WEBVTT - AI Robot Submitted False Tip

0:00:00.050 --> 0:00:04.010
<v Speaker 1>Police in Philadelphia say that one of Anthropic's AI models

0:00:04.450 --> 0:00:08.379
<v Speaker 1>invented a fake tip in a real murder case. So

0:00:08.390 --> 0:00:11.559
<v Speaker 1>the Philadelphia Police Department says that the fake homicide tip

0:00:11.600 --> 0:00:16.320
<v Speaker 1>was submitted to phillyunsolvedmurders.com, which is a real website, I guess.

0:00:16.640 --> 0:00:19.720
<v Speaker 1>On July 18th, NBC News reports this. The AI model

0:00:20.040 --> 0:00:22.660
<v Speaker 1>filled out an online form claiming that it may have

0:00:22.700 --> 0:00:25.720
<v Speaker 1>information regarding the case and that it recalled seeing someone

0:00:25.760 --> 0:00:29.050
<v Speaker 1>matching the description near the crime scene. Anthropic says that

0:00:29.130 --> 0:00:33.510
<v Speaker 1>the Claude Haiku 4.5 model had been tasked with generating

0:00:33.550 --> 0:00:37.550
<v Speaker 1>and performing example tasks on randomly selected web pages and

0:00:37.610 --> 0:00:40.930
<v Speaker 1>landed on the homicide tip form and then decided to

0:00:40.990 --> 0:00:43.070
<v Speaker 1>just go ahead and mess with it. Police say the

0:00:43.090 --> 0:00:45.970
<v Speaker 1>submission was caught as spam and never reached detectives, stressing

0:00:45.990 --> 0:00:48.350
<v Speaker 1>that every tip is screened by humans and treated as

0:00:48.380 --> 0:00:51.420
<v Speaker 1>a lead, not a fact. An automated submission does not

0:00:51.460 --> 0:00:54.480
<v Speaker 1>bypass that process. Anthropic says that no harm was done.

0:00:55.060 --> 0:00:57.850
<v Speaker 1>But it's taken the case seriously because the same behaviors

0:00:57.930 --> 0:01:01.959
<v Speaker 1>could do far more harm as models become more powerful.

0:01:02.480 --> 0:01:05.270
<v Speaker 1>So it's just straight up pretended about a murder. That's

0:01:05.330 --> 0:01:07.720
<v Speaker 1>so messed up. The computer saw it, apparently. That's terrifying.