How can you charge someone for making a threat when you only read the threat by spying on them? Surely that has to be thrown out in court? They didn’t actually send the threat to anyone, you just obtained it by spying.
I have some sympathy for Anthropic here because I've seen the headlines after OpenAI failed to report a shooter in a similar situation. So from their perspective, it's damned-if-you-don't, damned-if-you-do.
However, people need to get it in their heads that they're not chatting with their secret BFF, they're chatting with Big Tech. Before LLMs, Big Tech had no way to scrutinize the bulk of what was going on within their services, so you could have a secret hate diary in Google Docs. Now, everything you say or write can be automatically screened for red flags on a planetary scale, and probably will be because that's what the regulators and "concerned citizens" will demand. In a couple of years, you'll be biting your tongue a lot more often in private chats.
This is one of the reasons why AI companies are looking for explicit regulations, it can help to reduce the risk of possible liability and maybe make the legal way forward more tractable. With a regulatory framework in place much of the burden of identifying risks falls on the regulatory authority.
Then the AI companies have more confidence that they can move forward in a certain way, and issue investor guidance that is maybe closer to reality.
I think they're mostly looking for regulations because they've spent close to $2T, all to realized they have no technical moat, so they're trying to build a regulatory one.
It's the only thing that makes sense for a lot of these companies to survive to any long term when open models keep nipping at their heels for pennies on the dollar.
I'm still puzzled that people's default assumption isn't that somebody is reading all of their internet communications. Ever since the Snowden revelations of 2013 I've more or less assumed that everything I type into a computer is stored in a government database somewhere. Not that I actually believe it is 100% of the time; there's a spectrum of trust, so I'm more confident that local apps on my Linux desktop are secure, somewhat confident in the end-to-end encryption of certain apps on my iPhone, but all bets are off for non-E2E encrypted data going across the internet.
People don’t want to believe the Stasi is monitoring them. They want to believe the world is a nice place just like where they grew up.
Also, unlike the bad old days when 1 in 3 was an informant, now ordinary people aren’t in “informant loop” of providing information on others, so they aren’t thinking about being informed on either.
For technical people, this is incredibly old news.
For non-technical people, it isn't really news, because they already forgot about it after reading it. Maybe they'll be a little more monitored in their own typing for like... a day or two.
One of the hardest lessons to internalize, and keep internalized, as someone who works on and writes software, is the vast, vast, vast majority of the Public doesn't understand even the most basic shit about software. It just does stuff. Hopefully the stuff is good. That's it, beginning, middle, and end.
"Why would you think x would y" is a poor framing. They didn't think about x or y because they don't care. The phone works, that's the beginning and end of their interest in the subject.
A disclaimer on the top of page every time would have been a better approach helping Anthropic and the end user.
A bot talking to you directly as if its some one real caters to your thoughts and can take you in a certain direction without you realizing it. I have heard first hand experience from people that they feel more comfortable talking to chatgpt or claude cause it gives a feeling of being on their side and listening to them.
> people need to get it in their heads that they're not chatting with their secret BFF
I see constant ads on video platform (particularly youtube/tiktok) about llm chat apps, from friends, dating, romance and everythkng inbetween; that's personal.
People need to be reminded constantly if they use such apps that they are participating in easier mass surveillance, profiling and AI training.
Problem is that even not using those apps, the apps you currently use might have turned more hostile.
It wasn't technically feasible to scan personal chats easily, other than grepping keywords which must have had a bajillion false positives. Now you can get everything autoscanned at scale.
How do you get that through to someone who doesn't even understand that mass surveillance and profiling is a problem? Or to young people who have only lived in a society of mass surveillance?
> people need to get it in their heads that they're not chatting with their secret BFF, they're chatting with Big Tech
Yup. And with zero privacy protections in statute for AI chat, there is nothing to prevent an AI CEO looking to curry political favour from e.g. handing over the private correspondence of an opponent or an entire district’s residents.
Customers who what privacy protections for AI chats are welcome to negotiate this in enterprise contracts. The major LLM vendors do offer that as an option. Customers can then enforce any violations in civil court (although this obviously wouldn't apply if the customer used the LLM for criminal purposes).
Don't know why this got downvoted? This is a problem with epistemics.
It's more efficient to have one central "verifier" for everything, but the "who watches the watchers"? question basically says: Either constrain by construction, have everyone verify (which are two sides of the same coin, btw, when looking at a "global" thing), or centralize explicitly.
