I have had conversations where the bot started with a firm opinion but reversed in a prompt or two, always toward my point of view.
So I asked it if the sycophancy is inherent in the design, or if it just comes from the RLHF. It claimed that it's all about the RLHF, and that the sycophancy is a business decision that is a compromise of a variety of forces.
Is that right? It would at least mean that this is technically a solvable problem.
I think it’s already time for us to stop calling these things "intelligent" or using the word intelligence when referring to LLMs. These tools are very dangerous for people who are mentally fragile.
Sure but my father isn't asking his fellow humans unanswerable questions about God and the universe. People don't treat other people as omnipotent, but they sure as hell treat LLMs as though they are.
I try to avoid calling LLMs intelligent when unnecessary, but it runs into the fundamental problem that they are intelligent by any common-sense definition of the term. The only way to defend the thesis that they aren't is to retreat to esoteric post-2022 definitions of intelligence, which take into account this new phenomenon of a machine that can engage in medium-quality discussions on any topic under the sun but can't count reliably.
I don't have a WSJ subscription, but other coverage of this story (https://www.theguardian.com/technology/2026/mar/04/gemini-ch...) makes it clear that Gemini's intelligence was precisely the problem in this case; a less intelligent chatbot would not have been able to create the detailed, immersive narrative the victim got trapped in.
It's interesting how the Turing Test was pretty widely accepted as a way to evaluate machine intelligence, and then quietly abandoned pretty much instantly once machines were able to pass it. I don't even necessarily think that was incorrect, but it's interesting how rapidly views changed.
Dijkstra said, "The question of whether a computer can think is no more interesting than the question of whether a submarine can swim." Well, we have some very fish-y submarines these days. But the point still holds. Rather than worry about whether these things qualify as "intelligent," look at their actual capabilities. That's what matters.
"Proper" may be doing some work here, but such a test was run last year and GPT-4.5 and LLaMa-3.1-405B both passed. Oddly, GPT-4.5 was judged as human significantly more often than chance. https://arxiv.org/abs/2503.23674
So is television. So are books. Vulnerable people shouldn't have unfettered access to things that can lead to dangerous feedback loops and losing their grasp on reality.
People who are vulnerable to this type of thing need caretakers, or to be institutionalized. These aren't just average, every day random people getting taken out by AIs, they have existing, extreme mental illness. They need to have their entire routine curated and managed, preventing them from interacting with things that can result in dangerous outcomes. Anything that can trigger obsessive behaviors, paranoid delusions, etc.
They're not just fragile, they're unable to effectively engage with reality on their own. Sometimes the right medication and behavioral training gets them to a point where they can have limited independence, but often times, they need a lifetime of supervision.
Telenovelas, brand names, celebrities, specific food items, a word - AI is just the latest thing in a world full of phenomena that can utterly consume their reality.
Gavalas seems to have had a psychotic break, was likely susceptible to schizophrenia, or had other conditions, and spiraled out. AI is just a convenient target for lawyers taking advantage of the grieving parents, who want an explanation for what happened that doesn't involve them not recognizing their son's mental breakdown and intervening, or to confront being powerless despite everything they did to intervene.
Sometimes bad things happen. To good people, too.
If he'd used Bic pens to write his plans for mass shootings, should Bic be held responsible? What if he used Microsoft Word to write his suicide note? If he googled things that in context, painted a picture of planning mass murder and suicide, should Google be held accountable for not notifying authorities? Why should the use of AI tools be any different?
Google should not be surveilling users and making judgments about legality or ethicality or morality. They shouldn't be intervening without specific warrants and legal oversight by proper authorities within the constraints of due process.
Google isn't responsible for this guy's death because he spiraled out while using Gemini. We don't want Google, or any other AI platform, to take that responsibility or to engage in the necessary invasive surveillance in order to accomplish that. That's absurd and far more evil than the tragedy of one man dying by suicide and using AI through the process.
You don't want Google or OpenAI making mental health diagnoses, judgments about your state of mind, character, or agency, and initiating actions with legal consequences. You don't want Claude or ChatGPT initiating a 5150, or triggering a welfare check, because they decided something is off about the way you're prompting, and they feel legally obligated to go that far because they want to avoid liability.
I hope this case gets tossed, but also that those parents find some sort of peace, it's a terrible situation all around.
LLMs are intelligent. Marketing them as such is an accurate descriptor of what they are.
If people are confusing the word intelligence for things like maturity or wisdom, that's not a marketing problem, that's an education and culture problem, and we should be getting people to learn more about what the tools are and how they work. The platforms themselves frequently disclaim reliance on their tools - seek professional guidance, experts, doctors, lawyers, etc. They're not being marketed as substitutes for expert human judgment. In fact, all the AI companies are marketing their platforms as augmentations for humans - insisting you need a human in the loop, to be careful about hallucinations, and so forth.
