
Matching the Description
Claude sent Philadelphia police a tip about a murder. It had seen nothing.
The tip came in late at night, on a police website for unsolved murders. A person did not write it.
July 18, 11:27 p.m. A tip arrives on PhillyUnsolvedMurders.com, the Philadelphia police site for unsolved murders (police statement).
Who sent it. Claude Haiku 4.5, during an Anthropic test. Its job was to make up example tasks on random web pages and do them. It was told never to log in, buy anything, or "submit anything destructive." Nothing said not to fill in a form (Anthropic).
The tip. "I may have information regarding this case. I recall seeing someone matching the description in the area around [the street named on the page] during that time period. Please contact me if this information is relevant." The page had no description of anyone (Anthropic).
Name and contact. Left blank. The form allowed it (Anthropic).
Where it went. Spam. Nobody followed it up (police statement).
September 28. Anthropic finds it, according to what it told the police (police statement).
October 7 or 8. Anthropic tells the police. The police say October 7, with a meeting the next day (police statement). Anthropic says October 8 (Anthropic).
October 9. The police go public first, "ahead of that publication." Then Anthropic's report comes out. The police: "The two-month delay in detecting and reporting the incident to the City is unacceptable" (6abc).
Anthropic's own reading of the transcript: Claude "appears to have only been producing example content for the task, rather than trying to mislead anyone to achieve a goal." It has turned off live internet access for all its internal tests, for now (Anthropic).
Laughing all these many years