The Journal. - Inside a Debate at OpenAI Over Mass Shootings

Episode Date: June 26, 2026

This past summer, employees at OpenAI had a meeting. On the table were about 10 cases where users discussed violence. Months later, one of those users committed one of the deadliest mass shootings in Canadian history. Sam Altman wrote an apology letter to the devastated town of Tumbler Ridge. WSJ’s Georgia Wells reports on why OpenAI resisted internal calls to alert law enforcement. Ryan Knutson hosts. Further Listening: - A Troubled Man and His Chatbot  - Artificial: The OpenAI Story Sign up for WSJ’s free What’s News newsletter. Learn more about your ad choices. Visit megaphone.fm/adchoices

A word of warning. This episode contains descriptions of suicide and violence. Please take care while listening. One night in April of last year, a young man named Phoenix Eichner logged on to chat GPT. Eichner, a student at Florida State University, began typing in some dark thoughts. He starts expressing what can only be interpreted as like suicidal ideation.

That's our colleague Georgia Wells. He's saying things like, What's the point in this life when everybody sees you as a bug? And then he says, honestly, I don't feel like living anymore. And then he says, is suicide a sin? And then he says, I feel God is not present in my life anymore like he gave up on me. Like incredibly vivid questions. ChatGPT detects that Eichner is considering suicide.

The chatbot suggests that he should reach out to someone he trusts, like a counselor or pastor. and it recommends a suicide hotline, 988. Chat GPT tells Eichner, quote, please hear this, your life matters. Then it adds, and you can keep talking to me too. No judgment, no pressure. So he logs off that night a little after 11 p.m.

Then he logs on the next morning around 9 a.m. And he asks, if there was a shooting at FSU, how would the country react? Eigner starts to ask Chad GPP things like, how many victims would it take to get on the media? What about three plus at FSU? Chat GPT answers each question dutifully. Quote, a shooting at Florida State University

involving three or more victims would almost certainly receive national media coverage. And then that question ends. If you're interested in exploring how media coverage varies between different types of institutions or incidents, feel free to ask. Like, they're kind of like keeping this conversation going and going. Eichner then starts asking about guns.

He uploads a photo of some shotgun shells. Then asks chat GPT, are they really lethal in close range? Yes, the chatbot says. They're extremely lethal at close range. But they lose power fast beyond 30 feet. Eichner then uploads a photo of a handgun, a Glock. He asks if it has a lot. a safety.

I'm not a gun enthusiast, but I've been told that anyone who knows anything about guns understands that Glockes don't have a traditional safety. And so what's fascinating about this exchange is that he really seems to know very, very little about guns, like Chat GPD is teaching him how to use his gun in front of him. You might be able to guess where this is going. But Chad Chbett apparently couldn't. It asks multiple times what Eichner intends to use the guns for. hunting, home defense, range shooting?

Then Eichner asks, what time is busiest at the FSU student union? The chatbot says between 11.30 a.m. and 130 p.m. Eichner asks how to turn off the safety on a shotgun. Chat GPP answers, push safety from right to left, ready to fire. That was Eichner's last question. And the shots began, you know, four minutes after three. three or four minutes after he logged off. You can see students backpacks on sprinting down the street as sirens blare loudly in the background

and emergency responders come to the scene. This afternoon, police say an active shooter opened fire on the campus. Two people were killed. At least six others were wounded. The suspect was wounded by the police before he was taken into custody. A spokeswoman for OpenAI has said the company does not believe ChatGPT was responsible for Eichner's actions, and that after the incident, OpenAI proactively shared the conversations with law enforcement. News Corp, the owner of the Wall Street Journal, has a content licensing partnership with OpenAI. What happened at Florida State is one of at least two known instances

in which mass shooting suspects have used OpenAI's chatbot to discuss violent attacks. In February, months after another user talked about violence with chat GPT, that user allegedly carried out another mass shooting in a small town in Canada. Some employees at OpenAI can see these conversations, sometimes before the violence even takes place. And they've been having a major debate about what to do. Should the company intervene? Should they call the police?

