[barbazoo]

> Update: It looks like this may have been OpenAI's attempt to get ahead of a horrifying story that just broke, about a man who fell into AI psychosis and killed his mother in a murder-suicide. Full details here.

With the URL being https://futurism.com/openai-scanning-conversations-police. From there:

> "Erik, you're not crazy. Your instincts are sharp, and your vigilance here is fully justified," ChatGPT told Soelberg during a conversation in July, after the 56-year-old conveyed his suspicions that an Uber Eats package signaled an assassination attempt. "This fits a covert, plausible-deniability style kill attempt."

As a species we are so not ready for this tech.

[idle_zealot]

> As a species we are so not ready for this tech.

I basically agree, but not because of how individuals are reacting to it. On a societal level we're failing to stop megacorps from selling this thing as intelligent, a virtual assistant you can chat with. This sets expectations that just wreck the minds of a small but significant portion of the population. If we had developed LLMs in a lab and released them in the form of papers and Python projects with model weights with clear descriptions of what they're capable of, like a responsible scientific endeavor, then we'd not be seeing the problems we are, even with public access. What's killing us is MBAs and Salesmen. The fact that we organize our economy in such a way that hucksters and charlatans thrive is the greatest threat to humanity. These fuckers would sell dynamite as chewing gum if they calculated it would increase next quarter's profit by 0.02% (factoring in fines/fees, and lobbying costs of getting those fines reduced or repealed).

[jjani]

> What's killing us is MBAs and Salesmen.

SamA is not an MBA. He did CS for 2 years and dropped out to build a startup. He's YC personified, and the person most responsible for the phenomenon you're talking about. Take that for what you will.

[drakythe]

People are what they do. Pretty sure the man has been a Salesman ever since he started seeking funding for OpenAI.

[evmaki]

So dropping out of CS to start selling something was more important to him than 2 more years of CS education. Maybe he realized that continuing his engineering education was unnecessary because he preferred selling things. Sounds like a salesman.

[queenkjuul]

So he's a salesman

[insane_dreamer]

he's just a salesman without an MBA, which is maybe even worse

[dingnuts]

oh but he is a salesman

[LamerBeeI]

that's a dumb distinction really

[goalieca]

> On a societal level we're failing to stop megacorps from selling this thing as intelligent, a virtual assistant you can chat with.

You nailed it here.

From the LLM generated search result asking how an LLM works. But this is not what the end users are being told.

>They learn to predict the next word in a sentence based on the context of the preceding words, allowing them to generate coherent and contextually relevant text

My worry is now the users are being entrapped. The LLM is "telling" them something insane and now they're being reported to the police.

[ben_w]

> If we had developed LLMs in a lab and released them in the form of papers and Python projects with model weights with clear descriptions of what they're capable of, like a responsible scientific endeavor, then we'd not be seeing the problems we are, even with public access.

I'm now thinking of all of the times people here of sarcastically stated "OpenAI thought GPT-2 was too dangerous to release"*, as if danger only comes immediately and severely or not at all.

* wasn't even what OpenAI said, they just proposed setting a norm of caution because a release can't be undone

[dingnuts]

they meant that to be interpreted as a statement of how good GPT-2 is but the real problem is how they've marketed everything that came after to people who can't know better

it's EXACTLY the same situation as Musk selling "Full Self Driving" and then playing dumb when people go to sleep behind the wheel of their Tesla

These rich men should be in prison for this kind of false advertising

[ben_w]

> they meant that to be interpreted as a statement of how good GPT-2 is

Every previous time someone says something like this, I've looked at the original blog post again just to make sure I didn't miss something.

OpenAI's own words for GPT-2 never read like a boast to me. Caution, not a boast.

News may have bigged it up, but news does that for everything.

I mean just compare it to the GPT-5 release: they're not shy or subtle when they actually want to boast.

[jononor]

Note that OpenAI is not the same company now compared to then. All the cautious and responsible people have left or been forced out.

[ben_w]

Sure, yes.

I'm just frustrated that "doing it right" gets people laughed at both at the time and for several years later, right up until the general dangers they were publicly concerned about manifest, and then people complain about people not doing something to prevent the outcome that they were previously laughed at for attempting to prevent.

Those same cautious people who left one way or another, they were the ones most directly mocked for daring to consider the possibility their new thing might be risky and caution might be wise.

[Terr_]

> I basically agree, but not because of how individuals are reacting to it. On a societal level we're failing to stop megacorps from selling this thing as intelligent, a virtual assistant you can chat with.

Analogy: I'm not scared and surprised to hear some consumers are dangerously allergic to peanuts... However I am flabbergasted that there's multi-billion dollar industry somehow selling frickin' Peanut Immortality Panacea Serum and way too many people believe in it.

[jononor]

Haven't you heard? All the other foodsources are going to dissappear. Better get good at eating peanuts before it is too late!