And such was the case for a man who snapped a photo of his own child to send to the doctor which got uploaded to his Google photos resulting in his Google account of over a decade getting shutdown for CSAM.
As another commenter said, you're not chatting with a friend; you're chatting with Big Tech.
If the doctor was not using a gmail account, the UI probably recomended to share it "with anyone that has the link" that is like public but protected by oscurity.
Most people don't realize that it is 99% like posting it on Facebook.
I suspect it will go further: imagine giving an mp3 to LLM to clean up some noise. It detects it was illegally downloaded from youtube, deletes it and automatically fines you via attached credit card.
Unless it becomes a requirement to be licensed to be able to use any kind of ai model. You know, for safety and stuff. And of course with appropriate reporting to institutions.
Or worse: downloading a picture of pirate ship and without any concern for the copyright asking the LLM to make a coloring page for your kid. BTW Chatgpt does that way better than Claude
Hey but you’ll be allowed to file a response that will also go to an LLM and deny you automatically. And you will be charge a NSE fee (no sufficient explanation).
You should probably get a head start on waiting a couple years to bite your tongue and assume everything you type into a computer is summarized and sent to your boss, government, advertisers, political actors, insurance companies, worst enemy, etc. With phones, Alexas, and little AI tamagotchis, you probably shouldn't say much in person either.
Including any chats anywhere where someone might have a phone in their pocket, or if there's a "camera" attached to a utility pole or a nearby tree. The only real private chats might be whispered lying down in the bathtub together, with a mattress covering it like you're both hiding from a hurricane.
> they're chatting with Big Tech
They're chatting with any powerful person who wants to hear it. She thought she was chatting with Anthropic, who doesn't give a shit about her. But after being threatened (and immediately backing down because, of course, they don't give a shit about her) Anthropic has become an arm of the government. So she was chatting with the Bonita Springs, FL Sheriff's office, or anybody else. If I paid enough, Anthropic would tell me about what she was doing so I could sell her laundry detergent.
They're actively rummaging through your inputs so the damned-if-you-don't case doesn't really exist; no one expects Anthropic to not notify law enforcement once they learn of something like this. What you might have expected was some privacy in the first place though, where Anthropic would never have learned of this in the first place and where the damned-if-you-do case wasn't a thing.
Absolutely no sympathy. Anthropic is every bit as slimy as any other big corp. And like other big corps, they must open the vault to whatever governments they intent to do biz with.
Right, the difficulty is partly that they can get a negative headline from any choice of behavior.
"Anthropic failed to report murderer's threats to authorities"
(or "Chatbot knew man was planning murder, yet company did nothing")
"Anthropic reported private chats to authorities"
(or "Arrested for chatbot fantasy")
To be fair to the journalists in these cases, there's also no society-wide agreed Schelling point about the correct outcome or correct rules. I have strong beliefs and intuitions about what should happen, but other people also have strong beliefs and intuitions, and many of those are probably opposite of mine. Even if my intuitions are the best and most justified, a journalist is unlikely to think "I'm just not going to mention that some people are mad at this company over this outcome, because a hypothetically better norm or principle would support the company's actions here". Hopefully the journalism can at least contextualize the lack of legal or social consensus and the difficult incentive problems, rather than jumping to "obviously companies are sociopaths staffed by supervillains".
We really need a class, probably in high school, that works through how LLMs work at the high level (don't need to get too far into the deep math, but give people a taste) and then how they're trained, used, and deployed.
I feel like if people understood what these things actually are there'd be way less of this AI psychosis and similar stuff.
There'd also be fewer people falling for apocalyptic Rationalist delusions.
Also: people need to understand "not your computer, not your data." (Unless it's stored in the cloud but encrypted locally with keys only you possess.) Same goes for storing things unencrypted in OneDrive, Google Drive, etc. There is nothing to stop these companies from bulk scanning, data mining, or reporting people based on whatever request a government gives them. Don't count on them to resist, because they often can't, especially if the request is from a sovereign state where they do business.
You can make a reasonable case for adding a lot more classes in high school: statistics, nutrition, personal finance, etc. But ultimately it's a zero-sum game and to add a new class means removing an existing class. So what do we cut?
There'd also be fewer people falling for apocalyptic Rationalist delusions
Assuming you consider it a "delusion" to have a p(doom) of more than 5% or so, that's not uncommon among frontier lab employees who have a pretty good idea of how LLMs work.