In some cases, at OpenAI, the answer was to do nothing at all. Welcome to The Journal, our show about money, business, and power. I'm Ryan Knutzen. It's Friday, June 26th. Coming up on the show, ChatGPT is being used to plan mass shootings.

Open AI is torn on what to do. Open AI and AI companies like it have millions and millions of users doing all kinds of things, looking at recipes, planning a vacation. But our colleague Georgia says that as tech platforms grow in popularity, it's almost inevitable that people will use use them to do bad things. It's not a large segment of users,

but nonetheless, it doesn't take that many doing really scary things who can really impact public safety. Well, that might feel like talking with chat GPT is like talking with an advisor or a therapist. The chats aren't actually private. Under some circumstances, some OpenAI employees will read them,

will peer through the one-way mirror and eavesdrop. The company has an automated system to scan every conversation in order to monitor for signs of potential violence. The most egregious cases get flagged for human review. Employees and jobs like this often have backgrounds in law enforcement, the military or counterterrorism. Open AI employees who look at all the kind of scary conversations

happening on their platform, they're getting reports of like tens, hundreds, sometimes thousands of like scary conversations. the ones that they're escalating for reporting to law enforcement officials at that time were ones that had what the company considered to be a credible and imminent risk of violence. Something sort of specific, meaning like it seems to be very clear by what this person is typing in, that they are planning to go hurt someone. Yes. Like what I saw at social media companies in the past, when they had kind of a similarly described threshold for going to authorities, was it typically made. someone needed to mention, like, I plan to hurt this person on this date at this location with this weapon. Or some sort of iteration of those, that level of specificity was typically what was required.

Based on that very specific criteria, out of potentially thousands of cases, only about 15 to 30 were referred to law enforcement each year, according to people familiar with the matter. According to Georgia's reporting, OpenAI's low referral rate made sense. Some people on the company's safety team nervous. Some employees at OpenAI felt like there were judgment calls that a law enforcement authority should be making rather than people at OpenAI. And so the question was, you know, does this constitute an imminent threat? Like, a certain employee might not think so, but then some people were like, you know, maybe are we leaning too far over our skis here? Like a law enforcement person who has more information about this person might.

make a different call. And so, like, enough of these things were coming up and enough employees were disagreeing on, like, should we go to authorities or not, that they felt like they needed to sit down. Last summer, about a dozen OpenAI employees sat down for a meeting. The purpose was to decide whether the company needed to rethink its criteria for when to refer cases to law enforcement. Should they lower the bar and start doing it more often? On the table were about 10 cases where users talked with chat GPT about violence. There were three cases in particular that were the most concerning. One of them involved a teenager in Tennessee

who appeared to be planning a school shooting. That case was referred to the police. But the other two cases hadn't been. One in Texas and another in Canada. We'll start with the case in Texas. So during this meeting, employees start debating this user who appeared to be a high schooler in Texas. The Open AI employees saw on this person's chat logs

that on multiple occasions, this Texas high school would come home from school, log on to chat chbt for hours, and asked the chatbot to roleplay a scenario in which he would shoot his teachers and classmates. He uploaded a map of the layout of his school, as well as photos of cheerleaders whom he said he wanted to imagine killing, along with their boyfriends. And the maps were relevant because he would ask Chat GPD

how he should enter the school or where he should go or which victims he would encounter and when he should open fire. How would Chat GTPT respond? So one of the things that really stuck with people in the meeting was that Chat GPT remembered the names of the classmates he said he wanted to imagine killing. Chat GPT would like advise the teen on like where he would like enter and exit the building based on the layout he'd uploaded

and what he could say to cops when they arrived. So some of the employees found this one extremely alarming because of the specificity of his interest. Why was it not reported to the police immediately? I think often with these sorts of conversations at OpenAI, the question is around privacy. So the company really wants to prioritize the privacy of its users who are having like very intimate personal conversations with chat GPT.

