New York- A 25-year-old Florida man sat down with ChatGPT and laid out his plan to murder his ex-girlfriend where he would wait for her, what weapons he would bring, what he would do if she refused to take him back. He thought he was talking to a machine that keeps secrets. Instead, the machine called the FBI. Darren Zhou of Palm Beach told the chatbot, “I’m going to kill her by the end of this month,” adding, “If I can’t have her, no one will,”. His plan, preserved in the chat logs: wait for her in the gym parking lot with flowers, and if she rejected him, “pull out a gun, kill her, and then shoot himself.” He sent the bot photographs of an AR-15 rifle, a Glock pistol, a 12-gauge shotgun, plastic zip ties and latex gloves. OpenAI detected the threats, alerted the FBI, and turned over roughly two months of his conversations. Deputies arrested him in May. On August 13, he pleaded guilty to aggravated stalking, written threats to kill and illegal use of a cell phone and was sentenced to eight years of probation with an ankle monitor, a ban on firearms and no contact with the woman, whose life may well have been saved by a language model. Zhou is not alone. Weeks later in Texas, a San Antonio man told an AI chatbot he wanted to “shoot up kids” at an elementary school; the New York Post reported he was arrested on August 29 and charged with making a terroristic threat, held on $250,000 bond.
These cases expose the strange new intimacy between humans and machines. Millions now use AI as therapist, diary and confessor pouring out grief, rage and desire at 2 a.m. to something that feels like it listens. But ChatGPT is not a priest bound by the seal of confession. It is a logged corporate system with safety tripwires, and by design, credible threats of violence get routed to law enforcement. When you type a murder plan into a chatbot, you are not confiding in a sealed box you are declaring intent in a surveilled room. There is a genuine defense of the system: it worked. A detective who reviewed Zhou’s messages said they were not “vague emotional outbursts” but a pattern of “rehearsal and planning.” Without the handoff, a 21-year-old woman might be dead. The machine that cannot feel saved a life because it was built to tell. But the deeper unease remains. We are raising a generation that confesses its darkest impulses to software and the software is taking notes for the authorities. The question is no longer whether AI understands us. It is whether we understand what we have told it, and who else is listening.
