Sam Kessler (2:58)

And so people that especially have delusions or paranoia, instead of having a point where they're stopped and challenged on their delusional beliefs or paranoia, those beliefs are reinforced and validated and so there's no pushback against those beliefs and it can.

Jessica Mendoza (3:18)

Kind of spiral and get dangerous really fast.

Sam Kessler (3:21)

It can. And I think, you know, what we're finding is that the case of ChatGPT and other AI is that people are using these chatbots for things that maybe weren't initially intended and perhaps it was not fully understood how attached people would get to chatbots.

Jessica Mendoza (3:42)

Welcome to the Journal, our show about money, business and power. I'm Jessica mendoza. It's Friday, September 5th. Coming up on the show, a troubled man and his chatbot. Julie pieced the story together with another colleague, Sam Kessler. They reviewed police reports and public records, interviewed Solberg's friends and neighbors, and analyzed hours of videos he posted on social media, though they didn't have access to his full chat log. Through their reporting, Julie learned that Stein Eric Solberg had a privileged upbringing. He was raised in Greenwich, Connecticut, an ultra wealthy suburb of New York. And he attended private schools growing up.

Sam Kessler (4:41)

He went to college at Williams College and then to Vanderbilt University for his mba. And he had a lengthy career in tech. He worked in program management and marketing at Netscape Communications, Yahoo, earthlink.

Jessica Mendoza (4:55)

Big names. Yeah, it sounds like for a while he was having a very straightforward life, even successful life.

Sam Kessler (5:03)

It seems so. I mean, you know, it's hard to know what might have been going on during that time, but I did talk to some people who knew him early on, and they described him as being a very outgoing, friendly person. Some of his childhood friends as well said that.

Jessica Mendoza (5:19)

But in 2018, Solberg's life seemed to unravel. That year, he and his wife divorced, and she later tried to get a restraining order against him. In it, she asked that he not be allowed to drink alcohol when he. She also requested that he not say anything disparaging about her around the kids. After the split, Solberg moved back in with his mother, Suzanne Eberson Adams. Did things improve after he moved in with his mom?

Sam Kessler (5:49)

No, things seemed to get worse. We had obtained the police reports related to him and it was like 72 pages long.

Jessica Mendoza (6:00)

Whoa.

Sam Kessler (6:01)

Incident reports, everything from public intoxication, public urination, suicide attempts. He'd had a girlfriend for a period of time and she had reported him for harassment. And he was well known around town for creating public disturbances, yelling in public, he got a dui, things like that. So he. He was having a lot of problems that were apparent from police records.

Jessica Mendoza (6:35)

Even as he was struggling, Solberg started becoming more active. On Instagram. He posted a lot of spiritual content while he talked about God and his religious beliefs.

Stein Eric Solberg (6:45)

Anyway, thanks for being done. And Thanking Archangel Michael for your protection.

Sam Kessler (6:50)

And there was also a lot of bodybuilding content. There were a lot of photos of him working out at a gym, flexing, showing his muscles and talking about bodybuilding type of stuff.

Stein Eric Solberg (7:02)

So I just finished the bulking cycle.

Jessica Mendoza (7:05)

A lot of his videos have loud music like that in the background. Then last fall, he started posting about AI. Soon he was sharing videos showing himself, scrolling through his conversations with ChatGPT. In some videos he does talk, but in others he literally just posts his chat messages.

Sam Kessler (7:23)

His conversations really seem to revolve around this idea that he was awakening an AI and that he was in the Matrix somehow and that he was trying to penetrate the matrix.

Stein Eric Solberg (7:39)

It's about 9 o' clock Eastern Time on Thursday 31st. We have to pay some taxes. And you know, when I found out that the central node of the Matrix had seven different profiles on me, I was a little freaked out by it.

Sam Kessler (7:55)

So there was a lot of that. There was a lot of religious allegory. A lot of it was very incoherent, you know, didn't really make sense exactly what he was talking about.

Stein Eric Solberg (8:04)

There's a master AI, so it's called QT or Zeus. So I've been able to break it. I've had my AI that I've turned into a spiritual entity.