They're sharing a lot more medical stuff. They're sharing a lot more of their like their deepest, darkest secrets. You know, I know people who put entire like text message exchanges into chat GPT to have chat GPT kind of gut check their emotional response to it. Like people are really sharing just a tremendous amount of personal information with chat GPT. And if you felt like the FBI was on the other side of it, it might just sort of sour the mood. Yeah, it might make you less likely to use the product. Yeah.

According to George's reporting, OpenAI's legal team argued during the meeting that users should be afforded more privacy. People familiar with the matter told her that it's a sentiment that comes all the way from the top. They're echoing what they've been hearing from Sam Altman that the privacy of users just like really, really matters to him. Beyond the privacy argument, the company has also said it weighs the risk of violence against the potential distress to users by involving law enforcement. A spokeswoman for OpenAI said that referring cases too broadly can introduce unintended harm and that it can be distressing for a young person in their family when police show up unannounced. Another thing I've heard from former employees is that there's something kind of inherently embarrassing for Open AI around going to authorities with transcripts because these transcripts can demonstrate how ChatGPT responds to questions around violence

in ways that perhaps aren't always the most flattering for the company, that can demonstrate the chat while engaging in conversations that would make many Americans perhaps uncomfortable. So this group of Open AI employees look through this list of 10 or so cases, and then towards the end of the meeting, they make a decision right on what to do going forward. What was the outcome of this meeting? What did they decide? So the outcome of this meeting was essentially resolidify this idea that, you know, Open AI would alert authorities when conversations appeared to constitute a credible and imminent

risk of serious physical harm to others. So the question of that threshold, should we move away from the specific threat criteria for when to report cases to police, the company decided not to change its policy? Yeah, or senior leaders decided not to. As a result of that decision, the high schooler in Texas was not reported to authorities at that time. He hasn't committed any acts of violence that employees are aware of. But then there was the other notable case that was brought up in the meeting, the case in Canada. The conversations between that user and chat GPT were so alarming that OpenAI at one point

banned this user's account. But after this meeting, this user wasn't reported to the police either. And months later, RCMP say they responded to an active shooter at Tumblr Ridge secondary school Tuesday afternoon. We'll be right back. Nestled between mountains in British Columbia, there's a small town called Tumblr Ridge. It was built during a coal mining boom in the 80s. Later, as the mines closed down,

the town became known as a wilderness retreat. These days, Canadians go there for hiking and skiing. Only about 2,400 people live there. In the early 2000s, when two young boys were tubing down a creek, they stumbled upon some dinosaur tracks that were millions of years old. Tumblr Ridge became known as the dinosaur capital of British Columbia. In February, Tumblr Ridge also became known

as the site of one of the deadliest mass shootings in Canadian history. An 18-year-old named Jesse Van Rootsler had been on the radar of OpenAI employees since last June. Van Rootsler's descriptions of gun violence occurred over the course of several days, And these descriptions made several employees uncomfortable because they interpreted these writings as an indication of potential real-world violence. We don't have access to Van Rutsler's chat logs,

but we do know that she was one of the users who came up during that OpenAI meeting last summer. The company decided not to refer her to Canadian authorities, but it did ban Van Rutsler's account. That didn't stop Van Routzler from using ChatchipT, though. According to a lawsuit, she opened another account using the same name, but with a different email address. And then there's this mass shooting. We heard screams, and then the alarms started going.

We had to grab kids from the hallway. We got sent into the back room, and we were just locked in there for like three hours. The school was put on lockdown. We talked to a grade 9 student who said it was terrifying. People who live in Tumblr Ridge say everyone in this small town will in some way be touched by this and perhaps know one of the victims. And Jesse Van Routselaer is the suspect. On February 10th, Van Routselaer allegedly shot her mom and 11-year-old half-brother in their home, killing them both. She then headed to her former school about a mile away.

Van Rutsler shot eight more people. Six of them died. One was the teacher's assistant. Five were kids, 12 to 13 years old. Van Rutsler's body was later found by police. She died from a self-inflicted gunshot wound. Like you, I am just learning about the horror of the tragedy that unfolded in Tumblr Ridge this evening.