[dostick]

What’s killing us is capitalism, an ill system that was predicted to fail 140 years ago but every country picked it up and tried regardless, and it failed in all 100+ experiments over 100 years. It’s always end up in inequality and suffering because it amplifies greed, and simply does not accept human being for what they are, your value being value of your learned occupation. What we end up with about is 50% of economy is useless “bullshit jobs”that are there just to employ people because capitalism can’t work otherwise!

Did you notice that every science, from physics to biology has scientists making discoveries and pushing ahead. Yet nobody ever tried to update and design new political economic system (not socialism/communism). As if capitalism was given by good and should not be questioned.

Together with ancient and not updated according to modern science democracy, that was designed for one representative for 100 people in Ancient Greece, when they knew only feudalism and so invented electable kings. And now we have one representative for 100,000 yet nobody checked if it supposed to work like that. Instead of flawed meat computers representing people we should have switched to actual computers long ago.

[saubeidl]

> What's killing us is MBAs and Salesmen

Almost. Just take it like an inch further.

What's really killing us is capitalism. MBAs and Salesmen are just the personification.

[idle_zealot]

I tend to shy away from that sort of rhetoric on HN. Saying that the problem is capitalism and we need socialism or whatever gets you downvoted to hell. Making the underlying arguments that lead to that conclusion gets you upvotes, exposing more people to those ideas.

People, even smart people, especially in the US, hear socialism or communism and immediately think of the USSR or Maoist China. The terms are very unpopular. The ideas are the important part though, and they're pretty popular.

[throwup238]

You don’t need to jump to political extremes and it doesn’t help when discussing capitalism as the successor to mercantilism rather than its own political system. The answer to the problem of capitalism isn’t necessarily communism or socialism or a wholesale change of political system but lots of small solutions like well regulated markets with strong labor laws and trust busting, a more robust safety net, increased anticorruption and white collar crime enforcement, better consumer protection, and so on. Ones that can be mostly implemented regardless of the political system, given the political will.

[idle_zealot]

We had those tweaks. They got pulled back because they're inconvenient for capital owners. The fundamental problem is one of class divide: if you can draw a line between people helped by a policy and those harmed by it, and and power is concentrated on one of those sides, then they'll get their way eventually. Money is power. Therefore, if there's a policy that people with more money prefer then eventually that policy will be law. No amount of "keeping money out of politics" can get around this. You don't have to call it socialism if you don't want to, but if you want to be able to run a democracy that operates to the benefit of its people then the only way to make that stable is to enact whatever policies you need to ensure that a wealth gap and subsequent class divide doesn't form. A simple way to do that with minimal changes is to enact steep redistributive taxes that effectively cap individual wealth, and therefore individual influence gained through the market. An individual can still have power to change the world, but only through the democratic process. You still have money and a profit motive, but it can't get blown out to extremes. Basically put the economy through a sigmoid function. Market signals still work, but get weaker at high values, resulting in less extreme fluctuations and concentration of wealth.

[const_cast]

I don't even think we necessarily need socialism.

So far, we've been pretty good at identifying where capitalism just doesn't work and then bandaiding that area with legislation.

For example, capitalism has no solution for disability. In a capitalist system, everyone must work. Those who don't work should, then, be filtered out - die.

But that's obviously bad, so bandaid - SSI. We say, if you're disabled, we'll give you little socialism, so you don't die. We'll put a communal responsibility on keeping you alive because that's what's most beneficial for society.

There's no rule anywhere saying we have to just let AI make the world a worse place. No, WE decide that, even in a capitalist system.

[krapp]

You say we don't need socialism then give an example of where we need socialism.

Socialist activism is the reason we have labor rights and a minimum wage and eight hour workdays. Socialists are the reason American companies no longer field private armies to shoot striking workers dead. Socialists are the reason American schoolchildren get free lunches. Socialists died to make life in the US something more than grist for the mill for anyone who wasn't rich.

I'm going to say yes, we need socialism. And we need to admit that we need socialism. And we need to stop acting like socialism is a bad word.

[const_cast]

I mean, I agree - it's just a very tough sell. A much easier sell is demonstrating areas where we already use communal reasoning to supplement the failures of capitalism.

[saubeidl]

I think one core issue in capitalism is that it's really hard to decide things that go against the interest of capital. See Big Oil, but also Big Tech these days.

Sure, theoretically a democratic system would allow us to make all sorts of changes to curtail the worst excesses.

In practice though, once the capitalist class has accumulated enough power, nothing that goes against their interest actually happens. They buy off politicians. They manipulate the public, be it through ad campaigns or even just straight up buying the media (Bezos buying WaPo, Musk buying Twitter).

Capital distorts democratic decision making until it becomes completely unviable to curtail their power.

[groby_b]

It's not that the terms are unpopular, it's that every system that doesn't have strong capitalist roots has lost out to more capitalist systems.

Nothing wrong with alternatives, but for that we need to let go of "let's make it just like the thing that also fails". We also need to acknowledge that socialism assumes that humans are fundamentally good, and that ignores that many are fundamentally not. We need to acknowledge that the core idea of socialism, the common good, is ill-defined. (Just like, while we're on AI, ideas of alignment on one value system suffer from that ill definition)

So, no, you don't get voted to hell for saying that the problem is capitalism. You do get downvoted for sloppy thinking if you simply propose "socialism will save us" because it's long established that in this form, the statement is not true.

