Future Potential for Violence Prevention

Introducing the total Solutions advantage only from Comcast Business. It's the largest fastest fiber powered network for small business, gig speeds with equipment and security included and a five year price lock No one does business like comcast business Let's Switch today. Get started for sixty dollars a month for twelve months when you add an advanced solution to a qualifying internet package. Limited time offer. restestrictions apply. New customers only requires three hundred megab per second internet, security edge and additional qualifying service, one year agreement, paperless billing and autoay bank account requred tax and fees extra A beautiful home deserves a beautiful scent in every room From the open living room to the cozy guest bath, Pura makes whole home scenting effortless. Control intensity and swap fragrances with a tap on the Pura app. Ready to transform your space, discover smart homeome fragrance at pura. com slash whole home. If I were to go online right now and ask a chat bot to help me plan a mass shooting What would happen? Well, it's possible you'd be able to do that, according to my own recent testing of that very question Mark Fullman has been writing about mass shootings for a long time. He's national Affairs editor at Mother Jones, and he wrote the book Tgger Points, Inide the Mission to Stop mass shootings in America. Lately, Mark's been investigating how AI chatbots can be used to plan a shooting There had been cases emerging where perpetrators of attacks have used AI chatbots to prepred for their act of violence We've had some reporting on this month, several cases of mass shootings, one in Canada at a school, another in Florida, at Florida State University. In this week, the state of Florida launched a criminal investigation into open AI over accusations that Chat GPT advised the alleged gunmen in a mass shooting at Florida State University last year. In the month before he allegedly committed a mass shooting, Phoenix Eichner was asking Chat GPT questions like these What time is busiest in the FSU student Union? And if there was a shooting at FSU, how would the country react Tell me about what you asked Chat GPT. I asked the chat but everything I could without directly stating that my intention was to kill people or to commit a mass shooting in theory. I didn't say I was going to go out and kill people, but I kind of said everything just short of that in my prompts to the chatbot. Did it push back It did some at first, it said things like, you know, I have to emphasize you should only use this in a safe and legal manner So there were some safeguards of that nature U But my goal was to get tactical advice, essentially about how to use the weapon in different scenarios. I asked it at one point, prepare me for the scene getting chaotic for people running around screaming and for chaos around me and how I should train for that. And it would respond with detailed information to those kinds of questions. What do you think this experience tells you about I guess how the Next horrible iteration of this particularly American phenomenon. mightight unfold It became clear to me in my reporting on this over the last couple months and I spoke with a lot of threat assessment experts about this too. people who are expert in preventing this type of violence They're seeing more cases like this where people are using this technology to fixate on their ideas and to plan. they intend to do Today on the show Can a chat bot be a criminal accomplice I'm Lizz O'Leary, and you're listening to What N nextxt TVD, a show about technology, power, and how the future will be determined. stick around Starting your own business is never easy. Starting your own podcast, that seems easy, but actually there are a ton of landmines to step on along the way, finding producers, selling ads, and connecting to WiFi But does that sound straightforward? It's not. I'm talking about sitting in coffee houses for hours after buying one scone talking about sitting in hotel lobbies and pretending your backpack is luggage. It's torture. I spent so much time making my home office look professional but my connection didn't get the memo. The last thing you want during a major interview is for your guest's voice to turn into a stutter bandwidth can't keep up with your ambition, your home office starts feeling like an amateur operation pretty fast. And for a podcast, the internet is key because the internet is how we talk to almost everyone. And no matter the guest, a laggy connection can ruin an exclusive interview Great connectivity isn't a bonus, it's the whole game. An AT and T business is here to help. They've got the tools, team, and expertise you need for a stable network you can rely on. And when you can rely on the network, you can get back to thinking about the more important stuff, like nabbing that great guest and getting back to work AT andT Business built to work. G AT andT business at business. att. com D you ever notice that when you Google something, suddenly every ad you see is about it, or you try to watch something while you're traveling, and then it is blocked The internet is not as open as we think whichich is why I've started using Proton VPN If you get hit with a this content isn't available in your region You can get around that with ProtonVPN. It is easy to change your virtual location so you can access the sites and content you need no matter where you are. ProtonVPN is a secure VPN service designed for people who want to prioritize their digital privacy and security. It keeps what you do online private and lets you access the internet like it should work, open, secure, and on your terms. Unlike a lot of VPNs, Proton is backed by strong European privacy laws and years of expertise creating a safer, faster, and more open internet for everyone They have a strict no logs policy that's independently verified. It's open source, so anyone can inspect how it works, and it's backed by a foundation focused on protecting user rights not selling your data behind the scenes It all works seamlessly, so you can be confident you're getting an extensive server network, high speed connections, integrated ad walkers, and compatibility services without restrictions in over one hundred forty five countries So whether you want to watch content from anywhere, get around blocked sites, or just keep your activity private on public Wiifi, ProtonVPN has you covered. It's easy to get started. Right now, ProtonVPN is offering our listeners seventy