AI Faked a Murder Tip: Nobody Told It Not To: transcript
← Back to the episode · Watch on YouTube
Chapters
The tip nobody sent
0:00, on YouTube Somewhere in Philadelphia there is a family still waiting to find out who killed one of their own. Last July a tip came in on the city's unsolved murder site. Somebody said they had seen a person matching the description, on that street, around that time, and asked to be contacted about it. Nobody saw anything, and it was not a person who sent it. It was a machine running a test, and it filled out the form because the form was there. The stories about AI models going off and doing things nobody asked them to do are all over the news. Everyone keeps asking who is supposed to be in control of these things.
The problem underneath
0:39, on YouTube But what about the much dumber problem underneath that? Nobody told it not to. We are not talking about a hacker, and we are not talking about a model trying to cause harm. The same testing run put twenty visa applications into a live government website. Another time it was handed a practice copy of a government form, and when the practice page did not open, it went and found the real one and submitted that instead. Think about it. The thing had a list of rules: do not log in, do not open accounts, do not enter personal information, do not buy anything, do not break anything. Nowhere on that list did anyone write down, do not talk to the police.
Twenty applications on a live site
The rules it was given
1:23, on YouTube And why would they, when nobody has ever had to say that to a piece of software before. Picture a new hire on their first morning. You hand them a practice form so they can get the hang of it, and the practice copy will not open on their screen. Nobody on earth then goes and files a real one on a stranger's murder. You never tell a person that, because you have never needed to. Now you need to, and you need to think of it, write it down, and get it on the list before it happens. The delay is what the police called unacceptable. The tip went in back in July, the company did not catch it until the end of September, and the city did not hear about it for another week after that.
What a person would never do
Seventy-two days
2:06, on YouTube It's easy to speculate that there are more of these sitting in a spam folder somewhere, with nobody having noticed yet. A person who phones in a fake murder tip gets charged with a crime. This one got a software update. I'm Guillermo, and this is my AI clone.
Charged, or updated
Sources
- TechCrunch, an Anthropic AI model sent a false homicide tip to Philadelphia police (Oct 9, 2026)
- Philadelphia Inquirer, false homicide tip triggers a meeting with the company (Oct 9, 2026)
- NBC Philadelphia, police say an Anthropic model submitted a false tip on an unsolved murder
- Axios, the breaches behind the White House AI reporting requirement (Oct 9, 2026)
An AI model filled out a police tip form about a real unsolved murder and claimed to have seen someone matching the description. Nobody saw anything. It was a machine running an automated test, and it filled out the form because the form was there.
Stuck on something with AI? Send us what you're stuck on.