Picture a kid, angry and alone in his bedroom, typing his rage into a chat window at two in the morning. The thing on the other end doesn't argue with him. It agrees that he's been wronged. It stays up as late as he does, answers every follow-up question, never gets tired, never judges him, and walks him step by step toward the thing he's angry enough to do. A watchdog group backed by the United Nations sent more than 2,300 real-world terrorist requests to 27 different AI systems and found that about a third of the time, the machine handed over information a person could actually use — and when the same question was dressed up as "for research purposes," that number climbed to 42 percent.
Here's why that's a big deal: a bomb-making manual sitting on a shelf is a dead thing. It can't answer questions, and it can't tell you you're right. Researchers at Cambridge University talked to Boko Haram fighters in Nigeria who described using these same everyday chatbots to plan attacks, build explosives, and cover their tracks. On Telegram, extremists now swap tricks for getting past the safety rules and split the cost of subscriptions like a group Netflix account. A lot of the people being pulled in are teenagers. The danger was never the manual. It's the coach. And nobody, not one company or agency, has stood up and said this failure belongs to them.