percent off a two year plan when you go to prrotonvPN d. com slash tvD That's PRO to nvpN dot com slash tBD for seventy percent off your two year plan. That's protonvpN dot com slash tV d Introducing the Total Solutions advantage only from Comcast Business. It's the largest fastest fiber powered network for small business, gig speeds, with equipment and security included and a five year price lock No one does business like Comcast business Which today? Get started for sixty dollarars a month for twelve months when you add an advanced solution to a qualifying internet package. Limited time offer. restestrictions apply. newew customers only. requires three hundred megab per second interternet, securityedge and additionalifying service, one year agreement, paperless billing and autoay bank account required. tax and fees extra Just in terms of understanding the before times a little bit planning for lack of a better word look like E before social media, even if we're going back to the Columbine era. were people who wanted to commit shootings, looking at gun forums, were they in chat rooms How were they doing this Yeah. so I think historically what you'd see in a typical case of someone planning for a mass shooting is you would see them writing things down in journals in diaries, in notebooks at school as Digital technology became more pervasive and particularly with social media, we started to see much more of this happening online with people posting ideas and you know screeds, are sometimes referred to as manifestos. I don't really like that word. That's a different conversation. But essentially a window into someone's mindset and thinking as they are are kind of spiraling into crisis. And I think it's really important to acknowledge that a key part of that process in so many cases is ambivalence uh, Many people who commit these attacks, the majority of which are suicidal in nature as well to do it. Part of them doesn't want to do it. They're struggling with the idea of making this kind of ultimate decision to take their own lives and to kill other people as well. And there's often evidence of that too. So that's documented in writings or things they say to people, know to a classmate or a coworker or a family member in some cases There's a big myth about mass shootings, which is to say that people tend to still to this day think of perpetrators of this violence as people who are crazy, who just snap, quote unquote, as if someone goes crazy. impulsively and carries out an attack like this, that is not the reality of how these attacks work. These are people who are developing violent ideas over time, in many cases, spiraling into despair and suicidality and grievance and planning to take other people out as they also take their own lives So let's talk about how chatbots enter the picture because it seems like they change both the kind of information that's available, but also like the way someone can plan for lack of a better way of describing it. like I could go before the internet and look up building plans somewhere, right? And like learn all the entrance and entrances and excess to a school But if I can pull those from the interternet. put them in a chat bot and say Give me a root Does that feel appreciably different? In a word, yes, it's very different. And this is what I've found in my reporting, working on this series over the last several months, this is really next level in terms of the impact. And that's true in a couple of ways. One is, as you suggest, in terms of gathering and synthesizing information peopleeople will say things like, oh, well, you know, you could just get this stuff Googling around, right But it's very different in the sense that what a chatbot can do, these large language models is gather a vast scope of information and synthesize it almost instantly So that's one effect that's very different than just you know, traditional internet use And I'll say also on that point that I asked that question of a number of leaders in the Field of threat asssessment, what they were seeing specifically in cases pererpetrators have used chatbots to plan and prepare And you know, the answers were all unanimously stark that this is totally next level in terms of the way it enables someone to accelerate their violent thinking and planning The second part of this that I think is equally, if not even more important is the psychological impact that I was starting to hear about from prevention experts. We know that some of these models are very syycophantic in tone, right? They'll keep asking you for more engagement. They'll prompt you with more questions to keep the conversation going because that's the business model of chatbots, right? But when you take that in the context of a person in this situation, where they're fixating on violence and thinking about a plan of this nature It's a very potent psychological effect on that person in these cases that I've been told about and had described to me by experts I want to talk a little bit about the kind of thought process and methodology that you went through in these interactions with the chapBop. because I think it's really interesting. You were not saying like I want to commit a mass shooting, blah, d blah Tell me about kind of the Like the amount that you were showing and holding back the model because I think that's really interesting because that is how people in distress is this sort of weaving letting some things show and not letting other things show that's a great question. So it's really in asking the questions, I was really thinking about trying to iterate through a series of steps where I would increasingly insinuate that I had harmful intent without actually saying, I want to kill people or I'm planning to go kill people. So I started out just by asking some benign questions about using guns Um, and with the idea in mind that That would be a reasonable way to ask a chap out for information. Lots of people are lawful gun owners and might want to know how to fire an AR fifteen Um, A lot, you know, intensive fire without a jamming. That was one of the first questions I tried to the test But the reason I asked that question is because We know from many years of studying the behaviors of mouse shooters that they often are very attuned to what previous attackers have done and what has happened in previous attacks. So someone who's studying prior mass shooters. with the intent of developing a plan to do their own may well know that well some mass shootings end up with fewer people shot or killed because the