Large-scale online deanonymization with LLMs

[barnhart]

Sounds more like an Oldsmobile than a novel.

And an interesting Olds, let's see 90 barrel carbs, 8 speed transmission and quad exhaust...

The original model number scheme from the 1940s was Series 90 (larger than Series 60, 70, 80) + 8 cylinder = 98, later spelled out as Ninety Eight, even after the downsizing of '85 when the only available engine was a V6. Sadly, the 98 was discontinued after '96 and didn't quite make it to the eponymous model year. The Four could refer to 4 doors.

I thought it might follow the 442 pattern.

[ken_sylvania]

There is a saying among anti tech thinkers that goes something like "the future is here, it's just unevenly distributed." Meaning the tech of the future is already being employed against people with the least agency who can do little about it. An example might be monitoring prisoners or surveillance in crime prone neighborhoods. Eventually the group who have little power to object becomes larger until it encompasses nearly everyone but the very top.

My guess is this has been a tool available for some people for a while already, we are amazed only because it was hidden.

Worth noting: the government is suing Anthropic because they do not allow their LLMs to be used for two specific things:

1. Mass surveillance
2. Making the decision whether an autonomous weapon should kill someone without prior human approval

I don't think LLMs should be used for either purpose. It alarms me that the government is publicly suing Anthropic over this, and even threatening to use wartime production powers if Anthropic does not back down.

Also alarming: Anthropic is alone in this. OpenAI, Gemini, and Grok do not have restrictions like these.

I hadn't heard about that and I can't find any information about it online. Care to share your source for this?

[Soloist]

There is a saying among anti tech thinkers that goes something like "the future is here, it's just unevenly distributed." Meaning the tech of the future is already being employed against people with the least agency who can do little about it. An example might be monitoring prisoners or surveillance in crime prone neighborhoods. Eventually the group who have little power to object becomes larger until it encompasses nearly everyone but the very top.

My guess is this has been a tool available for some people for a while already, we are amazed only because it was hidden.

Worth noting: the government is suing Anthropic because they do not allow their LLMs to be used for two specific things:

1. Mass surveillance
2. Making the decision whether an autonomous weapon should kill someone without prior human approval

I don't think LLMs should be used for either purpose. It alarms me that the government is publicly suing Anthropic over this, and even threatening to use wartime production powers if Anthropic does not back down.

Also alarming: Anthropic is alone in this. OpenAI, Gemini, and Grok do not have restrictions like these.

I hadn't heard about that and I can't find any information about it online. Care to share your source for this?

https://www.anthropic.com/news/stateme ... ent-of-war
There’s a few stories floating around about it, but this is their take on it.

[mike]

One of the big problems with using LLMs in these kinds of ways is not only the deanonymization, but also the fact that AI will always make some mistakes. It is of course getting better all the time, but there is no way that AI will always be completely accurate, and this will sooner or later result in people being wrongly framed for things they didn't say.

[ken_sylvania]

Worth noting: the government is suing Anthropic because they do not allow their LLMs to be used for two specific things:

1. Mass surveillance
2. Making the decision whether an autonomous weapon should kill someone without prior human approval

I don't think LLMs should be used for either purpose. It alarms me that the government is publicly suing Anthropic over this, and even threatening to use wartime production powers if Anthropic does not back down.

Also alarming: Anthropic is alone in this. OpenAI, Gemini, and Grok do not have restrictions like these.

I hadn't heard about that and I can't find any information about it online. Care to share your source for this?

https://www.anthropic.com/news/stateme ... ent-of-war
There’s a few stories floating around about it, but this is their take on it.

Doesn't say anything about the US government suing Anthropic though. Just talks about the dispute and a threat to not use their product. I knew about that but not that the government was suing them.

[Soloist]

I hadn't heard about that and I can't find any information about it online. Care to share your source for this?

https://www.anthropic.com/news/stateme ... ent-of-war
There’s a few stories floating around about it, but this is their take on it.

Doesn't say anything about the US government suing Anthropic though. Just talks about the dispute and a threat to not use their product. I knew about that but not that the government was suing them.

I haven't heard about suing, I've heard about them having 72 hours to comply otherwise they get "forced" by the war powers act. I haven't kept up on all the specifics.

[JohnH]

Doesn't say anything about the US government suing Anthropic though. Just talks about the dispute and a threat to not use their product. I knew about that but not that the government was suing them.

In this day and age, people often sue once the government stops paying them millions or billions of dollars. Apparently, that now extends to the U.S. government deciding to just stop buying something because it doesn't meet their needs (or, in this case, because the seller has decided they are a unit of government who can tell the military what to do.)

[Soloist]

Doesn't say anything about the US government suing Anthropic though. Just talks about the dispute and a threat to not use their product. I knew about that but not that the government was suing them.

In this day and age, people often sue once the government stops paying them millions or billions of dollars. Apparently, that now extends to the U.S. government deciding to just stop buying something because it doesn't meet their needs (or, in this case, because the seller has decided they are a unit of government who can tell the military what to do.)

yeah well next time an Amish company is sued to build weapons I'll make sure to tell them what side you are on.

[JohnH]

yeah well next time an Amish company is sued to build weapons I'll make sure to tell them what side you are on.

I am not aware of a great many Amish weapon-builders, but if one were, I would imagine they would be subject to either (a) the government not buying weapons from them, or else (b) the government using its lawfully appointed powers to seize production from such a weapons factory if the Amish decided not to sell to the government.

But in any case, these aren't even munitions. This is just computer software & automated services (that, as far as I understand it, are conducted on government-controlled computing facilities in government data centres and isolated from everything else). The law is pretty clear that the government can seize production of something if it is needed for national defence or national security purposes.

A more fitting analogy might be an Amish farm that decides to refuse to sell food to food processors that might eventually send such food to soldiers on the front lines.

[JohnL]

Sounds more like an Oldsmobile than a novel.

There’s your proof right there that AI autocorrection can’t even get a simple book title right. I typed in “George Orwell's book Nineteen Eight-Four”. Those old typewriters were a lot more honest and dependable than some software writer.

Alas it appears the old typewriter can't even get the simple book title right either.

You see. Autocorrect. It's a digital AI minor demon that plagues all software.