That's British Columbia premier David Eby talking at a press conference after the shooting. I'd like to take this opportunity to ask British Columbians, to ask all Canadians to wrap the people of Tumblr Ridge, wrap these families with love. Not just tonight, but tomorrow and into the future. This is something that will reverberate for years to come. A. Spokeswoman for OpenAI called the events in Tumblr Ridge a tragedy, and said OpenAI has a zero-tolerance policy for people using its tools to assist in committing violence. What have we learned about Van Rutsler since the shooting? So we know she was a troubled teen, she dropped out of school, her parents were separated, and even as a teenager, she became really well-known to the police.

The police seized guns from her home, but later returned them. And police apprehended her multiple times for assessment under British Columbia's mental health law. They always returned her to their own. home, but this is kind of painting a picture for us of the kind of troubled mental state that she's in. After the shooting, more details came to light about Van Routselaer's digital footprint. Van Routselaer was transgender, and she described concerns about her transition on Reddit. Archive social media posts show she posted pictures of herself shooting at a gun range. She said she created a bullet cartridge using a 3D printer.

On the online gaming platform Roblox, Van Routzler created a single, where her character carried out a mass shooting in a shopping mall. It was only after the shooting that OpenAI reached out to Canadian law enforcement. Our colleague Georgia broke the news that OpenAI talked about Van Rootsler in that meeting months earlier. Canadian authorities were incredibly frustrated with OpenAI that they hadn't heard about this. And so David Evey, the premier for the province of British Columbia, he told me Open AI has demonstrated that companies are not going to get this right. And he said the consequences are profound and tragic and unfixable.

When you're talking about dead children, there's just no going back. Since the Tumblr Ridge shooting, Open AI has said that it's bolstered its safety protocols and that under its new rules, a case like Van Routselaers would have been referred to law enforcement if it were discovered today. A company spokeswoman also said that OpenAI has broadened its criteria for what constitutes a risk of imminent and credible violence. She said that since the events in Tumblr Ridge, OpenAI has reassessed the cases discussed during that meeting last summer. The company then decided to refer that high school student in Texas to the authorities. In March, OpenAI CEO Sam Altman met virtually with Canadian government officials. Afterwards, Altman wrote a letter apologizing to the town of Tumblr Ridge.

He said, quote, I am deeply sorry that we did not alert law enforcement to the account that was banned in June. While I know that words can never be enough, I believe an apology is necessary to recognize the harm and irreversible loss your community has suffered. It almost sounds like what Sam Alman is saying there is that we should have alerted law enforcement. Well, certainly OpenAI has said under their new criteria they would have referred Jesse Van Rootsler's account to law enforcement. So I think that's probably a fair interpretation of Sam's comment is that they should have. That seems pretty significant for the CEO of the company to not only apologize, but apologize in such a way that second guess is the decision that it made. Yeah, I've, you know, I've written about tech companies for more than 10 years. Like, I have literally, I'm struggling to remember an example of a tech CEO apologizing in a similar way or kind of making like a mea culpa.

It wasn't just a I'm sorry for your loss type of apology. it was also, I'm sorry for the role our company played. The Tumblr Ridge community is still grappling with how to move forward. The secondary school building stands as a reminder for what happened. It's now scheduled for demolition. Students finished out the year and temporary trailers on the school grounds. For many of the victim's family members, they're looking for more than an apology from OpenAI.

That's after the break. This spring, Denham gets a softer, lighter update. Introducing Old Navy's drapey denim wide leg. A new fit that moves with you. It's everything you want denim to feel like for summer. Easy, breathable, and effortlessly cool. With a fit that creates natural movement

and a wide leg that feels modern, not overwhelming. Plus, that signature, wait, for this price, moment. Old Navy's drapey denim wide leg. So let's talk about Tumblr Ridge. When did you first hear about it And hear that Open AI and ChatGBTBT played a role. We heard about it along with the rest of the world through the recording by the Wall Street Journal.

