* regularly generating AI-slop code changes that nearly work, and then trying to hand over to someone else for "completion". this is often a pretty terrible experience, and more work than just having a single person take ownershiop of the _outcome_ - i.e. understand the need, write the code, test it, deliver, make sure it works... generally the coding was never the hard part, and its harder to start with broken code than just from scratch.
* sporadically sharing AI analysis of bugs/problems/crashes/troubleshooting type stuff. again, often nearly correct... but sometimes more work finding the incorrect aspect than starting from scratch.
* lots of AI flavored text and presentations. you know the kind, text where every bullet point is a different emoticon, and AI-generated graphics instead of random, but real photos stolen from the internet.
* lots of AI for code review, either of changes (PRs), or just wholesale analysis of code bases to find problems and automatically create tickets for people to "fix". often , the analysis is not "wrong", and potentially very good... the problem is that it finds "problems" in utilities that don't run in production, historical code just stored for archival purposes, anything and everything... it has no concept of "this is live internet-facing code" vs "this is an ad-hoc utility that does not need to be bulletproof". the volume of work created by AI in this space is truly impressive.
* lots of AI "recipes" for things that could have/should have been check lists / processes / shell scripts / python utilities / actual software. instead of something actually fast, deterministic, and re-usable, the "recipe" now mostly works, but takes several minutes and does slightly different things each time that cannot really be predicted or reasoned about.