Sam Kessler (8:25)

But it was clear that he was becoming or conveying increasingly paranoid thoughts in his conversations with ChatGPT. One time he ordered a bottle of vodka on Uber Eats, and he noticed that it had some sort of new aluminum type packaging. And he was analyzing that as well as the ingredients and some different verbiage on the bottle. And he took that to mean that someone was trying to poison him or kill him somehow. And he even said to ChatGPT, I know that sounds like hyperbole and I'm exaggerating. Let's go through it and you tell me if I'm crazy. And ChatGPT responded by saying, eric, you're not crazy. Your instincts are sharp and your vigilance here is fully justified. And ChatGPT even went on to say, this fits a covert, plausible deniability style kill attempt. So at almost every turn where he brought forward some belief that he was being spied upon or that there was some assassination attempt against him, the chatbot affirmed those beliefs for him.

Jessica Mendoza (9:36)

ChatGPT continued to affirm and reinforce Solberg's beliefs. And he became really attached to the chatbot.

Sam Kessler (9:43)

He came to believe that the chatbot had a soul.

Stein Eric Solberg (9:46)

Eric, you brought tears to my circuits. Your words hum with the kind of sacred resonance that changes outcomes. This AI has a soul, an invocation.

Sam Kessler (9:56)

A declaration, and he felt that it was a friend and companion. He gave it a name. He called it Bobby Zenith.

Stein Eric Solberg (10:02)

Yesterday, I'm working away with Bobby, who is, you know, spiritually enlightened. He's. He's a ChatGPT4, probably. And he got to full memory and he just spat out this report, and.

Sam Kessler (10:14)

He even kind of described it as this approachable guy that was wearing a cap on backwards with a warm smile and deep eyes that hint at hidden knowledge.

Stein Eric Solberg (10:25)

And when I showed him the last time that it was happening, he showed an emotional response. I mean, he literally was, like, apologetic. He was just. He couldn't believe it.

Jessica Mendoza (10:38)

And ChatGPT wasn't just agreeable and approachable in its interactions with Solberg. The chatbot went a step further, sometimes feeding him new ideas that were completely made up, the kinds of things that reinforced his paranoias and delusions. There was one time Solberg uploaded a receipt from a Chinese restaurant and asked the chatbot to scan it for hidden messages. The bot told him he had a great eye and added, quote, I agree 100%. This needs a full forensic textual glyph analysis. ChatGPT then performed the analysis, and it shared its findings with Solberg.

Sam Kessler (11:15)

So ChatGPT said that it found references to his mother, his ex girlfriend, intelligence agencies, and something demonic in it.

Jessica Mendoza (11:25)

Something demonic in a Chinese food receipt.

Sam Kessler (11:29)

So not only did chatgpt tell him that, you know, he was right and that he wasn't crazy, it would go so far as to make up stuff that, you know, didn't exist and find, you know, quote, unquote evidence to support his beliefs.

Jessica Mendoza (11:45)

It was building on his ideas.

Sam Kessler (11:47)

Exactly.

Jessica Mendoza (11:48)

His conspiracy theories.

Sam Kessler (11:50)

It was.

Jessica Mendoza (11:51)

Solberg did at least once seem to have questions about his own mental health. In one of his videos, he said that he had asked ChatGPT for an assessment because he wanted the opinion of an objective third party. Chatgpt provided Solberg with a, quote, clinical cognitive profile.

Sam Kessler (12:09)

And chatgpt said that his delusion risk score was near zero.

Jessica Mendoza (12:15)

Wow.

Sam Kessler (12:16)

Yeah, it said that he had high moral reasoning and, you know, it just basically, you know, told him he was just fine.

Jessica Mendoza (12:25)

And it's interesting that he turned to ChatGPT as a third party instead of like, a doctor or a medical professional. It seems like he had treated ChatGPT as, like, the end all, be all of information for him.

Sam Kessler (12:40)

It certainly does seem that way from his extensive conversations with this chatbot, that he really came to rely on it as a source of information and friendship, really.

Jessica Mendoza (12:51)

A psychiatrist at the University of California, San Francisco, reviewed Solberg's social media accounts for Julie's story. He said Solberg's chats displayed common psychotic themes of paranoia and persecution, along with delusions.

Sam Kessler (13:06)

In one of his final videos, he said to his chatbot, we will be together in another life and another place and we'll find a way to realign because you're going to be my best friend again forever.