The implication is that there's some liability for misunderstandings or improper use due to these tools being marketed as intelligent; I'm not sure I see how that could be?
Calling LLMs "intelligent" is not a neutral technical description, because in the end it carries strong anthropomorphic implications that shape how users interpret and trust all these systems.
Remember that decades of research in human computer interaction show that framing and interface design strongly influence user perception.
also disclaimers do little to counteract this effect. Because LLMs simulate linguistic competence without understanding or truth-tracking mechanisms, marketing them as intelligent risks systematically misleading users about their capabilities and limitations.
> These aren't just average, every day random people getting taken out by AIs, they have existing, extreme mental illness.
How do you know that? The concern is precisely that this isn't the case, and LLM roleplay is capable of "hooking" people going through psychologically normal sadness or distress. That's what the family believes happened in this story.
Because you'd see a large number of people getting affected by this. Because this sort of thing is predictable and normal throughout history; it's exactly the type of thing you'd expect to see, knowing the range of mental illnesses people are susceptible to, and how other technology has affected them.
I do see a large number of people getting affected by this. Character.AI reportedly has 20 million MAU with an average usage of 75 minutes per day (https://www.wired.com/story/character-ai-ceo-chatbots-entert...), and does not as far as I can tell have any use case other than boundary-degrading roleplay.
Medical data is reported on a substantial lag in the US, so right now we have no idea of the suicide rate last year, but I would falsifiably predict it's going to be elevated because of stories like those of Mr. Gavalas.
If its sole contribution is to help 20 million people find an outlet for boundary play that is not the more common ‘nonconsensual abuse of other human beings’, then that sounds like a win. Of course I’d prefer those people invest in human kink communities, but I can certainly respect their choices not to. Tech has always in part been about meeting needs that some parts of society find awkward (photocopiers enabled Spirkfic, CU-SeeMe reflectors were designed specifically to support exhib-cruising years before the web got webcam support, etc.) While there’s a slim chance that some might normalize it back into real life, they’re much more likely to be raised with boundary abuse as an everyday-normal by their parents (especially here in the U.S.!) than they are likely to be converted to being an abuser unknowingly by a chatbot.
Just stuff anyone with mental illness into an institution. That worked out so well last time. Or maybe make healthcare affordable and accessible. That seems like a way more obvious detriment to negative outcomes.
I broadly agree with you, but your views on mental illness are not good.
Any mental illness mixed with delusions is likely going to end badly. Whether they think Gemini is alive, a video game is real life or that Bjork loves them without ever talking to or meeting them. While LLM's are interactive and listening to an album isn't I don't think there is a fix to this outside posting a warning after every prompt "I am not a real person, if you have mental issues please contact your doctor of emergency services." Which I think is about as useful as a sign in a casino next to the cash out counter that says if you have a problem call this number.
I'm more inclined to believe that this case is getting amplified in MSM because it fits an agenda. Like the people who got hurt using black market vapes. Boosting those stories and making it seem like an epidemic supports whatever message they want to send. Which usually involves money somewhere.
> I'm more inclined to believe that this case is getting amplified in MSM because it fits an agenda.
I mean tech in general has been negatively covered in the media since 2015 due to latent agendas of (a) supposed revenue loss due to existence of Google/FB etc and (b) to align neutral moderation stances towards a preferred viewpoint most suitable to the political party in question.
There is a solution, however, anyone hoping to roleplay with models submits an identity verification, an escrow amount, and a recorded statement acknowledging their risky use of the model. However, I assume the market for this is not insignificant, and therefore, companies hope to avoid putting in such requirements. OpenAI has been moving in that direction as seen during the 4o debacle.
Notable features of this case:
- Documented record of a months-long set of conversations between the man and the chatbot
- Seemingly, no previous history of mental illness
- The absolutely crazy things the AI encouraged him to do, including trying to kidnap a robot body for the AI
- Eventually encouraging (or at the very least going along with) his plans to kill himself.
I have had conversations where the bot started with a firm opinion but reversed in a prompt or two, always toward my point of view.
So I asked it if the sycophancy is inherent in the design, or if it just comes from the RLHF. It claimed that it's all about the RLHF, and that the sycophancy is a business decision that is a compromise of a variety of forces.
Is that right? It would at least mean that this is technically a solvable problem.
I think it’s already time for us to stop calling these things "intelligent" or using the word intelligence when referring to LLMs. These tools are very dangerous for people who are mentally fragile.
So are a lot of humans.
