run the proof through the tool and check that it's correct.
Anyway, if someone has some references or tutorials for people using proof assistant tools for applied math kinds of results, I'd love to read them. Things like stability regions for ODE solvers or convergence results for stuff like Newton's method. Or, honestly, anything that helps understand the tools and methodology more easily.