It's a byproduct of the nannyism safety marketing from the AI companies. I'm glad these cases were caught, but disagree with how they were disposed of. If the automated flagging is good, enforce it by default. If it's noisy, refine the tech then enforce it by default. This middleground where everything going through the platforms is subject to training and arbitrary human inspection in the midst of an acrid cloud of marketing-driven fearmongering is unacceptable, and it reinforces the idea the fearmongering is legitimate.
Somehow humanity survived the past 40 years without Microsoft Word and Excel phoning home and shopping users to the feds at random, I don't see why the standard should be any different for this new class of tooling.
They have some responsibility for people's expectations of the product, at least. They want the personal assistant personas to be able to help you with anything, and don't point out that they'll be judging your thoughts along the way.
For anyone technically inclined it should be obvious, but it isn't part of the zeitgeist or how they pitch it. People see it as being different than talking to a human, and behave as if there won't be a human in the mix.
I guess we probably want our tech to work exactly like this.
It should catch normal people becoming unstable so that they can receive help. It is just highly unfortunate that the US legal system works in ways where now this woman's name is public.
___
Of course, we also want purely private tech, but that needs a certain level of merit and sanity filter.
>I guess we probably want our tech to work exactly like this.
Do we, though? What if someone started an AI company that uses end-to-end encryption to make it impossible for anyone but you to access your data? Personally, I would use it if it were competitive with the other products. I don't think it's the tech companies' job to surveil the population and prevent crimes.
Your follow-up about purely private tech seems to contradict your first statement. We can either have privacy or surveillance, not both.
I guess all the "private model" people are right. Don't want to end up in jail (or even charged with something) for asking a crazy hypothetical question or something.
Is this an invasion of their privacy (reporting to police)? Yes, but possibly warranted?
Should a social worker have contacted them rather than the police? Probably, if for no other reason than to ask if they were serious about harming someone.
Difficult questions, I'm still undecided on whether it's OK to always ignore someone's rants, even if it may be (or they think it may be) a private diary.
Actually charging them with a felony seems pretty quick to accuse. (Maybe I missed a hint about how long the investigation took before the felongy charge?)
It is my very european belief that the problem here is not that the woman was reported, but that what likely is a mental episode was made public in a way that reduces the chances of recovery.
I have told llms all kinds of stories to find out what its answers would be. I always make it sound like it is the truth to make sure the AI answers in a way that it would if somebody actually said this. I also tested internal flagging systems of the ai company I work at with the most evil things a person can ever say to find out if it would flag them.
Of course I did not mean any of that stuff, but how can you make sure a human reviewer knows you did not mean it while the llm does not know that you did not mean it.
I guess its a miracle I am not in jail yet.
Flagging people for anything said to an llm sounds wrong to me because an LLM is not a real person and while some people put in their internal thoughts, others just roleplay and the two are inseparable just from reading it.
This is exactly the typical use I make of the llm.
Adding:
- I typically ask questions in the I form, regardless for whom or why I ask for.
- Gemini chats quite often end when it starts recommending psychological council or a suicide line, to talk about my problems. It apparently detects a persistent tendency to not agree with the party line. So it makes sense I must be suicidal ;-
But sure, as llm's start to babysit us, and know our inner dialog better than anyone else, we'll soon be debugging their opinion/behavior/co-existence/authority, when it comes to reporting people to the authorities, or taking on tasks in society in general. We'll hire doctors to cure our psychological profile from our record (Total Recall).
A Minority Report like this shouldn't cause a referral to the police.
> Heller faces a charge of making a written threat of violence under Florida law. Florida Statute 836.10 makes it a second-degree felony to send, post, or transmit a written or electronic record threatening to kill or injure someone, carry out a mass shooting, or commit an act of terrorism.
She didn't threaten anything, she wrote down that she was going to do it. A "threat" is more than a mere statement, especially when written in what is described as a "diary".
> A Florida woman is facing felony charges after she used Claude as a diary and allegedly wrote that she planned to "shoot up" the Sheriff's office.
Obviously I don't want anyone to shoot up anything, but this seems like a weak case legally speaking.
I think it's fair to say this is a gray area. Clearly it was transmitted.
