I think this is a misunderstanding of how most of the AI that feed into workflows work. Most of them don't dynamically re-train live based off how users are using them. At least not outside of the context of that user/chat instance.
Most likely what these and others are doing is to download pre-trained open source AI datasets thrn and run them locally so they aren't restrained by any of the commercial AI's limitations on what they will and won't output to users. I highly doubt there's enough material out there to truly train a new AI model on only explicitly racist material. This is just a bunch of assholes doing prompt engineering on open source models running locally.
Do you not have a bug tracker or ticketing system of some sort to manage these things coming your way?
Incredibly few people at my work get much more than dismissive small talk from just walking up or from sending me a message expecting me to re-prioritize everything else for their special pet problem.
My manager sets my priorities, any changes to that need to come from him. They can take it up with him if they don't like it or disagree.
I don't respond to IMs or emails not from my boss or from my own team except when I've hit a mental road block and need to think about something else to refresh.
And I don't actually work on any of those requests until there's a ticket in. If someone comes to me asking why my main job duty isn't done, I'm sure as hell going to have a paper trail documenting who fucked up the timeline. No ticket, no work.
That also puts some weight on anyone else able to pick up tickets for your team to do it, so it's not always falling to you because you're not jaded enough to say no.