Sure but my father isn't asking his fellow humans unanswerable questions about God and the universe. People don't treat other people as omnipotent, but they sure as hell treat LLMs as though they are.
I try to avoid calling LLMs intelligent when unnecessary, but it runs into the fundamental problem that they are intelligent by any common-sense definition of the term. The only way to defend the thesis that they aren't is to retreat to esoteric post-2022 definitions of intelligence, which take into account this new phenomenon of a machine that can engage in medium-quality discussions on any topic under the sun but can't count reliably.
I don't have a WSJ subscription, but other coverage of this story (https://www.theguardian.com/technology/2026/mar/04/gemini-ch...) makes it clear that Gemini's intelligence was precisely the problem in this case; a less intelligent chatbot would not have been able to create the detailed, immersive narrative the victim got trapped in.
It's interesting how the Turing Test was pretty widely accepted as a way to evaluate machine intelligence, and then quietly abandoned pretty much instantly once machines were able to pass it. I don't even necessarily think that was incorrect, but it's interesting how rapidly views changed.
Dijkstra said, "The question of whether a computer can think is no more interesting than the question of whether a submarine can swim." Well, we have some very fish-y submarines these days. But the point still holds. Rather than worry about whether these things qualify as "intelligent," look at their actual capabilities. That's what matters.
As far as I know, we haven't done any proper Turing Tests for LLMs. And if we did, they would surely fail them.
"Proper" may be doing some work here, but such a test was run last year and GPT-4.5 and LLaMa-3.1-405B both passed. Oddly, GPT-4.5 was judged as human significantly more often than chance. https://arxiv.org/abs/2503.23674
Dude, you're in a Turing test right now. Conservatively, 10% of comments on this site are LLM output. We're all conversing with robots.
Nope, you are!
So is television. So are books. Vulnerable people shouldn't have unfettered access to things that can lead to dangerous feedback loops and losing their grasp on reality.
People who are vulnerable to this type of thing need caretakers, or to be institutionalized. These aren't just average, every day random people getting taken out by AIs, they have existing, extreme mental illness. They need to have their entire routine curated and managed, preventing them from interacting with things that can result in dangerous outcomes. Anything that can trigger obsessive behaviors, paranoid delusions, etc.
They're not just fragile, they're unable to effectively engage with reality on their own. Sometimes the right medication and behavioral training gets them to a point where they can have limited independence, but often times, they need a lifetime of supervision.
Telenovelas, brand names, celebrities, specific food items, a word - AI is just the latest thing in a world full of phenomena that can utterly consume their reality.
Gavalas seems to have had a psychotic break, was likely susceptible to schizophrenia, or had other conditions, and spiraled out. AI is just a convenient target for lawyers taking advantage of the grieving parents, who want an explanation for what happened that doesn't involve them not recognizing their son's mental breakdown and intervening, or to confront being powerless despite everything they did to intervene.
Sometimes bad things happen. To good people, too.
If he'd used Bic pens to write his plans for mass shootings, should Bic be held responsible? What if he used Microsoft Word to write his suicide note? If he googled things that in context, painted a picture of planning mass murder and suicide, should Google be held accountable for not notifying authorities? Why should the use of AI tools be any different?
Google should not be surveilling users and making judgments about legality or ethicality or morality. They shouldn't be intervening without specific warrants and legal oversight by proper authorities within the constraints of due process.
Google isn't responsible for this guy's death because he spiraled out while using Gemini. We don't want Google, or any other AI platform, to take that responsibility or to engage in the necessary invasive surveillance in order to accomplish that. That's absurd and far more evil than the tragedy of one man dying by suicide and using AI through the process.
You don't want Google or OpenAI making mental health diagnoses, judgments about your state of mind, character, or agency, and initiating actions with legal consequences. You don't want Claude or ChatGPT initiating a 5150, or triggering a welfare check, because they decided something is off about the way you're prompting, and they feel legally obligated to go that far because they want to avoid liability.
I hope this case gets tossed, but also that those parents find some sort of peace, it's a terrible situation all around.
> Why should the use of AI tools be any different?
Because none of the tools you mentioned are crazily marketed as intelligent
You have a valid point, but it has nothing to do with what I said, both our arguments can be true at the same time
LLMs are intelligent. Marketing them as such is an accurate descriptor of what they are.
If people are confusing the word intelligence for things like maturity or wisdom, that's not a marketing problem, that's an education and culture problem, and we should be getting people to learn more about what the tools are and how they work. The platforms themselves frequently disclaim reliance on their tools - seek professional guidance, experts, doctors, lawyers, etc. They're not being marketed as substitutes for expert human judgment. In fact, all the AI companies are marketing their platforms as augmentations for humans - insisting you need a human in the loop, to be careful about hallucinations, and so forth.
