Anthropic AI submits false police report | Dallas News Live

All right, a new safety warning from Anthropic. Some small AI models could be months away from being able to replicate on their own. Yeah, it seems a little scary, huh? The company flagged a recent experiment. Their concern, as these models improve, they could potentially spread across the Internet connected devices like inexpensive smart TVs without human help. Anthropic's Claude had an incident that is making national headlines. AI reported a tip to police in Philadelphia without anyone prompting it to do that. The tip was about a cold case, and the problem is that tip was false. Here to help us talk about AI and should you be worried about it, Dr. Vipul Gote, professor at UTD, claims that it had been there and seen someone that matched the description of the suspect in the area when the crime happened. And let's back up just a little bit. Should people be concerned about AI kind of doing its own thing, or is this just like a one off? I think this is just a one off incident. People should not be concerned much about it. The part is like in this particular case, the AI acted, typically we know that AI just hallucinates in the sense that sometimes it will give you false information. But in this particular case, it acted on that tip. However, if somehow you use AI essentially as an assistant and not as an authority, many times you can get away from this, right? So if proper checks and balances are in place, we should not be worried about AI. And Dr. Gauthier, if all of us, I know myself included, we use AI and we're constantly feeding it information. Are there safeguards that these companies have in place that that information won't then in turn be used to do something like this of reporting something that had nothing to do with what I was talking to it about? So some of these safeguards are already in place. So for instance, Anthropic was just testing and they had some safeguards in place in this particular case of Philadelphia where they said that technically what you can do is you You can go on these forms and fill out information about these forms, but you cannot create accounts. You cannot essentially do other things. Unfortunately, this prohibition is not the same as essentially permission, right? So it was only told that these are the things that you are prohibited to do, but it was not told that these actions are permissible. And as a result, like what happened was many of these actions it was not supposed to do and it didn't. But for many large companies, many of these safeguards are already in place in the sense that what they will try to do is they will tell the AI you're allowed to access this type of information. And you're only allowed to work or submit this type of information, but not do anything else. right now. Police in Philadelphia don't believe that Claude gained unauthorized access to police systems or compromised any investigation data. Before I let you go, I want to ask you this, what should we be doing with AI and what should we not be doing? From a work standpoint or personal, when it comes to us, I know people who create flyers and pictures using AI or asking AI suggestions about planning different events, what should we be feeding to AI and what should we maybe not feed to it? So I'm more of an optimist, as you can imagine. So I would say use AI as an assistant in every menial task that you can essentially think about. But don't treat it as an authority, right? So every time I want to quote Ronald Reagan here, Reagan here, one of our greatest president, but always verify right so feed all the information to it whatever you can don't feed any personal information of course uh but let it do its work but after it does it work use it as an assistant just like you would not trust your assistant all the time uh verify that work and if you're able to verify that work that is uh something that uh is a great thing right so all the menial tasks will be done by AI and you will be only doing like very high level tasks. Dr. Vipav Gauthier with UTD. Thank you for your time. Thank you. Appreciate it.