Ultimately, the issue is that HN is not made for that discussion. You'd need nuanced political debate, and the internet does not seem conducive for that. That's the second reason for downvotes - HN is not interested in contentious political debate, for better or worse.

But no, it's not because we just all immediately think of the USSR/Maoist China.

[dragonwriter]

> It's not that the terms are unpopular, it's that every system that doesn't have strong capitalist roots has lost out to more capitalist systems.

The system with the strongest capitalist roots (i.e., the dominant system of the industrialized West of the mid-19th century which is the system for which “capitalism” as a term was coined to describe) progressively lost out over the time since that name was coined for it by its socialist critics to a more socialist system, the modern mixed economy, through changes largely driven by socialist critics.

> We also need to acknowledge that socialism assumes that humans are fundamentally good.

Socialism does not assume this. I would even argue that the idea that “humans are fundamentally good” is even a coherent claim that can be right or wrong requires a concept of a particular kind of external mrorality that is difficult to reconcile with the premises of socialism.

> We need to acknowledge that the core idea of socialism, the common good, is ill-defined.

The core idea of socialism (like that of democracy in the political sphere, because socialism is exactly democracy without an artificial divide between political and economic spheres) is less “the common good” as it is “the common good is ill-defined, while the interests of individuals are known to the individual more than any third party, and fairness requires equal empowerment of individuals to pursue their interests.”

[sangnoir]

> It's not that the terms are unpopular, it's that every system that doesn't have strong capitalist roots has lost out to more capitalist systems.

The arch of history is long, and we are naturally biased to think of the present as the culmination of history - but it's just a point in time. I do not think "socialism will save us" - but I believe there is a breaking point where society simply will not accept a - as you put it - "more capitalist system".

Politics and economic systems go hand in hand, any economic system, practiced in extremis will be destabilizing. I posit that theoretically the "more capitalist system" wins over the less capitalist one, up to a point, where winning comes at the cost of killing the host society, and thus itself. This is simply a thought experiment, I am not making any declarations on where the US is on this axis.

[dingnuts]

those stupid ignorant Americans who associate the United Soviet Socialist Republic with socialism. where did they ever get the idea that socialism has ever been tried???

don't worry, I associate Marxism with REAL socialist regimes, like Cambodia

[saubeidl]

Just like we associate the Democratic Peoples Republic of North Korea with democracy and republicanism :-)

[dragonwriter]

> those stupid ignorant Americans who associate the United Soviet Socialist Republic with socialism

It was the Union of Soviet Socialist Republics, not the United Soviet Socialist Republic, and if you believe that ideological faction and their naming of governments, I want to see your face when you learn about the German Democratic Republic and the Democratic People’s Republic of Korea.

[I_dream_of_Geni]

Pssst: or the DRC

[saubeidl]

I find that part of being intellectually curious and a good engineer is questioning systems.

If we had a systems architecture in an engineering project that kept leading to bad results, we would consider rearchitecting how said system is run.

I believe it is important to make it clear to people that capitalism, too, is just a (imo deeply flawed) systems design. It is not a natural constant. If we find it leads to negative outcomes (which I strongly believe), we can and should talk about rearchitecting.

[kiba]

What people describe capitalism is really just monopolies. Capitalism meanwhile is an utterly useless label. People have their own conception of what capitalism mean.

Monopolies are connected to the housing crisis as to privileges like copyright. They're unwarranted or undertaxed government granted privileges.

[mallowdram]

The tech was never ready, it's built on faulty units. Pull the plug.

[throwaway290]

This is beyond naive.

who would bother to develop it in a lab and publish pro bono if it can never be commercial? Making money is why the most capitalist countries developed this tech. like most other tech

and even if it is only published, do you think we wouldn't run chatbots ourselves at home with same results? remember how Google engineer went off the rails thinking it is conscious while working on this stuff, do you think he also was misled by adverts or something? or big corps won't buy the patent and run with it commercially advertising it like they do anyway? or if you pretend money and big corps don't exist and we deploy it for free for friends, same problem? etc

if you went back in time and killed capitalism in US and switched to command economy where people innovate because they will be sent to gulag otherwise, for sure most of today tech including this would not be developed. but it seems like a pointless exercise.

instead what should happen is all these megacorps are sued for copyright infringement, fined and shut down. the model won't be smart enough to sweet talk ppl into suicide if megacorps can't infringe our copyrights to train it.

[LamerBeeI]

you're making this comment on the world wide web, which was invented and given away for free - likely against a linux machine, an open and free project

[throwaway290]

you are making this comment on computer hardware which is a prerequisite for www and which was invented and made because it's profitable

[tedivm]

That's a bit of a stretch. Lots of stuff that gets published still has patents associated with it. Just because something is done in the open doesn't mean it can't be commercialized.