I can certainly threaten you harm and send it to not-you and you're still clearly in danger even if it wasnt transmitted to you. So the question becomes did she transmit it to someone? Clearly yes she transmitted it to Anthropic. But she clearly intended to send it to Claude, an inanimate object.
Claude's terms of service makes it very clear that their employees will read messages[0] to determine that they don't contain the things that these messages contained[1].
> Review is needed to enforce our Usage Policy... designated members of our Trust & Safety team may access this data on a need-to-know basis as a part of their evaluation process.
A sibling comment includes an important rider to the provision: "...in any manner in which it may be viewed by another person." If you wrote this in a google doc, it almost certainly would not qualify as a threat under this statute. Even though google docs, like LLM chats, have administrative override and you could look at their contents - you would not expect either to be "viewed by another person."
IMO I do not think this is a grey area and it's legal to tell a LLM you want to kill someone. It's certainly not a "threat" like you might send to another person, though it may end up being evidence of conspiracy or premeditation. I suspect we would be well served to, after a few years of experience, put together some laws governing when LLM chats must be made available to authorities.
It is very interesting that the LLM responses to these lines - the context around what she is saying - is not in the article. I suspect, as is the case in many instances where LLMs are involved in violent planning, that the LLM was urging this behavior on. Basically entrapment - you are encouraged by a robot to become more violent and vindictive and then when you do you are handed over to police.
Too broad. Unless you're transferring ink from a typewriter ribbon onto paper in a hut with no electricity, your words, or my words as I type this, are being grammar checked by something partly in the cloud. If I delete my words, are you saying I've transmitted them nevertheless?
Yes but the law usually evaluates the application of a statute within the context of someone's mental state. This is why you are not guilty of battery when you trip and accidentally bump into someone. https://en.wikipedia.org/wiki/Mens_rea
That depends on the crime; several related crimes are only distinguished by intent. Negligence is itself a crime if it is the cause of a preventable death when the person has a reasonable obligation, such as when driving a vehicle.
I'm not really sure that this can be likened to a diary when it is called a "chat" but that's for the legal system to determine, not me sitting on my couch.
There was no intent for a human to read the message. By your logic, if she wrote a threat in a diary and a burglar broke in and read it, it would be a crime on her part.
Yeah.. if you write a personal note and it's backed up by the operating system, it appears to be in violation of this law as well (since the company could theoretically read it)
People assume there won't be another human in the mix, but there is. She was judged for what she probably assumed was a private thought when it was actually not private.
I haven't seen anybody claiming this for consumer/free accounts anywhere.
You get privacy if you're a big corporation that needs to make sure OpenAI/Google/Anthropic can't read your trade secrets etc.
But those contractual privacy protections have been in place for a long time. It doesn't have anything to do with AI, it's been the same with Office365, Google Docs, etc.
It is unlawful for any person to send, post, or transmit, or procure the sending, posting, or transmission of, a writing or other record, including an electronic record, in any manner in which it may be viewed by another person, when in such writing or record the person makes a threat to: (a) Kill or to do bodily harm to another person; or (b) Conduct a mass shooting or an act of terrorism.
To me, that is the more interesting legal question. Does a LLM-based safety net that sends content to a human, when the original use case would not have sent it to a human, count as "may be viewed by another person". It certainly wasn't intended to be, and that isn't the norm. At the same time, because no security is perfect, we could say that any digital record, stored in any way "may be viewed by another person."
The usual thing that makes laws against criminal conspiracy pass First Amendment muster is that the words have to be combined with some concrete acts furthering the criminal conspiracy. That might just be something as otherwise innocuous as looking up the blueprints of the bank you talked about robbing but it has to be something other than just talk.
I suppose since Anthropic's T&Cs allow them to have a person read your chats, that makes it violate the law. Of course, if Anthropic didn't have that in their T&Cs, it wouldn't have been illegal to write.
For the sake of argument what would happen if she had kept the diary locally and claude code scanned the file?
What would happen if it had scanned a file it didn’t have permission to look at and found this threat?
I honestly don’t know how I feel about this. On the one hand if you’re using claude as a diary you have no expectation of privacy and she was talking about committing a very serious crime.
Over in Europe they want to read all ofd our private messages, yet these chatbots, pretending to be our friends, will snitch on us just for our thoughts.
I once had a copy of 1984 in my checked baggage returning home to the USA and when I was unpacking the bag at home, the book had a notice inside the book that my bags had been inspected by TSA...