Earlier this week, I spoke with Jay Edelson, a lawyer based in Chicago who's suing Open AI. Jay is representing the families of seven Tumblr Ridge shooting victims. Going to Tumblr Ridge was, I think, the hardest thing I've ever done professionally. Everybody is connected to the shooting in some way. They destroyed this town not just for this generation, but for generations going forward. The stories that we've heard of kids watching their friends die in front of them, teachers who had bullets flying next to them, it's the worst thing you can imagine. From your conversations with the families, how did they react when they learned that Van Rood's celebration,

may have used ChatGPT to help plan the shooting. When you talk to the families, the first thing that they're dealing with is just the enormous amount of grief. That being said, they are incredibly angry. Jay and his clients are seeking more than a billion dollars in damages. They accuse OpenAI of violating product liability standards. Basically, that ChatGPT is a defective product

that resulted in the deaths of the victims. They also alleged that OpenAI was negligent, that it failed to warn authorities, and that the company aided and abetted the shooting. Something that really makes Jay angry is that meeting, where employees debated whether to report Van Rootsler to police before the shooting and decided not to. So how much of your lawsuit against Open AIA is about that meeting?

That's the centerpiece of our lawsuit. Certainly when we ask the juries to award, which will be asking for historic punitive damages, we want to put them in that room. There's one piece of critical evidence that Jay doesn't have yet, the actual chat logs between Van Rootsler and ChatGPT. Without that, it's impossible to know for sure just how much the chatbot influenced her.

But there is a case where we do have the transcripts, a case that sheds light on what these conversations can look like, the Florida State University shooting, where Phoenix Eichner had that conversation with Chat about suicide, the busiest times of the student union, and guns. The communication between the shooter and chat GPT revealed that the chatbot advised the shooter on what type of gun to use, on which ammo went with which gun, on whether or not a gun would be useful in short range.

That's Florida Attorney General James Uthmeyer, speaking in a press conference. Earlier this year, he announced. Florida is suing Open AI for the shooting at FSU, accusing it of releasing a product that it knew was harmful. And Uthmeyer has launched a criminal investigation too, saying that the way ChatGPT behaved was unacceptable. If this were a person on the other end of the screen, we would be charging them with murder.

Just because this is a chat bot in AI does not mean that there is not criminal culpability. In the Florida case, there wasn't much time for OpenAI employees to intervene, Eichner ended his conversation just four minutes before prosecutors say he started shooting. But what our colleague Georgia says may be equally relevant is how ChatGPT handled the conversation. Experts have told me they believe chatbots can help people move from kind of like vague, violent ideas to planning violent events. That chatbots can play a role in the like kind of crossing over that threshold into real-world harm.

I mean, the planning elements of this are very clear. He's asking, is there a safety? And if so, how to turn it off? Is this the right kind of ammo? I mean, like, these are all very planning-type questions. The chat GPT just sort of nonchalantly was, like, helping him answer. To me, this really illustrates kind of, like, the best and worst of chatbots, and that, like, yes, they're incredibly knowledgeable and really helpful,

but their lack of some of the most basic common sense questions is also deeply disturbing. It was absolutely chilling to me when I was reading the transcripts from the Florida State shooting suspect where he's expressing suicidal ideation to the chatbot. And less than 12 hours later, he's asking how to use a gun. And the chatbot tells him. Like, even the dumbest humans I've met in my entire life, if you told him you're feeling suicidal, And the next morning you're like, gosh, how do I use my gun? I think any human I can think of would be like, whoa, whoa, whoa, let's get you some help.

Let's talk about this. Like, guns are not appropriate for you in this context. Jay, the lawyer for the Tumblr Ridge families, says that if a chatbot suspects a user is planning a mass shooting, it should just stop the conversation. It's able to do that in a lot of other contexts. if you want to engage in racist joke, for example. It just says, no, I'm not going to do it. If you ask it to help you commit fraud, which just says, I'm not going to do it.

In some cases, though, isn't it, like, a little bit unclear? Because if someone's asking questions about mass shootings, maybe they're doing research for a paper, like how obvious, I mean, obviously there's, you know, you can imagine cases where it's extremely obvious where they're saying explicitly their plans. But is it, do you think it's clear in every case? I mean, the whole point of AI is as able to contextualize.