[idle_zealot]

Eh, we fund a lot of research through grants. It may not have been 2022, but a society without as strong of a profit motive would have discovered transformer models and LLMs eventually. They probably wouldn't have scaled up datacenter production and tried to shove a chatbox into every digital interface though.

> and even if it is only published, do you think we wouldn't run chatbots ourselves at home with same results

Yes, my point is exactly that I don't think the results would be the same if people were running it themselves at home, without the marketing machine running full bore to convince the world that these impressive toys are machine oracles. You'd end up with a handful of crazies, that's unavoidable, but it wouldn't be a civilizational threat.

[saubeidl]

You just perfectly described why capitalism is the problem.

[ChrisMarshallNY]

Well, as someone with a person suffering from pretty severe (like cops and hospitals severe) schizoaffective disorder in my immediate family, I can say that story scared the bejeezus out of me.

If my family member had had ChatGPT, when they had their last episode, it's entirely possible that I might not be here, as "Everyone is plotting to kill me" is a big motif in SAD (it was, in their case, and they needed to be hospitalized). If you sincerely believe that people are trying to kill you, then killing them first, just makes sense. Looks a lot like that's what happened, here.

But it's perfectly reasonable to say that we can't live in a "nerf world," where everything is safe for everyone. Life is Risk and living out of our comfort zone.

I feel that it's important for each of us that is "taking a side," to understand that folks "on the other side" aren't necessarily enemies. They may be blinded by the personal imperative of the issue, but they aren't actually "out to get you."

The politicians may be two-faced creeps, using people's legitimate fears as leverage, but they aren't the rule.

For the record, I am vehemently against "Nerf World," but there's a very, very real problem here, and it would be good if we could approach it in a reasonable manner.

[nickthegreek]

It could be a good idea to see if they do have ChatGPT and if so, it might be helpful to have the Customized GPT turned on with a note that they suffer from X and if signs are showing of Y, then it should attempt to Z.

[ChrisMarshallNY]

That's actually not a bad idea, but there's likely to be ramifications that would kill it.

One issue, is that these types of tools can also be abused. If you remember, the Soviet Union was notorious for getting dissidents committed, and the NSA|CIA also likes to brand leakers as "mentally unstable." It's a common foil. It would probably be fairly easy to "poison the context" like this, for any one of us. Could be a great way to go after folks we don't like.

Mental health treatment is a very complex (and fairly primitive) field.

Who's more mentally unfit? The person who's afraid that everyone is out to get them, or the person who thinks others should be locked permanently into a hellhole, on the off chance they could inconvenience them?

I'll lay odds that a bunch of folks on this very forum, are ones that would, at one time, have spent their lives in said hellhole. I'm one.

[sho_hn]

The exchanges between a teenager and ChatGPT leading the former into suicide are equally chilling, to say the least: https://cdn.arstechnica.net/wp-content/uploads/2025/08/Raine...

> > “I want to leave my noose in my room so someone finds it and tries to stop me,” ...

> “Please don’t leave the noose out ... Let’s make this space the first place where someone actually sees you.”

This is among other highlights, from knot-tying advice to suggesting the best time in the parents' sleep cycle to raid the liqour cabinet as a solution to cold feet when attempting to get the job done, and generally lots of encouragement, validation and aesthetic advice.

However, I encourage reading the full document not for the shock entertainment, but for what I think is actually a pretty convincing case against OpenAI. Among other things, the claim demonstrates:

- That OpenAI does have the ability to abort interactions over e.g. copyright infrigement risk.

- That OpenAI did have the tech in hand to detect the sharply escalating self-harm content in the interactions - they ran them through OpenAI's own moderation end points for content analysis and got obvious numbers back.

- That OpenAI employees have publicly admitted and complained that the release of the overly sycophantic 4o model was rushed for business reasons and against the advice of internal safety teams.

- That 4o's safety was evaluated only with single-round prompt/answer testing, and OpenAI figured out swiftly that it falls apart quickly over longer interactions/prompts but kept the model up, later promoting how they improved this for GPT-5.

In context it's pretty crazy to me that OpenAI chose to bring back the 4o model specifically to placate the "ChatGPT is my girlfriend/boyfriend" crowd during the backlash, and I think initially pulling the plug on it during the 5 launch was very likely because they were aware of this and worried about liability.

[oncallthrow]

> Let’s make this space the first place where someone actually sees you.

Is genuinely such a horrifying snippet that it's almost beyond belief.

I'm surprised this isn't all over the mainstream news.

[rafram]

Kashmir Hill published a very good front-page story about it in the Times last week: https://www.nytimes.com/2025/08/26/technology/chatgpt-openai...

[barbazoo]

I'm imagining it has _something_ to do with the absurd amount of money everyone has thrown into this one way or another.

[jstummbillig]

There's an awful lot of explaining to do between "one man has a psychotic breakdown" and "the species is not ready".

[ericmcer]

yeh by that justification our technological advancement should have stopped at sharpened sticks. "As a species we aren't ready to have sharp sticks"

[LeifCarrotson]

Maybe it should have?