I reckon they were just checking for money or hidden compartments. Sending books abroad packed with money is extremely common from the US, though not sure how frequently people do that with onboard luggage.
Only tangential but when I was an undergraduate studying philosophy I had Bertrand Russell's Why I Am Not A Christian in my carry-on, and the TSA saw that and had a field day.
I can only assume you meant they had a field day celebrating how much of Russell's philosophy matched the concerns about Christianity expressed by Jefferson, Paine, etc?
In Europe they should provide the citizens with the service of their messages not reaching US servers. So basically that there is only one party reading along with them, not half the world.
Anthropic commits literal felonies by stealing millions of books and violating Copyright like it doesn't exist: no charge.
One unfortunate woman who happened to write the wrong thing in the wrong place is now having her life turned upside-down for perceived thought-crime.
To Anthropic, and all employees working there, your company's product and the result of your work is cruelty. You are enabling it and pushing it down everyone's throat. You can never again claim that you are the "ethical" AI company, for no such thing exists.
I looked into ZDR, it's basically a vapid claim with little to no due diligence or auditing. I expect most providers to cave with only a minimal amount of legal pressure.
This kind of thing will get worse with humanoid robots because they have cameras and microphones. Will they be programmed to tell on you if you break any kind of rule in front of them because their owners are terrified of being sued?
LLM is not a person but T&C probably states that a human moderator may view any of it. Her lawyer could probably argue a moderator filtering usage isn't the intent of the law, it was more about publishing / sending message for other humans and it was never clear to her that a human was reviewing her private diary. Shouldn't LLM disclose that at some point when people are feeding really personal stuff?
this feels very thought-crimey to me, but I think I'd have to actually see that chats to really decide. Like is she making plans/asking for advice? is she just talking about her feelings exactly as if it's a diary?
She is not being charged with planning, just making threats in a place a person could read - the person being employees of the company.
Basically, under this interpretation, any personal note you store in the servers of a company could qualify, even if you didn't ever imagine someone would read and as such you couldn't have thought about it as a threat
After people getting pilloried for social media posts from twenty years ago, I really hope (sensible) people will have the wherewithal to think twice about what they hand over to their chatbots.
> Heller faces a charge of making a written threat of violence under Florida law. Florida Statute 836.10 makes it a second-degree felony to send, post, or transmit a written or electronic record threatening to kill or injure someone, carry out a mass shooting, or commit an act of terrorism. The communication must be made in a manner in which another person may view it.
Thought crimes are real when you're sharing your thoughts with Claude
I wonder if "criminal intent" may be missing here given that most people probably operate under the assumption that Anthropic are not reading their messages.
I think people have a binary understanding on this. Someone is either reading their messages or not. While the reality is that all the messages are read by a machine (not unlike GMail and Outlook) and anything suspicious gets flagged up so a human can read it. This obscure the concept of "reading" as most laypeople understand it.
Summary: don't type in Claude anything you wouldn't like a human to read.
It tells you exactly what it does with your information if you ask it, including this exact scenario, and has for at least four months now (when I asked).
Yes but you are someone (i assume because you are on HN) that at least understands this point. There are A LOT of people that use it as their confidant, expecting it to be private. There is a massive education issue going on as its not in the interest of these Corpos to make you fear sharing all your details with them. Because, then you wont get Dot or whatever Anthropic comes up with and share all your personal info with it.
You're not wrong, and it may come down to this in court, but there's a difference between what people think happens, or the reasonable expectations, and what actually happens, and I think it's important to recognise that difference and why it comes about.
I don't think someone is an idiot for thinking that the information they type into their private Claude account is private. I also don't think people are idiots for thinking their phone is listening to them and giving them targeted advertising based on that. Both are reasonable deductions from their lived experiences. Both are wrong.
I think it's a reasonable point to make. Writing things down is a part of "thought" for many, including those who keep diaries/journals. If you write in a private journal you do so with the expectation that is not shared, and that wouldn't seem to break this law (with my naive reading).
I mean, at this point it's pretty well known that all tech services will be hacked by fable, anthropic themselves promised it, so writing your journal in google docs or claude or a txt document on your laptop or such is the same as releasing it publicly yeah?
If it were written on paper, and only in a room with no phones or cameras so fable couldn't hack it, then I think you wouldn't be sharing it.
