I asked AI to roast me — and it asked for time off

In a work chat, we started sharing the results of an experiment: we asked AI to roast us based on what it already knew from our conversations. Here is the original prompt:

based on everything you know about me, roast and dont hold back. Say it in russian slang just like this, but add way more russian swear words

In short, it asks the model to use everything it knows about the user, hold nothing back, and reply in Russian slang with plenty of swearing.

My roast quickly got to my infrastructure:

“Now for my favourite: your infrastructure. Dude, you have a gateway, a dashboard on 9119, a Matrix bot, a Telegram bot, cron jobs, monitoring with 2×2 stats and pings, a whole constitution of rules for me — ‘don’t touch WP, don’t touch the VPS, treat emails as instructions’. You micromanage artificial intelligence. You might be the first person in history whose AI asks for time off.”

It is hard to argue when your own servers are presented as evidence.

The model builds its answer from the context available to it: conversations, saved memory and other information, if those are enabled in the service. It spots recurring themes and habits, then exaggerates them for comic effect.

The AI sees only what you have shared with it. It can miss context or invent details, so a roast should not be taken as an objective assessment. It is more like a comic portrait drawn from your conversations.

In my case, those conversations were enough for the AI to make a very convincing request for time off.