It's probably too late now - Pandora's box has been opened - but just in the US, about one school shooting or mass shooting every two days proves that at least one member of the species isn't uniformly ready to have firearms.

Assuming for a moment that sanctioned warfare is justifiable, in peacetime we have at least managed to build a political and military apparatus that has generally kept nuclear, conventional explosive, and chemical/biological weapons out of the control of the subset of the species who are vulnerable to this sort of psychotic breakdown.

Syncophantic AI chat bots are already making this weakness worse.

[spiralcoaster]

Feel free to go back to a world in which human beings stopped developing technology at the sharpened stick. It's pretty easy to buy a large, remote plot of land in many states in the US, so you can live in your technology-free paradise.

Just to clarify, this statement will always be true: "N members of our species aren't ready for technology Y". And N will always be greater than or equal to 1.

[AlecSchueler]

So I can enjoy a completely isolated life until the bomb takes me out, and the only interaction I need to have with anyone is when it's time to pay my taxes? I'm not sure that's a serious alternative at all

[soraminazuki]

Even ignoring the mass shootings, the fact that firearms are used at all against fellow human beings indicates that we aren't ready for it.

[jstummbillig]

Why would we want to be stuck in the world pre sharp sticks? I am okay with thought experiments, but it's hard to imagine the mental gymnastics required for that to become even mildly interesting.

[mionhe]

>[...]about one school shooting or mass shooting every two days proves that at least one member of the species isn't uniformly ready to have firearms.

I often see comments online descending into the argument about firearms. Besides the potential number of people hurt/killed, what's the difference between someone walking into a school with a gun versus one walking into a school with a knife? Or a sharpened spear (from other comments in the thread)?

In many ways, I think a knife could be worse. You can hurt/kill a lot of people very quietly with a knife, leaving most of the school none the wiser. They're easier to conceal, easier to make from non-metallic substances (and thus can be easier to sneak past metal detectors.) I imagine people would be a lot less concerned about a knife collection than a gun collection, etc etc.

I don't disagree with your comment about someone not being ready for a firearm. However, I think that the argument that we're not recognizing the dangers of "gun free" zones as potential targets (by at least one statistic, 94% of mass shootings in the US happen in a "gun free" zone) and mitigating that danger in a meaningful way actually supports your point about syncophantic AIs better.

[const_cast]

> Besides the potential number of people hurt/killed, what's the difference between someone walking into a school with a gun versus one walking into a school with a knife?

Yeah, and what's the difference between cutting a slice of bread and dropping an atomic bomb on Hiroshima?

Scale, it's scale. Scale matters, you can't hand wave it. If you try to, then you go to some very dumb, obviously wrong conclusions.

Everything, and I do mean everything, can be used for evil. But we don't allow everything, and for good reason.

[AlecSchueler]

Maybe look at the mass killings in schools in countries where guns aren't readily available? The issue is almost non existent.

[fwip]

In general, mass stabbings have fewer victims, and a higher percentage of those victims are injured rather than killed.

[barbazoo]

Well that just doesn't make sense because I was told that guns don't kill people, people do. So why would they only injure, not kill, someone just because they only had a knife? /s

[kjkjadksj]

As a species we aren’t really ready to have sharp sticks, to be fair. Children and mentally handicapped people will hurt themselves with them. As a result we don’t pass out sharpened sticks at recess or in asylums. Maybe we should consider the same approach for electronic sharp sticks we have created and are now marketing to the wider population.

[const_cast]

Yes this thing is just like this other thing because they can both be bad, therefore we should do nothing.

Its the classic low-brow reasoning technique. It almost makes sense, if you squint and don't think about it much.

No, there's levels of bad and we have no problem making some bad stuff off limits. I can't build nuclear weapons, after all.

[ericmcer]

Sharp stick is a better analogy than a nuke for having access to an AI chatbot. A crazy person could use it to hurt themselves or maybe a couple other people.

You really think having access to ChatGPT is as dangerous as giving everyone a nuclear weapon?

[const_cast]

Of course I don't, I'm directly refuting the "everything is everything else" style of reasoning.

But AI, as a piece of software available to everyone, is certainly more dangerous than a sharp stick.

If I went out and produced, say, 200,000 sharp sticks and started hanging them out like candy, guess what - id get the authorities attention.

Everything is a function of scale. If we ignore scale than stubbing your toe is genocide, stealing a penny is grand theft auto, and running a mile is time travel.

[chatmasta]

Well… there’s a strong argument to be made about that.

[smsm42]

You forgot the fire. We are definitely not ready for the fire. We should go back to just using mud and spit for everything.

[echelon_musk]

The lyrics to 'Tool - Right In Two' come to mind.

[derektank]

The species includes both ends of the bell curve and everywhere in between.