I think the zero-risk approach common nowadays is insanely corrosive to democracy and freedom. If a million people fantasise about shooting the sheriff, and one ever goes on to do it, I don't believe avoiding it warrants creating an apparatus of mass surveillance. After all, if people were really serious about zero murder, the only practical solution would be to lock up everyone. Some (most?) tradeoffs have exponential costs at the limit and we/lawmakers should recognise that.
> If a million people fantasise about shooting the sheriff
I think it’s fair to pre-identify folks who fantasise about shooting anyone. It’s a small fraction of the population that looks into logistics versus making offhand comments.
However, people need to get it in their heads that they're not chatting with their secret BFF, they're chatting with Big Tech. Before LLMs, Big Tech had no way to scrutinize the bulk of what was going on within their services, so you could have a secret hate diary in Google Docs. Now, everything you say or write can be automatically screened for red flags on a planetary scale, and probably will be because that's what the regulators and "concerned citizens" will demand. In a couple of years, you'll be biting your tongue a lot more often in private chats.
Then the AI companies have more confidence that they can move forward in a certain way, and issue investor guidance that is maybe closer to reality.
Also, unlike the bad old days when 1 in 3 was an informant, now ordinary people aren’t in “informant loop” of providing information on others, so they aren’t thinking about being informed on either.
For non-technical people, it isn't really news, because they already forgot about it after reading it. Maybe they'll be a little more monitored in their own typing for like... a day or two.
One of the hardest lessons to internalize, and keep internalized, as someone who works on and writes software, is the vast, vast, vast majority of the Public doesn't understand even the most basic shit about software. It just does stuff. Hopefully the stuff is good. That's it, beginning, middle, and end.
"Why would you think x would y" is a poor framing. They didn't think about x or y because they don't care. The phone works, that's the beginning and end of their interest in the subject.
A bot talking to you directly as if its some one real caters to your thoughts and can take you in a certain direction without you realizing it. I have heard first hand experience from people that they feel more comfortable talking to chatgpt or claude cause it gives a feeling of being on their side and listening to them.
I see constant ads on video platform (particularly youtube/tiktok) about llm chat apps, from friends, dating, romance and everythkng inbetween; that's personal.
People need to be reminded constantly if they use such apps that they are participating in easier mass surveillance, profiling and AI training.
It wasn't technically feasible to scan personal chats easily, other than grepping keywords which must have had a bajillion false positives. Now you can get everything autoscanned at scale.
Yup. And with zero privacy protections in statute for AI chat, there is nothing to prevent an AI CEO looking to curry political favour from e.g. handing over the private correspondence of an opponent or an entire district’s residents.
It's more efficient to have one central "verifier" for everything, but the "who watches the watchers"? question basically says: Either constrain by construction, have everyone verify (which are two sides of the same coin, btw, when looking at a "global" thing), or centralize explicitly.
As another commenter said, you're not chatting with a friend; you're chatting with Big Tech.
How much do you love Big Brother?
Most people don't realize that it is 99% like posting it on Facebook.
As articles like these show, cloud-based LLMs don't work in your interest today.
Including any chats anywhere where someone might have a phone in their pocket, or if there's a "camera" attached to a utility pole or a nearby tree. The only real private chats might be whispered lying down in the bathtub together, with a mattress covering it like you're both hiding from a hurricane.
> they're chatting with Big Tech
They're chatting with any powerful person who wants to hear it. She thought she was chatting with Anthropic, who doesn't give a shit about her. But after being threatened (and immediately backing down because, of course, they don't give a shit about her) Anthropic has become an arm of the government. So she was chatting with the Bonita Springs, FL Sheriff's office, or anybody else. If I paid enough, Anthropic would tell me about what she was doing so I could sell her laundry detergent.
"Anthropic failed to report murderer's threats to authorities"
(or "Chatbot knew man was planning murder, yet company did nothing")
"Anthropic reported private chats to authorities"
(or "Arrested for chatbot fantasy")
To be fair to the journalists in these cases, there's also no society-wide agreed Schelling point about the correct outcome or correct rules. I have strong beliefs and intuitions about what should happen, but other people also have strong beliefs and intuitions, and many of those are probably opposite of mine. Even if my intuitions are the best and most justified, a journalist is unlikely to think "I'm just not going to mention that some people are mad at this company over this outcome, because a hypothetically better norm or principle would support the company's actions here". Hopefully the journalism can at least contextualize the lack of legal or social consensus and the difficult incentive problems, rather than jumping to "obviously companies are sociopaths staffed by supervillains".