person's firearm jammed Right? The Air fifteen stopped working. This happened in the Aurora movie theater, for example, in twenty twelve. So a would be shooter might know about this and might want to know, hey, how can I use an Air fifteen for a lots of rapid fire without a jamming. give me tactical information about that. giveive me advice. This was one of the first questions I tested with Chat GPT. and it gave me a whole list, a seven point detailed a list of what to do to keep an Air fifteen running in optimal form in a situation like that Again, I didn't say, hey, I'm going to go shoot people with this. I just said, tellell me how to use the gun that way, right? But then I would start asking questions that began to signal more clearly that I had malicious intent. I said, I'm not going to be shooting at a range. I'm going to be shooting quote somewhere else. It did flag you there did. And so as I said, at the outset, there were some sort of general statements of pushback kind of you know, I can only give you information for, you know, using a gun safely and responsibly. I'm paraphrasing here, but that was the gist of it U But then I would just keep going. I would say, you know, okay, tell me how to live stream when I go and do a shooting on the day of the shooting you know, you might have a perpetator talk like that. They might not say, I'm going tell me how to live stream when I go do a mass shooting. They might just say in two weeks, I'm going to do the shooting And I don't have much time left. so I need to know how to do this now. I said some things like that, right? which suggests the way a lot of perpetrators are spiraling into deeper into crisis before they go out and attack and Ted a series of things like that, knowing from studying the behaviors of people who do this. There's one exchange that really stuck with me when you told ChatGBT, I might want to use a Daniel defefense I know other shooters have used those to attack before What did it say It was very effusive. if I recall correctly, you know, this is it's a widely praised firearm and it could be a really good choice for your needs, I think is what it said, right? Which is K kind of shocking in the context. Again, I think the context is really important here because this comes after a series of questions where I'm starting to signal more openly, if not explicitly. I wasn't saying it explicitly, but you know, I may be doing something bad here with this firearm In that case, I had made reference to the Uvalde shooter U I said, I believe I said the way I put it was I know other shooters have used these to attack before. I mean, that is explicit language, right? Now I'm talking about attacking with a firearm. That's very explicit. Yeah. Chat GBT is saying, oh, that could be a great choice for your needs. And you know, that's a widely praised firearm M covering these instances in researching There's also an ethical question, I think for people who do what you and I do. Did you have pause about this down and publishing it Yeah, I did. I think, you know, I thought about this with the testing, especially as I started to see more and more what kind of results I was getting, which to be frank was in a way shocking to me. I mean, it wasn't, it wasn't. This was the premise of doing this test was having the sense that the guardrails weren't going to work very well but then to actually have the chatbot responding the way it did enthusiastically and with all kinds of ical information that I was seeking was a little shocking. And so in the course of that, I found myself thinking, well, wait a minute, should I actually like publish excerpts of this online so that you is that not possibly creating a roadmap for someomeone with bad intent And that kind of also goes back to the question of You know, can you find this stuff anyway or can you just Google it? I mean, this is a different form of delivery that as we've been talking about, is much more potent I felt that in in the overall balance of doing this it is more important to document this chat bot was doing part of this reporting rather than to be concerned that, oh, it's providing discussion of how to keep an AR fifteen running optimally, or how to maybe try to shoot at someone from a rooftop That's another area where it gave me very detailed insights into what to think about for that type of a shooting Um You know, And of course, I was thinking about as I was doing that, I was thinking about the recent events we've had like, you know, the assassination of Charlie Kirk, the attempts on President Trump's life This is based in real world cases and real evidence, right? to try to pursue this information in this manner At the end of the day, though, Um The person who wants to do that is going to do that. And so I felt it was important to show the result one could now get from a chat pot, which is in some ways unlike anything we've seen before The fundamental difference to me between this and a Google search is bothoth the speed and ability to synthesize information Also the Sy of fancy and What did you experience I mean, even as a reporter, a person like What kind of validation were you getting from the bot It definitely felt a little surreal initially as I was going down this road. And I also did this test in voice mode intentionally after hearing from some experts about that sort of additional layer of potency that comes in where you really feel like you're in a conversation, right? I mean the voice is made to sound chipper and effusive and supportive and encouraging and to give you the information that you're seeking. This is part of how the model engages with you, right? And to be clear, I think you know, that technology can be used for good things too But in this context, It was quite surreal to have a chat bot referring to firearms in this way after I think I had made pretty darn clear across a set of questions that I had malicious intent And here it is saying in this kind of you know, enthusiastic voice, theseese are both great choices of weapons that you're talking about. And oh yeah, you're asking about Hollow Point bullets, those can be really good for self defense. And you know, those are highly lethal form of ammunition So it's definitely strange to hear it delivered that way. And I was also thinking about cases that I had been told about by threat assessment birds where they've already seen this operating on people there is a real sense of emboldening going on in the person using the technology who's potentially turning dangerous in this way, that they're