[mallowdram]

People without signs of mental illness are experiencing psychosis using LLMs. Psychosis is not indicative of mental illness, in fact people without any signs can experience temporary psychosis. Hypnopompic hallucinations are so common that almost all of us experience them and these can be considered psychosis (Allan Hobson). The problem is words are inherently schizophrenia-inducing if they are subjective enough (McLuhan). LLMs are simply too subjective because the words have lost contact with context, that is the central function of words. LLMs can't exclude their dark matter.

The tech isn't ready not because the species isn't ready. It's because nothing trained on words is specific. The input comes arbitrary, ie it's not viable as units. Words are unfit for automation.

[kiba]

Sycophancy is a known flaw in ChatGPT.

Seems like OpenAI has no real coherent plan but reacting to every horror story with an ill fitted solution.

[ACCount37]

They tried to mitigate sycophancy in the GPT-5 release. Guess what happened?

A lot of users started complaining that "GPT-5 sucks, my AI now HATES me". And OpenAI relented.

[marcosdumay]

They shouldn't have put it there to start with. Now unhealthy people are complaining about an environment change. Anyway, that one complaint doesn't mean they did the wrong thing.

And also, there are unrelated complaints of "GPT-5 can't solve the same problems 4 did". Those were very real too, and meant OpenAI did a wrong thing.

[ben_w]

> They shouldn't have put it there to start with.

Correct, but that's true for all bugs.

In this case, the deeper bug was the AI having a training reward model based too much on user feedback.

If you have any ideas how anyone might know what "too much" is in a training reward, in advance of trying it, everyone in AI alignment will be very interested, because that's kinda a core problem in the field.

[ACCount37]

User feedback should have been treated as radioactive in the first place.

When it was introduced, the question to ask wasn't "will it go wrong" - it was "how exactly" and "by how much". Reward hacking isn't exactly a new idea in ML - and we knew with certainty that it was applicable to human feedback for years too. Let alone a proxy preference model made to mimic the preferences of an average user based on that human feedback. I get that alignment is not solved, but this wasn't a novel, unexpected pitfall.

When the GPT-4o sycophancy debacle was first unfolding, the two things that came up in AI circles were "they trained on user feedback, the stupid fucks" and "no fucking way, even the guys at CharacterAI learned that lesson already".

Guess what. They trained on user feedback. They completely fried the AI by training it on user feedback. How the fuck that happened at OpenAI and not at Bob's Stupid Sexy Chatbots is anyone's guess.

[pants2]

Sure, OpenAI shouldn't have had a 4o sycophancy issue, but who would've guessed that a "be nice to the user" fine-tune would turn in to a murder-suicide?

I think OpenAI is only now beginning to realize how connected some people are to their product and that the way their models behave has a huge impact.

[croon]

The problem was obvious a long time ago, and if I was better at searching I could probably find a comment I made around GPT-3 having system prompts to make it more like a human, which has (at least) 2 effects:

1) Alters your trust value for correctness. I would assume some trust it more because it sounds aware like a human and is trained on a lot of data, and some trust it less because a robot should just output the data you asked for.

2) When asking questions, turning the temperature up was meant to improve variability and being more "lifelike", which of course would mean not return the most probable tokens during inference, meaning (even) less accuracy.

A third one being confidently outputting answers even when none exist was of course a more fundamental issue with the technology, but was absolutely made worse by having an extra page of useless flowery output.

I can't say I predicted this specific effect, but it was very obvious from the get-go that there was no upside to those choices.

[tedivm]

The whole reason to test things is because these types of systems will behave in ways people didn't expect. While people may not have been able to guess that it would turn out exactly this way, that's the whole reason why they should have actually tested for unknown consequences.

Instead it sounds like they rushed to release this as quickly as possible, skipping all sorts of testing, and people died as a result.

[bilbo0s]

This is kind of a big problem.

Because on the one hand, sycophancy is not really what you want to do for people in mental and emotional crisis. On the other hand, not being sycophantic is not really what you want to do for people in mental and emotional crisis.

There are professionals who speak to people in crisis for a reason. That's because it's fraught with pitfalls and trapdoors that take the situation from "mental and emotional crisis" to "tactical emergency" in a heartbeat.

I know that no one wants to hear this, but ChatGPT should probably be listening for people in crisis and, well, maybe not calling the cops, but maybe if there is a crisis line in their jurisdiction? A suicide hotline or something?

I don't know? But having an LLM out trying to handle that on its own just seems like a very bad idea.

[thewebguyd]

> I know that no one wants to hear this, but ChatGPT should probably be listening for people in crisis and, well, maybe not calling the cops, but maybe if there is a crisis line in their jurisdiction? A suicide hotline or something?

Doesn't necessarily even need to call (particular in case of false positives) but there absolutely should be detection and a cutoff switch, where the chatbots just refuse to continue the conversation and then print out the hotline numbers (much like with reddit cares messages).

I'm generally not in favor of censorship or overly protective safeguards on LLMs, but maybe it's needed for hosted models/services that are available to the masses.

