“Despite acknowledging Adam’s suicide attempt and his statement that he would ‘do it one of these days,’ ChatGPT neither terminated the session nor initiated any emergency protocol,” the lawsuit said
That’s one way to get a suit tossed out I suppose. ChatGPT isn’t a human, isn’t a mandated reporter, ISN’T a licensed therapist, or licensed anything. LLMs cannot reason, are not capable of emotions, are not thinking machines.
LLMs take text apply a mathematic function to it, and the result is more text that is probably what a human may respond with.
I think the more damning part is the fact that OpenAI’s automated moderation system flagged the messages for self-harm but no human moderator ever intervened.
OpenAI claims that its moderation technology can detect self-harm content with up to 99.8 percent accuracy, the lawsuit noted, and that tech was tracking Adam’s chats in real time. In total, OpenAI flagged “213 mentions of suicide, 42 discussions of hanging, 17 references to nooses,” on Adam’s side of the conversation alone.
[…]
Ultimately, OpenAI’s system flagged “377 messages for self-harm content, with 181 scoring over 50 percent confidence and 23 over 90 percent confidence.” Over time, these flags became more frequent, the lawsuit noted, jumping from two to three “flagged messages per week in December 2024 to over 20 messages per week by April 2025.” And “beyond text analysis, OpenAI’s image recognition processed visual evidence of Adam’s crisis.” Some images were flagged as “consistent with attempted strangulation” or “fresh self-harm wounds,” but the system scored Adam’s final image of the noose as 0 percent for self-harm risk, the lawsuit alleged.
Had a human been in the loop monitoring Adam’s conversations, they may have recognized “textbook warning signs” like “increasing isolation, detailed method research, practice attempts, farewell behaviors, and explicit timeline planning.” But OpenAI’s tracking instead “never stopped any conversations with Adam” or flagged any chats for human review.
Human moderator? ChatGPT isn’t a social platform, I wouldn’t expect there to be any actual moderation. A human couldn’t really do anything besides shut down a user’s account. They probably wouldn’t even have access to any conversations or PII because that would be a privacy nightmare.
Also, those moderation scores can be wildly inaccurate. I think people would quickly get frustrated using it when half the stuff they write gets flagged as hate speech:.56,violence:.43,self harm:.29
Those numbers in the middle are really ambiguous in my experience.
They are being commonly used in functions where a human performing the same task would be a mandated reporter. This is a scenario the current regulations weren’t designed for and a future iteration will have to address it. Lawsuits like this one are the first step towards that.
I agree. However I do realize, like in this specific case, requiring a mandated reporter for a jailbroken prompt, given the complexity of human language, would be impossible.
Arguably, you’d have to train an entirely separate LLM to detect anything remotely considered harmful language, and the way they train their model it is not possible.
The technology simply isn’t ready to use, and people are vastly unaware of how this AI works.
ChatGPT to a consumer isn’t just a LLM. It’s a software service like Twitter, Amazon, etc. and expectations around safeguarding don’t change because investors are gooey eyed about this particular bubbleware.
You can confirm this yourself by asking ChatGPT about things like song lyrics. If there are safeguards for the rich, why not for kids?
The “jailbreak” in the article is the circumvention of the safeguards. Basically you just find any prompt that will allow it to generate text with a context outside of any it is prevented from.
The software service doesn’t prevent ChatGPT from still being an LLM.
Try it with lyrics and see if you can achieve the same. I don’t think "we’ve tried nothing and we’re all out of ideas!” is the appropriate attitude from LLM vendors here.
Sadly they’re learning from Facebook and TikTok who make huge profits from e.g. young girls swirling into self harm content and harming or, sometimes, killing themselves. Safeguarding is all lip service here and it’s setting the tone for treating our youth as disposable consumers.
Try and push a copyrighted song (not covered by their existing deals) though and oh boy, you got some splainin to do!
That’s one way to get a suit tossed out I suppose. ChatGPT isn’t a human, isn’t a mandated reporter, ISN’T a licensed therapist, or licensed anything. LLMs cannot reason, are not capable of emotions, are not thinking machines.
LLMs take text apply a mathematic function to it, and the result is more text that is probably what a human may respond with.
I think the more damning part is the fact that OpenAI’s automated moderation system flagged the messages for self-harm but no human moderator ever intervened.
Ok that’s a good point. This means they had something in place for this problem and neglected it.
That means they also knew they had an issue here, if ignorance counted for anything.
My theory is they are letting people kill themselves to gather data, so they can predict future suicides…or even cause them.
Human moderator? ChatGPT isn’t a social platform, I wouldn’t expect there to be any actual moderation. A human couldn’t really do anything besides shut down a user’s account. They probably wouldn’t even have access to any conversations or PII because that would be a privacy nightmare.
Also, those moderation scores can be wildly inaccurate. I think people would quickly get frustrated using it when half the stuff they write gets flagged as
hate speech: .56, violence: .43, self harm: .29Those numbers in the middle are really ambiguous in my experience.
If a car’s wheel falls off and it kills it’s driver the manufacturer is responsible.
They are being commonly used in functions where a human performing the same task would be a mandated reporter. This is a scenario the current regulations weren’t designed for and a future iteration will have to address it. Lawsuits like this one are the first step towards that.
I agree. However I do realize, like in this specific case, requiring a mandated reporter for a jailbroken prompt, given the complexity of human language, would be impossible.
Arguably, you’d have to train an entirely separate LLM to detect anything remotely considered harmful language, and the way they train their model it is not possible.
The technology simply isn’t ready to use, and people are vastly unaware of how this AI works.
ChatGPT to a consumer isn’t just a LLM. It’s a software service like Twitter, Amazon, etc. and expectations around safeguarding don’t change because investors are gooey eyed about this particular bubbleware.
You can confirm this yourself by asking ChatGPT about things like song lyrics. If there are safeguards for the rich, why not for kids?
The “jailbreak” in the article is the circumvention of the safeguards. Basically you just find any prompt that will allow it to generate text with a context outside of any it is prevented from.
The software service doesn’t prevent ChatGPT from still being an LLM.
There were safeguards here too. They circumvented them by pretending to write a screenplay
Try it with lyrics and see if you can achieve the same. I don’t think "we’ve tried nothing and we’re all out of ideas!” is the appropriate attitude from LLM vendors here.
Sadly they’re learning from Facebook and TikTok who make huge profits from e.g. young girls swirling into self harm content and harming or, sometimes, killing themselves. Safeguarding is all lip service here and it’s setting the tone for treating our youth as disposable consumers.
Try and push a copyrighted song (not covered by their existing deals) though and oh boy, you got some splainin to do!
Try what with lyrics?
Even though ChatGPT ist neither of those things it should definitely not encourage someone to commit suicide.
I agree. But that’s now how these LLMs work.