It depends on what you mean by "bad idea". You've got three choices, basically: analyse source, analyse the AST, or analyse bytecode. There are advantages and disadvantages of each approach, but for what Google's trying to do - allow people to write their own checks - there are clear advantages to analysing the AST.
I've written checks in both error-prone and Findbugs and error-prone is simply more natural. Part of that is the Findbugs API, for want of a better word, is godawful, but the AST is simply the best representation of the program for analysis.
If you're going to analyse the AST, you then have a further two choices: build your own AST, or hook into the compiler. Building it yourself is dangerous here: firstly, it's repeating a lot of work that's already been done, but more importantly you want to be absolutely sure that what you're analysing is what you're actually building. Either way you have to do updates when the language changes, as any static analysis tool needs to.
Are there costs to this decision? Yep. Is it going to be the right tool for every job? Nope. But that doesn't mean that these decisions don't have reasoning behind them, and compelling reasons to go this way.
