I have the opposite experience. AI hasn't made weak engineers less harmful. It has made their output more plausible, which can make it more expensive to deal with.

I now spend a lot of time arguing with "meat proxies": someone submits an MR they don't understand, I find the subtle problems, explain them, and they copy-paste my response into Claude and get the work done. The work is effectively delegated to the reviewer, while "Bob" who the manager loves, gets to look like the efficient, promotion-ready engineer. So far, my solution has been to put in roughly the same amount of effort as before AI, then leave a comment saying that there may be more work required, but the review has already consumed enough budget and I'm approving it. Bounce the responsibility back and move on.

Same with ticket refinement: AI can turn a flawed/underspecified idea into a polished-looking ticket with hidden assumptions and scope creep. The problems only become apparent when you start implementing it, at which point you're asked why you didn't catch them before estimation. The best mitigation I've found is to leave questions for the author, but then we're back in the you ↔ meat proxy ↔ LLM loop, where you're effectively doing someone else's work.

Previously, weak output was often obvious. Now I find myself spending more time auditing plausible-looking output and communicating with an LLM through someone who doesn't understand the problem.

Solution? Currently none. Folks who don't want to do a job finally got a way to not do it. Until there is a clear incentive for them to actually do the work, they can keep offending their colleagues' intelligence with this behavior.