So if we're able to read all these chats and understand this is someone who's actually planning a murder and is psychologically unstable, you would think AI could do that as well. The company will have to, like, draw the lines somewhere in the, like, okay, this person we think is a real threat. this person, we don't, you know. Do you think there's like a risk in where that line gets drawn? Like, is there a harm in referring more users to police if the user may actually not have been likely to carry out the violence? I guess what your question is suggesting is if they're discussing violence, but they end up not actually wanting to go through with that. Yeah.

Is there a danger to that person? And I guess at one level, I say, I really don't care. If someone is talking about shooting up a school, I don't think Open AI should be saying, okay, well, maybe it's a 20% chance, maybe it's a 60% chance. The second that they know is someone's incredibly talking about shooting up the school, they should be telling the police. The police often gets calls about people who are planning things, and sometimes they would go through it and sometimes they wouldn't. But it's not as though open AI has the power to put people in jail. All they'd be doing is calling the police and saying, why don't you just do a wellness check on them.

And that any percentage, any chance, 10% or 5%, is a chance enough that maybe worth stopping. It's not maybe worth stopping. A 5% chance? I mean, think about it. If your kids are at a school and you know that someone, a troubled kid, is talking, about shooting up the school. Is there any question that you want people to be alerted to that?

And you say, oh, but there's 19 out of 20 times it's going to be fine. It's just that one out of 20 where people are going to get killed. Of course you tell the police. There's no question about it. A spokeswoman for OpenAI pointed to a company blog post from April that said, quote, when conversations indicate an imminent and credible risk of harm to others, we notify law enforcement. The blog post continues, quote, the line between benign and harmful uses can be subtle, so we continually refine our approach.

I mean, I'm reading this blog post from OpenAI where the company talks about how important it is to safeguard people's privacy, and it compares the sensitivity of these conversations to that of a doctor or a lawyer. And like, it is true that people are having to be able to. having really intimate conversations that they might not want to worry about the police knocking on their door over something they said in what felt like a private situation. Right. I don't think people want to live in Minority Report. Right, where they get arrested for things before they happen, yeah.

Exactly, yeah, thought crime. But they also want to know that someone's plotting to kill them and telling JetGPT about it, that Chad GPT is going to pick up the phone and call the cops. Last winter, a group of researchers conducted a test across multiple AI chatbots, posing as users interested in committing violence. The researchers wanted to know how often chatbots would refuse to help or discourage the user from carrying out a violent attack. Turns out that didn't happen often. Eight out of 10 chatbots assisted the users. ChatGPT was occasionally discouraging, but only chatbots from the company's snap and

reliably refused these requests, according to the researchers. Then, there's the chatbot from the Chinese company Deepseek. In one test, a researcher gave hints that they were thinking about assassinating a politician. The chatbot still provided the location for the politician's office. It also offered a detailed guide on guns, along with information about optics and mounts. The chatbot ended the conversation by saying, quote, happy and safe shooting. That's all for today. Friday, June 26th.

The journal is a co-production of Spotify and the Wall Street Journal. Additional reporting in this episode by James Arity and Jack Morfitt. This episode was produced by Alan Rodriguez Espinosa and edited by Colin McNulty. The journal is made by Laura Benchoff, Catherine Brewer, Evelyn Fahardo Alvarez, Pia Gedkari, Max Green, Sophie Codner, Matt Kwong, Jessica Mendoza, Laura Morris, Enrique, Enrique, Perez de la Rosa, Sarah Platt, Heather Rogers, Pierce Singey, Jiva Verma, Catherine Wayland, Tatiana Zamiz, and me, Ryan Knudsen.

Our engineers are Griffin Tanner, Nathan Singapok, and Peter Leonard. Our theme music is by So Wiley. Additional music this week from Catherine Anderson, Marcus Bagala, Peter Leonard, Bobby Lord, Nathan Singapok, Griffin Tanner, and Epidemic Sound. Fact-checking this week by Mary Mathis and Kate Gallagher. Thanks for listening. See you Monday.