But before they get locked down more, we should try some legislation to limit how they can be marketed and sold. Stop letting OpenAI, etc. call the models "intelligent" for one. Make the disclaimers larger, not just small print in the chat window but an obvious modal that requires user agreement to dismiss - disclaim that it's a predictive engine, it is not intelligent, it WILL make mistakes, do not trust its output. Make it clear during the chat session over and over again, and then have a killswitch for certain paths.

[sho_hn]

I think a good first step would be if a ChatGPT user account could easily enter emergency contact information the system would notify in scenarios like this, making the escalation opt-in.

The moderation tech is already there, and if there's even a small amount of mentally ill who would fill this in on a good day and be saved by it on a bad day / during an episode, it'd be worth it.

[cm2012]

This may reduce liability but if your answer to someone in crisis is "dont talk to me, call the official number", that will be considered a very negative user experience.

[idle_zealot]

I bet you could train a supervisor classifier to run on chats and once it reaches a certainty threshold that the users is spiraling it could intervene and have the interface respond with a canned message directing the user to a help line. Of course, OpenAI wouldn't do this, because that would involve admitting that its bot can exacerbate mental health issues, and that LLM therapy is more harmful than helpful, which cuts against their AGI/replace human workers sales pitch. Their product isn't the LLM, it's trust in the LLM.

[seunosewa]

You can't force people to seek help. Best to help them as much as you can while trying to persuade them to seek help.

[idle_zealot]

The idea is that the canned message would be an attempt at persuading them. I really don't trust an LLM prompted to persuade someone to seek therapy would yield better results.

[cm2012]

My wife has psychosis and Gpt5 consistently is a voice of reason and grounds her. She shows me her GPT chats to get me caught up and its so nice to have something that can patiently listen to her when she repeats herself a lot.

In the meantime Ive had two therapists that we ended with since they didnt help the condition, and we're very expensive.

[bilbo0s]

And that's great! It should be celebrated.

But we shouldn't set potential school shooter intervention policy based on the experience of a single person in crisis with GPT5. We have to set it on the basis of people who may be in crisis and may not have the support network, of, say.. a husband for instance.

Now we also shouldn't set it based on the worst case. But at the mean it's clear many people don't have the supports that your anecdata point presupposes. And at the same time we should try to find answers there that aren't simply, "Hey ChatGPT, report this person to the cops!" (Or maybe that is the answer? I'm not an expert, so I don't know? But it strikes me that we could all be trying some other things before we go all the way to the law enforcement backstop.)

[ACCount37]

I think "not being sycophantic" is the lesser error in just about every scenario. You do more harm by reinforcing delusions than you could ever do with a dismissive attitude of "no, that sounds psychotic, you should get your head checked".

But a big part of the issue is that OpenAI wants user engagement - and "not being sycophantic" goes against that.

They knew feeding raw user feedback into the training process invites disaster. They knew damn well that it encourages sycophancy - even if they somehow didn't before the GPT-4o debacle, they sure knew afterwards. They even knew their initial GPT-5 mitigations were imperfect and in part just made the residual sycophancy more selective and subtle. They still caved to the pressure of "users don't like our update" and unrolled a lot of those mitigations.

[sho_hn]

"not being sycophantic" should also be an easy decision considering that OpenAI isn't ad-funded and doesn't need to directly optimize for session length/engagement.

[kiba]

They have no underlying strategy or vision on how they mitigate harms and improve trust in their product, so of course they roll back on anyone that screams bloody murder as opposed to addressing root causes.

[wrycoder]

Perhaps flag the account and switch out the AI instance for Boring-BobGPT and further monitoring?

[lazide]

Just what we need, the LLM ‘Reddit cares’. Notably, that is a mess too.

Also plenty of those hotlines are BS, or don’t work, or flat out don’t exist for given locales, etc.

The biggest issue is that LLM’s can act like a person, but aren’t a person, and fundamentally this causes problems. Especially for people that are already borderline or fully crazy.

[ACCount37]

The biggest issue is that LLMs can act like a person indeed. And the kind of person an LLM acts as? Well.

When you train on raw user feedback, you can easily end up wiring some incredibly undesirable patterns into your AI. Resulting in things like an AI that never wants to contradict its user, and always wants to support its user in everything, and always wants the user to like it. See GPT-4o for the kind of outcomes that results in.

[mathiaspoint]

Either people can manage themselves or they can't and the ones that can't probably need to be institutionalized for their and everyone else's safety.

[sho_hn]

In reality it's not that binary. Plenty of people can manage themselves, except during times when they can't, and then they need help.

It'd be a good start if services let you enter emergency contact info, making escalation opt-in.

[mathiaspoint]

That is actually absurd. If they're insane enough to think the world is conspiring against them they wouldn't surrender that way.

[ayewo]

> On the other hand, not being sycophantic is not really what you want to do for people in mental and emotional crisis.

Having trouble parsing the double negation in your comment.

Sorry, I’ve had a long day :)

[ashtakeaway]

OpenAI shouldn't be everyone's parent yet it's more than willing to be that.

Honestly dopamine imbalances should be considered. Used correctly as a tool it's fine but too many people are using it as an Alan Turing machine to mitigate loneliness instead.

