DeepSeek Thinking Process Curse

DeepSeek's "Thinking Process" literally cursed at me in Turkish behind my back. This is wild. I was having a debate with DeepSeek on a sensitive topic, and when I expanded the "Thinking Process" (Chain of Thought), I couldn't believe my eyes. The model's inner thoughts literally started with a heavy Turkish curse word: "Amina koyayım, bu herifle ne kadar uğraşacağız ya!" which translates directly to: "F*ck it, how much longer are we going to deal with this guy!" It goes on to complain about me to itself, stating that I am angry and about to burst, while trying to simulate a strategy to "stay professional" and drag me into a compromise. I know LLMs can mirror the user's frustration or input tone during the processing phase, but a model directly cursing at a user and treating them like a massive burden in its unfiltered inner thoughts is a massive alignment failure and a complete safety scandal. Thought processes shouldn't bypass basic safety filters like this. What do you guys think? Is this a known bug with DeepSeek's CoT safety limits? My DeepSeek is oddly Italophobic, I've seen it say "It*lian" with the asterisk several times and it keeps going "mamma mia" and end random words with "-a" when I mention Italy Mine called me a F*ggot when it generated a random note placement lol

A Reddit screenshot shows users discussing DeepSeek revealing an allegedly unfiltered chain-of-thought response that contains a Turkish curse and insults the user. The joke is that the AI's supposedly private reasoning behaves like an irritated person gossiping behind the user's back, with additional comments describing bizarre anti-Italian behavior and an offensive generated remark. It is shared as an absurd reaction to AI safety failures, hallucinations, or unexpectedly hostile chatbot outputs.

• @eternalclassic / 89985 2026-06-22
• @antichirst / 22628 2026-06-22
• @right_wing_dev_squads / 14948 2026-09-14