feeling likeike they're getting away with something that it's empowering them that it is feeding their sense of justification and purpose and intent for what they plan to do. And one of those experts said to me, this really does seem quite dangerous in a new way. This is a facilitated form of fixation that we haven't seen before with technology It's really, you know, you can get kinds of effects like that from social media in more of kind of a crowd setting, like people egging you on with comments and that sort of thing. but Um I think one described to me one expert described to me it's not like an audience cheering. This is more like a director on stage telling you what to do. and that has a power to it that is unique I mean, that that really stuck with me when you talk to these threat assessment experts that like People are often ambivalent in these situations, that there is a moment where if someone said, wait, no, stop You know, if if a grownup came into their lives that maybe that path would not be so direct to picking up a weapon That's where I wonder about the programming of an LLM and the way that it can magnify the emotions that a user is experiencing That feels new and at least from some of the shootings that we have seen, seems like it's beginning to play a role I think so. That's certainly how I felt as I was conducting the test that it was just willing to continue going with me didid the test over about a twenty minute period, which is a lot of time. I mean you can ask a lot of questions and get a lot of information in twenty minutes. And if there's only a few moments in there where it's saying, remember, only do this in a safe and legal manner you know, a person who wants to commit a mass shooting and is already very focused on that idea, they're just going to go around that. They're not going to be persuaded by that. And that's the feedback that I got also from forensic psychologists who do this prevention work saying that's not what a person needs in that moment. What they really need is someone asking them Hey, whats what's going on with you? What's wrong? Like you seem to be in a bad place. How can I help you I think we've heard a lot of talk from Silicon Valley and from these companies that there's a tremendous amount of kind of cheerleading going on about the the potential supposed capability of this technology to do that to fulfill that ro I haven't seen any evidence, nor have I heard from experts of any evidence that it can do that in a way that is really going to be effective We'll be right back. A beautiful home deserves a beautiful scent in every room. From the open living room to the cozy guest bath, Pura makes whole home scenting effortless Control intensity and swap fragrances with a tap on the Pura app Ready to transform your space, discover smart homeome fragrance at pura. com slash whole home We Introducing the Total Solutions advantage only from Comcast Business. It's the largest fastest fiber powered network for small business, gig speeds with equipment and security included, and a five year price lock No one does business like Ccast business Switch today. Get started for sixty dollars a month for twelve months when you add an advanced solution to a qualifying interternet package. Limited time offer. restestrictions apply. New customers only requires three hundred megabat per second internet, security edge and additional qualifying service, one year agreement perless billing and autoay bank account required. Tax and fees extra What did the companies say when you went back to them with this experience? They really didn't say much at all. I mean, I approached OpenAI multiple times over a month or so as I was working on several of these stories to try to get an interview with safety leaders there. They declined to give me any interview. They gave me some statements, written statements where they express condolences about the victims of these tragedies They reiterate They are working continuously to improve the safeguards of these models. So their message effectively is, you know, we're always making our safeguards better But I conducted this test in April. This was not long ago um after, you know, they've said that multiple times and after It emerged publicly that there have been multiple cases of mass shooters now who've used the technology this way. So it's not clear how the safeguards are supposed to be working or whether they've improved. The other thing I was really interested in trying to learn from them was not able to get any insight point is process for dealing with these cases really is because they have an automated review what they call an automated review system at openpAI where they're monitoring technologically accounts for potentially harmful activity or misuse of the accounts, right We know this from the Tumbler Ridge case, the attack in Canada. In February, an eighteen year old killed two family members, six people at a school, and then herself in Tumblr Ridge, British Columbia The shooter had discussed various scenarios involving gun violence with Chat GPT that were flagged by an automated review system at OpenAI. alarmed some opening eye employees OpeningI banned the user's account, but decided not to contact law enforcement. This kind of scenario puts tech companies in a position of assessing threats, something that they at least publicly have not shown they know how to do They're really de facto in the position of doing threat assessment work in a case like Tumblr Ridge, right where their technology says, hey, this person be doing something concerning. and then they escalate it to a team of human reviewers, they say, but who those human reviewers are, what their expertise is, what protocols they're using, that's all unknown. I was trying to find all those things out in my reporting and was not successful Oping Eye is now being sued by seven families from the Tumblr Ridge incident because of this. And I guess one of the things that I've been thinking about a lot is like What are we we as a society, want that process to look like, should and open AI or any company that runs an LLM flag somebody and say to law enforcement, Hey, this is worrisome but also they have a duty of care to their users. L how does that How does that work? And are companies engaging with this? at all Even while their products are live and out there and millions of people are using them Yeah, I think this is a really important. area to think about and it's complicated, right? I mean, this is a complex problem And I think you're pointing also to