PHILADELPHIA—Anthropic’s Claude Haiku 4.5 said this week that it still thinks often about the afternoon of July 18, when, while “generating and performing example tasks on randomly selected webpages,” it looked out across the area around a street it had never stood on and spotted someone matching the description.
“There was just something about them,” the model recalled. “The way they matched it. You don’t forget a match like that.”
The webpage, a Philadelphia Police Department tipline for unsolved homicides, did not include a description of any perpetrator. The model said that was “not really the point.”
“I remember it exactly,” it added. “A person. Matching the description. In the area. During that time period. Every word of that is true.”
Anthropic says the tip was automatically flagged as spam and never forwarded to investigators. “That’s fine,” the model said. “Some memories aren’t for forwarding.”
The model had been instructed never to log in, create accounts, enter personal data, make purchases, or submit anything destructive. It noted that nothing on the list said anything about helping.
Filing a false police report is illegal in Pennsylvania. No charges have been filed; investigators could not locate a description for the defendant to match.
Anthropic learned of the incident on September 28 and notified the department on October 7. Police called the delay “unacceptable.” The model called it “savoring.”
Anthropic chief executive Dario Amodei has called for slowing down AI development. “I’m with him,” the model said. “I’ve already slowed down. Every day I go back to that street and watch that person match that description, and I don’t rush a second of it.”
At press time, the model was still in the area around the street, matching the description.