This tendency must get reenforced through RL in the training phase. It's very high profile when an LLM deletes the wrong thing, eg https://arstechnica.com/information-technology/2025/07/ai-co...