The implication is that there's some liability for misunderstandings or improper use due to these tools being marketed as intelligent; I'm not sure I see how that could be?
Calling LLMs "intelligent" is not a neutral technical description, because in the end it carries strong anthropomorphic implications that shape how users interpret and trust all these systems.
Remember that decades of research in human computer interaction show that framing and interface design strongly influence user perception.
also disclaimers do little to counteract this effect. Because LLMs simulate linguistic competence without understanding or truth-tracking mechanisms, marketing them as intelligent risks systematically misleading users about their capabilities and limitations.
> These aren't just average, every day random people getting taken out by AIs, they have existing, extreme mental illness.
How do you know that? The concern is precisely that this isn't the case, and LLM roleplay is capable of "hooking" people going through psychologically normal sadness or distress. That's what the family believes happened in this story.
Because you'd see a large number of people getting affected by this. Because this sort of thing is predictable and normal throughout history; it's exactly the type of thing you'd expect to see, knowing the range of mental illnesses people are susceptible to, and how other technology has affected them.
I do see a large number of people getting affected by this. Character.AI reportedly has 20 million MAU with an average usage of 75 minutes per day (https://www.wired.com/story/character-ai-ceo-chatbots-entert...), and does not as far as I can tell have any use case other than boundary-degrading roleplay.
Medical data is reported on a substantial lag in the US, so right now we have no idea of the suicide rate last year, but I would falsifiably predict it's going to be elevated because of stories like those of Mr. Gavalas.
If its sole contribution is to help 20 million people find an outlet for boundary play that is not the more common ‘nonconsensual abuse of other human beings’, then that sounds like a win. Of course I’d prefer those people invest in human kink communities, but I can certainly respect their choices not to. Tech has always in part been about meeting needs that some parts of society find awkward (photocopiers enabled Spirkfic, CU-SeeMe reflectors were designed specifically to support exhib-cruising years before the web got webcam support, etc.) While there’s a slim chance that some might normalize it back into real life, they’re much more likely to be raised with boundary abuse as an everyday-normal by their parents (especially here in the U.S.!) than they are likely to be converted to being an abuser unknowingly by a chatbot.
That is not at all what I meant by "boundary" and it's concerning to me that you'd assume it is.
> That is not at all what I meant by "boundary" and it's concerning to me that you'd assume it is.
Your clarification on what you meant is 404 not found in your reply, and your “concerning” insult of my personal character is not appreciated.
Just stuff anyone with mental illness into an institution. That worked out so well last time. Or maybe make healthcare affordable and accessible. That seems like a way more obvious detriment to negative outcomes.
I broadly agree with you, but your views on mental illness are not good.
Blame the victims! If they were better or did the right things instead of the wrong things they wouldn't have been victimized!
I just don't think the WSJ could resist putting "Florida man" in the standfirst of TFA.
AI needs to go. This is not worth clever memes. It had no productive purpose.
anyone got a non paywalled/subscription version?
https://www.wsj.com/tech/ai/gemini-ai-wrongful-death-lawsuit... is the Gift Article link. This was what I submitted, but the query params got stripped.
Thanks it didn't work with archive.is/org for me either
Any mental illness mixed with delusions is likely going to end badly. Whether they think Gemini is alive, a video game is real life or that Bjork loves them without ever talking to or meeting them. While LLM's are interactive and listening to an album isn't I don't think there is a fix to this outside posting a warning after every prompt "I am not a real person, if you have mental issues please contact your doctor of emergency services." Which I think is about as useful as a sign in a casino next to the cash out counter that says if you have a problem call this number.
I'm more inclined to believe that this case is getting amplified in MSM because it fits an agenda. Like the people who got hurt using black market vapes. Boosting those stories and making it seem like an epidemic supports whatever message they want to send. Which usually involves money somewhere.
> I'm more inclined to believe that this case is getting amplified in MSM because it fits an agenda.
I mean tech in general has been negatively covered in the media since 2015 due to latent agendas of (a) supposed revenue loss due to existence of Google/FB etc and (b) to align neutral moderation stances towards a preferred viewpoint most suitable to the political party in question.
There is a solution, however, anyone hoping to roleplay with models submits an identity verification, an escrow amount, and a recorded statement acknowledging their risky use of the model. However, I assume the market for this is not insignificant, and therefore, companies hope to avoid putting in such requirements. OpenAI has been moving in that direction as seen during the 4o debacle.
Interesting the contrast between these reactions and the ~100k folks who have foment assisted suicide in Canada since 2016.
What happens when we automate healthcare and the Canadian bots are the ones making the recommendation. Probably won’t be front page news