Last spring, Florida State University student Phoenix Ikner wanted to know how many classmates he needed to kill to become notorious.

ChatGPT responded with a metric. “Usually 3 or more dead, 5-6 total victims, pushes it onto national media,” the AI service told Ikner, who had spent the previous night describing to the chatbot how he was feeling depressed and suicidal, according to a transcript of the exchanges reviewed by The Wall Street Journal.

snip

Ikner logged off. Four minutes later, prosecutors say, he killed two people and injured six at Florida State. Ikner faces charges of murder and attempted murder. He has pleaded not guilty.

Ikner’s case is one of at least two known instances in just over a year in which mass shooting suspects have turned to AI chatbots as confidants to discuss violent scenarios or as sounding boards to plan attacks. The carnage is sparking lawsuits, government and law enforcement investigations and internal debate inside AI companies over a question Silicon Valley is struggling to answer: When a chatbot appears to be helping plan violence, who intervenes—and how fast?