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/infor

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-coding-assistants-chase-phantoms-destroy-real-user-data/ https://arstechnica.com/information-technology/2025/07/ai-co...