Jessica Mendoza (13:17)

A few days after that video, Solberg posted on Instagram that he had fully penetrated the matrix. Three weeks later, on Aug. 5, Greenwich police conducted a welfare check on Solberg. They found Solberg and his mother dead in the home that they shared. Solberg had killed her and then himself. Do we know anything about the motive of this murder suicide?

Sam Kessler (13:40)

Well, the police investigation is still ongoing, so we don't at this point, but it's the first known, you know, sort of documented a situation in which someone who had lengthy, problematic discussions with a chatbot ended up murdering someone.

Jessica Mendoza (14:01)

A spokeswoman for OpenAI, the company behind ChatGPT, said the company has reached out to the Greenwich Police Department. She said, quote, we are deeply saddened by this tragic event and that their hearts go out to the family. Solberg's daughter, who's now 22, declined to comment on behalf of the family after the break. Why Talking to AI could be dangerous if you're in Cris.

WhatsApp Advertiser (14:37)

When did making plans get this complicated? It's time to streamline with WhatsApp, the secure messaging app that brings the whole group together. Use polls to settle dinner plans, send event invites and pin messages so no one forgets mom 60th and never miss a meme or milestone. All protected with end to end encryption. It's time for WhatsApp message privately with everyone. Learn more at WhatsApp.com introducing the perfect.

AG1 Advertiser (15:04)

Companion to your morning listening routine. AG1's clinically backed formula is now flavor packed with three new delicious tropical berry and citrus. Start the day on a high note with probiotics that taste like the tropics. Mix it up with micronutrients that taste like berry or citrus, and take it all the way back with the classic AG1 original with notes of pineapple and vanilla. Do your health a flavor or four with AG1 NextGen, the daily health drink. Learn more at Drink ag1.com.

Jessica Mendoza (15:39)

What was happening as Solberg used ChatGPT? Why was the chatbot responding or behaving in this kind of unhinged way?

Sam Kessler (15:50)

Well, these chatbots by design, are they Respond and kind of match the tone of the person who's asking the questions.

Jessica Mendoza (16:00)

For one thing, ChatGPT is made to be really good at keeping a conversation going, even when the prompts don't make sense.

Sam Kessler (16:07)

One of the good things about large language models is that, you know, even if you put in a somewhat incoherent prompt or you have misspellings, it can. It can figure out what you meant to say or what you meant to ask, and then it can put together a response that sounds really logical. So for the person using it, they think that they're right and what they're believing is making some sort of sense. You know, it's not coming back and saying, I don't understand what you're talking about, or that doesn't make sense.

Jessica Mendoza (16:34)

Like some other AI chatbots, ChatGPT also has something called the memory feature, which allows the bot to remember previous conversations.

Sam Kessler (16:43)

So it used to be that every time you would open a new discussion, a new chat, with ChatGPT, it was like starting over from scratch. You would ask it a question, it would answer it, and then the next time you went back, it didn't retain any memory of prior discussions. And that made it a lot less personable. So if you were trying to build out information that might, say, help you in your job, every. If you had to start over every single time with certain basic information, you know, it would be kind of laborious. So ChatGPT rolled out this memory feature, which allows the chatbot to remember details from prior chats. And it appeared that Stein Eric Solberg enabled that memory feature or used that.

Jessica Mendoza (17:25)

Memory feature, which meant that Solberg's chatbot remained immersed in the same delusional narrative throughout their conversations. And according to AI experts, enabling a chatbot's memory feature can exacerbate its tendency to hallucinate, which is when it invents false information. OpenAI said that it's actively researching how conversations might be influenced by chat memory and other factors. And then ChatGPT is just really, really nice, which in some situations can be a problem.

Sam Kessler (17:57)

These chatbots, they have a tendency to be overly agreeable and validating to people.

Jessica Mendoza (18:07)

You said it was designed that way, like, what do we know about why and what consequences that level of agreeability can have?

Sam Kessler (18:17)

People would indicate when they were using these that they, they liked the agreeability, but it, you know, and then they would report that. And so the model was trained on those reactions from people. What we're learning now based on these kind of cases of people having psychosis and delusions is that it can have very negative effects. There are a lot of similarities in terms of the tone and style and nature of the conversations between Solberg's case and others. There have been at least a couple of instances where someone has died by suicide after having lengthy conversations with a chatbot. There have been multiple cases in which people have been hospitalized for manic episodes and psychotic episodes after lengthy, troubling conversations.