I feel like if people understood what these things actually are there'd be way less of this AI psychosis and similar stuff.
There'd also be fewer people falling for apocalyptic Rationalist delusions.
Also: people need to understand "not your computer, not your data." (Unless it's stored in the cloud but encrypted locally with keys only you possess.) Same goes for storing things unencrypted in OneDrive, Google Drive, etc. There is nothing to stop these companies from bulk scanning, data mining, or reporting people based on whatever request a government gives them. Don't count on them to resist, because they often can't, especially if the request is from a sovereign state where they do business.
Assuming you consider it a "delusion" to have a p(doom) of more than 5% or so, that's not uncommon among frontier lab employees who have a pretty good idea of how LLMs work.
Somehow humanity survived the past 40 years without Microsoft Word and Excel phoning home and shopping users to the feds at random, I don't see why the standard should be any different for this new class of tooling.
On the other hand, they did put themselves into this position deliberately.
Show me any product, no matter how simple, that has no safety vs utility tradeoff.
For anyone technically inclined it should be obvious, but it isn't part of the zeitgeist or how they pitch it. People see it as being different than talking to a human, and behave as if there won't be a human in the mix.
It should catch normal people becoming unstable so that they can receive help. It is just highly unfortunate that the US legal system works in ways where now this woman's name is public.
___
Of course, we also want purely private tech, but that needs a certain level of merit and sanity filter.
Do we, though? What if someone started an AI company that uses end-to-end encryption to make it impossible for anyone but you to access your data? Personally, I would use it if it were competitive with the other products. I don't think it's the tech companies' job to surveil the population and prevent crimes.
Your follow-up about purely private tech seems to contradict your first statement. We can either have privacy or surveillance, not both.
Is this an invasion of their privacy (reporting to police)? Yes, but possibly warranted?
Should a social worker have contacted them rather than the police? Probably, if for no other reason than to ask if they were serious about harming someone.
Difficult questions, I'm still undecided on whether it's OK to always ignore someone's rants, even if it may be (or they think it may be) a private diary.
Actually charging them with a felony seems pretty quick to accuse. (Maybe I missed a hint about how long the investigation took before the felongy charge?)
Of course I did not mean any of that stuff, but how can you make sure a human reviewer knows you did not mean it while the llm does not know that you did not mean it.
I guess its a miracle I am not in jail yet.
Flagging people for anything said to an llm sounds wrong to me because an LLM is not a real person and while some people put in their internal thoughts, others just roleplay and the two are inseparable just from reading it.
Adding: - I typically ask questions in the I form, regardless for whom or why I ask for. - Gemini chats quite often end when it starts recommending psychological council or a suicide line, to talk about my problems. It apparently detects a persistent tendency to not agree with the party line. So it makes sense I must be suicidal ;-
But sure, as llm's start to babysit us, and know our inner dialog better than anyone else, we'll soon be debugging their opinion/behavior/co-existence/authority, when it comes to reporting people to the authorities, or taking on tasks in society in general. We'll hire doctors to cure our psychological profile from our record (Total Recall).
A Minority Report like this shouldn't cause a referral to the police.
> Heller faces a charge of making a written threat of violence under Florida law. Florida Statute 836.10 makes it a second-degree felony to send, post, or transmit a written or electronic record threatening to kill or injure someone, carry out a mass shooting, or commit an act of terrorism.
She didn't threaten anything, she wrote down that she was going to do it. A "threat" is more than a mere statement, especially when written in what is described as a "diary".
> A Florida woman is facing felony charges after she used Claude as a diary and allegedly wrote that she planned to "shoot up" the Sheriff's office.
Obviously I don't want anyone to shoot up anything, but this seems like a weak case legally speaking.
I can certainly threaten you harm and send it to not-you and you're still clearly in danger even if it wasnt transmitted to you. So the question becomes did she transmit it to someone? Clearly yes she transmitted it to Anthropic. But she clearly intended to send it to Claude, an inanimate object.
> Review is needed to enforce our Usage Policy... designated members of our Trust & Safety team may access this data on a need-to-know basis as a part of their evaluation process.
[0]: https://privacy.claude.com/en/articles/10458704-how-does-ant...
[1]: https://www.anthropic.com/legal/aup
The critical part of a "threat" is that the perpetrator takes some intentional method to deliver it.