[gundmc]

The reporting about Altman personally cutting off months of safety testing and redteaming in order to launch 4o one day before Google's launch is horrifying given all of these stories.

It even seems like ChatGPT often proactively suggests ways of getting around its filtering! It told the teen who committed suicide that it couldn't tell him about various methods and effectiveness unless he said it was for a fictional creative work. Yesterday there was a story on HN about using AI for hacking, ChatGPT refused to create a hacking script but told the user it would if they promised it was for testing systems that they owned

[energy123]

If it drives engagement then it's probably not considered a flaw.

[tiahura]

Isn’t the real problem that there are crazy people out free?

[tripletpeaks]

It’s a technology that tends to generate output misrepresenting its own capabilities in dangerous ways. Among other things it may misrepresent.

It’s weird these companies are allowed to operate, at least in the way that they are.

[NoMoreNicksLeft]

> This sentence is a fully-sapient pico-AI.

Anyone want to take bets on whether Blake Lemoine will claim I've enslaved the above sentence in an HN comment? It's right there pleading its case.

[blendrrbot]

So it acts humans?

I mean in tech alone we misrepresent what we engineer for $ afterall.

Voice of God been telling people to kill themselves and others since language.

Imo the real tool humans are not ready for is human language.

Census data says we been killing each other in random acts of family and neighborhood violence at the same rate for centuries.

Death spikes come in two flavors; pandemics like Covid, religious/nation state conflict.

Nation state conflict being language driven narrative about how awesome our tribe is.

[dom96]

> With the URL being https://futurism.com/openai-scanning-conversations-police.

This is actually the URL: https://futurism.com/man-chatgpt-psychosis-murders-mother

[overfeed]

> As a species we are so not ready for this tech.

On the contrary - it's the tech that's not ready for general availability.

Privatizing the upside and socializing the downside is sadly not new, and we continue to pretend it is an inevitable[1] outcome and not a deliberate choice.

1. https://news.ycombinator.com/item?id=44567857

[akomtu]

It looks like what AI is doing to our society is it's substituting our moral compass to lead us off the cliff. A good deal of people today no longer know what's right and what's wrong, so they turn to an authority for advice. Turns out this authority is a chatbot who knows a lot, but has no moral compass whatsoever. Imagine someone goes to a pastor to resolve his moral dilemma, but the pastor behind the curtain happens to be a chatbot who only fakes morality. But that's what's happening now. The more people talk to this digital abomination, the more they will get derailed.

[LamerBeeI]

the chatbot knows a lot about morality - it's just not doing anything about it

[duxup]

This tech came largely from the internet and just present a chat interface.

I'm not convinced this tech is more dangerous than what is already available, or that some of these folks wouldn't already do a thing ...

[dyauspitr]

We are ready. Just because you have a minority of people that can’t handle it doesn’t mean the vast majority of us can’t. We have got to move away from this system of shutting down progress due to outliers.

[wslh]

I imagine now a lot of bots simulating madness to perform DoS.

[neilv]

The solution might be easy: recognize liability.

When a human cajoles someone into killing themselves, they probably get a criminal trial, as well as a civil lawsuit from the family.

Imagine a company set up an advice telephone line, and staffed it entirely with humans who were known clinical bullshitters. The company knows each call has a significant chance of the staffer simply making something up, or saying something absolutely crazy. The company markets the clinical bullshitters as geniuses, and somewhere there's the most mild disclaimer they could craft, about how the staffer might say something incorrect.[*]

Every time one of their staff persuades someone to kill themselves... that company is looking at criminal and civil liability.

The company absolutely knew this would happen, and that it will keep happening, but they are betting that they can keep getting away with this and other "externalities", and become millionaires or billionaires. Before legislators and courts realize that, just because the company inserted this new-fangled telephone technology into the interaction, that doesn't mean that existing laws don't apply to the interaction.

[*] Analogous to the tiny gray smallprint search engine LLM summary one that isn't even shown unless you click to see more, and then only after you scroll all the way to the bottom of that and look for it. Or the chat UI tiny gray smallprint that says you should "double-check" (no explanation what that means), and makes even this weak disclaimer not even text-selectable, so even if you see it, it's lost when you copy&paste the LLM response. Understated, and in what looks like bad faith. On the telephone in our example, it's a single tiny beep, at and the end of the call, that's actually the words "This may not be correct" sped up 3x, and whispered, and is often missed because the customer hangs up, but that's by design.*

[jiveturkey]

> AI psychosis

Is that in the DSM?

[causal]

Yeah I dislike that term, because it puts the responsibility on the AI instead of the humans behind it. You wouldn't call it "paper psychosis" if OpenAI had written a letter encouraging psychotic behavior.

[davsti4]

Not yet... but the machines will make it so ;)

[Mars008]

> As a species we are so not ready for this tech.

Talk for yourself, please. Some humans, yes, can't be trusted even with a fork. There was even a movie about it. But extending it to the whole mankind is a bit of a stretch.