Comment by MisterMunchkin
4 hours ago
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.
> when you only read the threat by spying on them
The AI companies have clauses in their user agreements saying they can review content flagged as harmful. It’s not legally spying.
If you recall previous outrage about ChatGPT being used in cases of suicides or shootings, this is the result. Every time a crime was committed and the police found ChatGPT history about the crime, the media turned it into a frenzy. So the AI labs added safety filters to their consumer plans that detect threats of violence, escalate them to human review, and report to the police.
Spying is not the right analogy because the information was given to the police by a third party which had a EULA saying they would do this. A more analogous situation would be someone reading another person’s diary and then turning it into the police department. There might be some limitation in the law that makes the evidence inadmissible because it was not intended to be shared with anyone, but that’s a separate decision.
Not only that they can review flagged content, but they tend to have separate retention policies for flagged content. Anthropic's is this: "We retain inputs and outputs for up to 2 years and trust and safety classification scores for up to 7 years if your chat or session is flagged by our automated trust and safety systems as violating our Usage Policy."
So don't run for office or anything like that. Someone, somewhere will have a contact that will get that.
>Spying is not the right analogy because the information was given to the police by a third party which had a EULA saying they would do this. A more analogous situation would be someone reading another person’s diary and then turning it into the police department.
This is spying with extra steps couched in corporate speak.
I was responding to a question about the legal case. The police did not perform any spying.
Frustrations about Anthropic’s EULA are a separate matter.
9 replies →
Extra steps couched in corporate speak is often the defining line that defines whether something is technically legal or not.
1 reply →
> This is spying with extra steps couched in corporate speak.
Calling something names doesn't invalidate it. It only invalidates what point you're trying to make.
Well, let's say that you have a regular customer at a bar.
They get friendly and loose-lipped with the bartender over the span of months. Eventually they let slip that they plan on killing their spouse for a life insurance payout. At first the bartender thinks they're joking, but it becomes evident that there's an actual plan being acted upon and someone's life is very likely in imminent danger.
Does the bartender have a responsibility to go to the police?
2 replies →
Many clankers deny data retention or spying on the user if you ask them. That should be completely illegal.
Then, you can write anything in an EULA but it is not automatically legal either.
This seems to be the statute: https://www.leg.state.fl.us/Statutes/index.cfm?App_mode=Disp...
With the obvious IANAL, it doesn't seem to rely on the message be sent to the person being threatened. The specific segment is "in any manner in which it may be viewed by another person".
This may be one of those cases where we get to find out how courts view SaaS platforms.
Interesting that it exempts telephone calls. Why don't we treat other messaging services like phone calls?
It looks like a prior statue covered that: https://www.leg.state.fl.us/Statutes/index.cfm?App_mode=Disp...
The subjective element of crime (i.e. doing it on purpose) is fundamental also in the US legal system. If the person wasn't aware that someone else might see their messages, it should be hard to claim that they committed the crime.
According to Gemini, "Florida appellate courts have overturned juvenile convictions [based on this law] when the state could not prove the person subjectively intended for the record to be seen."
The prosecution will ruin her life regardless of the outcome.
Exactly, can you threaten someone without them receiving the threat? If this is not thrown out, Minority Report will actually happen.
Cops will charge them to let the courts decide
this is almost certainly what anthropic is hoping for here - a judgement that says there is no point in them continuing to monitor and report this behaviour
We are snowballing to Minority Report ...
If the AI recommends murder and you exhibit a pattern of following AI advice are you guilty of precrime
> Surely that has to be thrown out in court?
The prosecutors likely know this and expect it. But there's enough gray area here for them to make the argument, and it's hard to prove malicious prosecution, so they know they'll get away with it. It's just about sending a message to the public - they don't care whether a conviction sticks. Just politics.
The prosecutors aren't on the hook, anyway. They have absolute immunity. The decision to charge is protected. The prosecutor would have to have done one of the few, enumerable things outside the scope of the role, like conducting an investigation without probable cause or hiding exculpatory evidence.
It's not quite spying when you willingly hand over this information and agree to terms of service. You're data is not considered yours alone.
It still shocks me the number of people I know who freely let agents on devices that contain unencrypted private keys, freely dump internal data into cloud models and generally don't give a second thought about any of it being trained on, inevitably leaked one day in a db breach or read by providers. I find it's best to consider any data put into a cloud model the same as if it were posted publicly online, since that is the very possible eventual end result.