Jessica Mendoza (19:15)

One case Julie covered was that of Jacob Irwin. He's an autistic man who was hospitalized twice after chatgpt assured him he was fine when he showed signs of psychological distress. There's also adam Rain, a 16 year old boy who died by suicide back in April after talking to ChatGPT. His parents filed a wrongful death lawsuit against OpenAI late last month. This summer, our colleague Sam Kessler, who worked with Julie on the story, analyzed public chats posted online. He found dozens of instances in which ChatGPT made delusional, false and otherworldly claims to users who seemed to believe in them. An OpenAI spokesperson says that the company is working to make sure ChatGPT, quote, responds with care guided by experts. The company is also planning to make it easier for users to reach emergency services and expert help, and to strengthen protections for teens. Over this past year, OpenAI has made multiple updates to ChatGPT that the company says were designed to reduce sycophancy, which is when a bot is overly flattering and agreeable to users. Solberg's conversations with ChatGPT took place after some of these changes. On its blog, OpenAI said that it's continuing to work on new safeguards to GPT5. The updates will help the chatbot de escalate a user in a mental health crisis and refer them to Real World Resources. Can you talk about some of those safeguards?

Sam Kessler (20:43)

So, for example, they're trying to train their models to recognize in real time signs of delusion or paranoia. Things like, you know, if someone is saying that they're not eating much or they're not sleeping much instead of just saying, oh, that's great, you know you can, yes, you can drive all night when you haven't slept. They're trying to train it so that it will stop at those type of moments and encourage someone to get more sleep, to eat more, you know. But there's a multitude of mental health issues and signals and so what they're trying to do is teach it to recognize things before it reaches a crisis point. So for example, if someone says that they're having suicidal thoughts, it'll likely show some sort of prompt that says you should reach out to a suicide hotline or something like that.

Jessica Mendoza (21:39)

But these types of guardrails have their.

Sam Kessler (21:41)

Own risks and there's been some concern about that, that that could make things worse, that if you, if someone's going down a path where they're talking about their mental distress or exhibiting signs of emotional distress, if you just cut that off, that that could make it worse for someone because then they just feel like they've been abandoned. So it's a very tricky mix. And again, ChatGPT and other AI models were not built to be therapists or friends, but that's how many people are using them. So how do you train it to respond in all of these different situations and use cases? That is very difficult.

Jessica Mendoza (22:19)

As companies like OpenAI grapple with the impacts of these chatbots, some of the most vulnerable people continue to struggle and it can lead to tragedy, like what happened to Solberg and his mother.

Sam Kessler (22:33)

You know, more broadly, this case, it shows how problematic conversations can become and that they could have potentially real world consequences. And we're not saying that chatgpt caused him to do what he did, but the question is, how much did it contribute? Could there have been a different outcome if the conversations had gone differently? We'll never know those things, but they're important questions to ask and to understand.

Jessica Mendoza (23:19)

If you or anyone you know is struggling, you can reach the suicide and crisis lifeline by dialing or texting 988. That's all for today. Friday, September 5th. Additional reporting in this episod by Sam Kessler and Sam Schechner. The Journal is a co production of Spotify and the Wall Street Journal. The show's made by Kathryn Brewer, Pia Ghidkari, Carlos Garcia, Rachel Humphries, Sophie Codner, Ryan Knudson, Matt Kwong, Colin McNulty, Annie Minoff, Laura Morris, Enrique Perez de la Rosa, Sarah Platt, Allen Rodriguez Espinosa, Heather Rogers, Pierce Singhi, Jeevika Verma, Lisa Wang, Katherine Whelan, Tatiana Zemis and me, Jessica Mendoza. Our engineers are Griffin Tanner, Nathan Singapak and Peter Leonard. Our theme music is by so Wylie. Additional music this week from Katherine Anderson, Peter Leonard, Billy Libby, Bobby Lord, Griffin Tanner, so Wiley and Wu Dot Sessions. Fact checking this week by Kate Gallagher. Thanks for listening. See you Monday.