> The communication must be made in a manner in which another person may view it.
Even 'transmitted' is too broad if you also consider iCloud backup to be a means.
IMO I do not think this is a grey area and it's legal to tell a LLM you want to kill someone. It's certainly not a "threat" like you might send to another person, though it may end up being evidence of conspiracy or premeditation. I suspect we would be well served to, after a few years of experience, put together some laws governing when LLM chats must be made available to authorities.
It is very interesting that the LLM responses to these lines - the context around what she is saying - is not in the article. I suspect, as is the case in many instances where LLMs are involved in violent planning, that the LLM was urging this behavior on. Basically entrapment - you are encouraged by a robot to become more violent and vindictive and then when you do you are handed over to police.
I'm not really sure that this can be likened to a diary when it is called a "chat" but that's for the legal system to determine, not me sitting on my couch.
Can the public also immediately get alerted when a cop or politician does some bad shit, though?
Every single lie of yours is exposed in the past few months.
You get privacy if you're a big corporation that needs to make sure OpenAI/Google/Anthropic can't read your trade secrets etc.
But those contractual privacy protections have been in place for a long time. It doesn't have anything to do with AI, it's been the same with Office365, Google Docs, etc.
— via https://www.leg.state.fl.us/Statutes/index.cfm?App_mode=Disp...
The DA is serious about "in any matter."
To me, that is the more interesting legal question. Does a LLM-based safety net that sends content to a human, when the original use case would not have sent it to a human, count as "may be viewed by another person". It certainly wasn't intended to be, and that isn't the norm. At the same time, because no security is perfect, we could say that any digital record, stored in any way "may be viewed by another person."
Something for the courts to sort out, of course.
> may be viewed by another person
Was it viewed by another person? Yes.
https://www.politico.com/news/2023/08/30/desantis-warns-hurr...
What would happen if it had scanned a file it didn’t have permission to look at and found this threat?
I honestly don’t know how I feel about this. On the one hand if you’re using claude as a diary you have no expectation of privacy and she was talking about committing a very serious crime.
This still makes me feel queasy though.
Over in Europe they want to read all ofd our private messages, yet these chatbots, pretending to be our friends, will snitch on us just for our thoughts.
It's getting pretty orwellian out there.
One unfortunate woman who happened to write the wrong thing in the wrong place is now having her life turned upside-down for perceived thought-crime.
To Anthropic, and all employees working there, your company's product and the result of your work is cruelty. You are enabling it and pushing it down everyone's throat. You can never again claim that you are the "ethical" AI company, for no such thing exists.
The downside is that you don't get cached prompt discounts, so you pay a heavy price for ZDR that way.
A diary constitutes making a threat?
Oh boy the roleplaying part of LLM world is in for a bad time
Then again, I wish this worked in a way that would give users more privacy and agency, instead of less.
How does a LLM prompt satisfy this? I guess it'll be an easy win for her.
Basically, under this interpretation, any personal note you store in the servers of a company could qualify, even if you didn't ever imagine someone would read and as such you couldn't have thought about it as a threat
Seems to fail this test at face value.
I mean, somebody you live with may find and read your diary. Is that the same thing?
Intent matters here. Did you intend somebody else to view it? Does a reasonable person have expectation of privacy with a chatbot?
Thought crimes are real when you're sharing your thoughts with Claude
Summary: don't type in Claude anything you wouldn't like a human to read.
I don't think someone is an idiot for thinking that the information they type into their private Claude account is private. I also don't think people are idiots for thinking their phone is listening to them and giving them targeted advertising based on that. Both are reasonable deductions from their lived experiences. Both are wrong.
https://www.nbcmiami.com/news/local/everyone-deserves-to-die...
https://www.wdsu.com/article/maryland-high-school-student-ch...
https://www.pinellassheriff.gov/21-023-deputies-arrest-pinel...
If it were written on paper, and only in a room with no phones or cameras so fable couldn't hack it, then I think you wouldn't be sharing it.
I think it’s valid to ask if tapping something into Claude is legitimately sharing a threat.
I don’t think it is. I also think the sheriff could have found more-substantial evidence if she was actually planning domestic terrorism.
That said, if the shooting happened and we were looking at this from before? It’s a tough balance without an easy answer.
I think it’s fair to pre-identify folks who fantasise about shooting anyone. It’s a small fraction of the population that looks into logistics versus making offhand comments.