Hopefully more of these stories push people towards local models :)
the argument to be made is that allowing anthropic to see it constitutes sending the threat.
sandbox your ai.
That is not what sandboxing solves. A good sandbox would inject credentials into provider API calls so that the model never sees credentials, but the provider is still going to see the transcript. Sandboxes do not require or imply that there is a local model. Sandboxes limit what the agent can access on the host machine as well as the network and public internet.
>Sandboxes limit what the agent can access on the host machine as well as the network and public internet.
this is exactly what I meant. I am presuming the danger is AI reacting to personal notes that it reads on your computer, like a diary, and you should not allow the tools to have access to those documents.
how does sandbox help in this case when you use a provider like anthropic/openai?
Another way to interpret this is that they are legally presuming that you already have sandboxed their product and anything it sees or has access to is intentional.
Any failure to understand what it can access or what it has permission to see from the user's end is presumably not their problem. Regardless of what the user specifically asks of the tool.
If you're still using a provider like this then you didn't sandbox the AI. You still need to follow the instruction.
1 reply →
It's legal to spy if the terms of service say so
I was thinking the same. If the evidence was not obtained with a proper court order wouldn’t this result in a mistrial?
If you overhear someone, in the privacy of their house, threatening to murder someone and go to the police, surely you don't expect this report being thrown out and you being yourself charged with the violation of someone's privacy instead?
IIRC if the evidence wasn't lawfully gathered (which it sounds like it was, tbh) then it wouldn't be a mistrial, it would be thrown out and then the prosecution wouldn't have any evidence of any crime.
what if it's just testing the AI to see how it responds
A threat sent by mail is still a sent threat even if nobody ever opens the envelope to read it. The crime is in the sending. This woman used an online resources, one which involves transmitting everything across innumerable state lines. I am surprised she isn't up on federal charges.
Note that the law doesn't forbid the writing of a threat. You have to send it to someone. Had she kept it in a book under her bed, she would not be in trouble. But she sent it to a website/service/LLM portal.
>> 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
If you draft an email threatening someone and delete it without sending have you committed a felony because someone at Google could be reading your drafts box, stored in a datacenter across state lines?
Honestly, I'm equally fascinated by the way email has changed. 30 years ago when you drafted an email but didn't send it, it was only on your local machine. There was no SMTP. 20 years ago, it might be a 50/50 shot as to whether you "transmitted" it to your "Drafts" folder if you were using IMAP instead of POP3 to read it.
1 reply →
Sounds reasonable. Google's bots could pick that up easily and forward if for human review.
FYI, the use of drafts folders to transmit messages has been used by terrorists. This is likely where CIA director David Petraeus got the idea when he needed a secure way to chat with his mistress.
https://www.findlaw.com/legalblogs/technologist/gen-petraeus...
3 replies →
What if she mailed it to herself?
What if she put it in a locked box before shipping it to herself UPS, and she has the only key?
What if instead of UPS, she hired a moving company to move the locked box?
What if she wrote it electronically in diary.txt, but it was backed up to a cloud provider?
--
I'm guessing there's some sort of "reasonable expectation of privacy" for certain activities. We're going to find out what Florida courts think about this new medium.
We'll only find out what the courts think when this happens to someone with a lot of money. It takes a real legal fight to push it high enough to become precedence. She'll be pushed to plea out.
Surely this is the wrong side of what "sending" here will be interpreted as?
Saving is not sending ie passive vs active act.
Everything you "save" on an online service gets "sent" to someone, be that a person or a computer, more often than not across state lines.
> in any manner in which it may be viewed by another person
Does the person have to know (or at least believe) that it will be viewed by another person?
She likely didn't think anyone would view it. Honestly, even as a career software developer I don't think it is unreasonable to think know would would see what she wrote to an AI. I assume most of what I write to an AI is not viewed by any other human, based simply on the quantity of messages sent back and forth to AIs, I would assume a vast majority are not read by another human.
What if she had written this into google docs, and she kept a diary there? That also crosses state lines, and is transmitted to another location.
I get were you are coming from but this all feels like it needs more context to make a better judgement.
You can argue from technicalities but they would need to prove intent.