Hegseth to Anthropic: Nice company you got there…

[jon-nyc]

@Horace said in Hegseth to Anthropic: Nice company you got there…:

@jon-nyc said in Hegseth to Anthropic: Nice company you got there…:
Another framing would be that the DoD refuses to delegate authority to Anthropic to decide what is or is not ethical or illegal.

If you're right, then Anthropic can just write a letter saying they take no position on the legality and ethics of mass surveillance of US citizens and the matter is solved.

After all (according to you) the DoD has not demanded that they provide the capability.

[Horace]

The Pentagon's demand is that Anthropic agree to "all lawful use" without the company imposing its own additional restrictions.

That does not depend on Anthropic setting fire to its ethics branding.

[Horace]

An analogy would be if guns had broad strokes AI built into them intended to prevent unlawful discharge, but that AI was known to provide false positives. For the general public, maybe the cost / benefit of such an AI would be acceptable, but for the military, no such gun would ever be bought. If it was a great gun, the military would either demand that the guardrails be removed, or it would not do business with that company. But I do agree that punishing the company with a supply chain threat designation is overreach.

Presumably anybody could agree that AI guardrails against "illegal mass surveillance" could not possibly discern with perfect accuracy whether it is being used for that. There will necessarily be false negatives and false positives. The military doesn't want to deal with that. And that's reasonable.

[Horace]

We can also pick apart the concept that there is such a thing as unambiguous "illegal mass surveillance", to say nothing of an AI's ability to know that it's a part of such a scheme. I am very skeptical of jon's frequent claims of what is or is not obviously illegal. The actual law is, more often than not, ambiguous, and the certainty of illegality is just a rhetorical trick.

how unambiguously defined is "illegal mass surveillance" in US law?

Claude:
It's not well-defined at all, and that's a significant part of the problem.

There's no single statute that says "mass surveillance is illegal" in clear terms. What exists is a patchwork. The Fourth Amendment prohibits unreasonable searches, but what counts as a "search" in the context of AI-powered data analysis is unsettled law. The Posse Comitatus Act restricts military involvement in domestic law enforcement, but has exceptions and has been interpreted narrowly. Executive Order 12333 governs intelligence collection and has restrictions on surveilling US persons, but it's an executive order — the current president can modify or revoke it. FISA governs foreign intelligence surveillance but has been interpreted to permit bulk collection programs that many would call mass surveillance. The NSA programs Snowden revealed were happening under legal frameworks that the government argued were lawful.

The fundamental ambiguity is that "surveillance" at scale using AI doesn't map neatly onto existing legal categories. Is it surveillance if an AI model processes publicly available social media posts to identify patterns? What about analyzing metadata rather than content? What if the system flags individuals for human review rather than taking action autonomously? Existing law was mostly written for wiretaps and physical searches, not for an AI that can process and correlate vast datasets in ways that don't clearly fit the old definitions.

So when Anthropic says it wants to prevent "mass surveillance" and the Pentagon says it will follow "the law," they're both gesturing at a legal landscape that's genuinely ambiguous. Anthropic is essentially saying the law hasn't caught up and they want to err on the side of caution. The Pentagon is saying that ambiguity is for courts and Congress to resolve, not a vendor.

This is actually one of the stronger arguments for the Pentagon's position on this specific point — if the law is unclear, it's not obvious that a private company should be the one drawing the line.

[jon-nyc]

@Horace said in Hegseth to Anthropic: Nice company you got there…:

The Pentagon's demand is that Anthropic agree to "all lawful use" without the company imposing its own additional restrictions.

So they are insisting that they provide the capability? You explicitly said otherwise a few posts back.

[89th]

Turns out it's really hard to build a hammer and then say “you can't use a hammer that way” and also succeed in the hammer business.

I predict a compromise here much like starlink separate product for DoW use case.

[Horace]

@jon-nyc said in Hegseth to Anthropic: Nice company you got there…:

@Horace said in Hegseth to Anthropic: Nice company you got there…:

The Pentagon's demand is that Anthropic agree to "all lawful use" without the company imposing its own additional restrictions.

So they are insisting that they provide the capability? You explicitly said otherwise a few posts back.

Any piece of technology can be used illegally. They are insisting that imperfect guardrails not be imposed on them. That is not the same as an insistence that they are provided with the ability to do illegal things. The AI will provide them with the ability to do illegal things with or without the imperfect guardrails. But what they have not done, is to insist on an ability to break the law. If it were possible for the AI to know what the law is exactly, which of course it isn't, as the law in question is unsettled, then in theory the pentagon would agree to the guardrails. But that thought experiment depends on an impossible universe.

[jon-nyc]

So they’re not insisting that the company provide the capability, they’re insisting that they not NOT provide the capability.

Ok glad we’ve cleared that up.

Shame on me for the tortured framing.

[Horace]

@jon-nyc said in Hegseth to Anthropic: Nice company you got there…:

So they’re not insisting that the company provide the capability, they’re insisting that they not NOT provide it.

Ok glad we’ve cleared that up.

Shame on me for the tortured framing.

It's incoherent to say that the difference between the version of AI anthropic would like to provide, and the version the DoD wants, is that one has the ability to be used illegally and the other does not. No such categorical separation exists. The DoD is insisting on the absence of imperfect guardrails, literally. That is not the same as "insisting on the ability to break the law". If all they wanted was the ability, they could use the version Anthropic suggests. They are insisting that Anthropic not be in the loop regarding whether something is legal or illegal. They have pledged to follow law, such as it is.

[jon-nyc]

Ok, so let’s leave the legality to the future lawyers since it depends on actual use.

For tomorrow’s deadline, they’re insisting that Anthropic NOT NOT provide a certain capability. Which is different than insisting they DO provide that same capability. In fact, the latter is dishonest tribal rhetoric.

Ok, I’m learning. Don’t give up on me yet.

[Horace]

No, you're not really learning. But I'm patient. You may not have understood my previous post; feel free to read it again. The difference in versions that Anthropic wants to provide and that DoD wants them to provide, is not one of categorical "can or cannot be used for illegal surveillance". This is an important point. The DoD only wants imperfect guardrails removed.

It is simply and objectively false to think that the version Anthropic would like to provide, will perfectly prevent itself from use in "illegal" surveillance while allowing itself to be used in legal circumstances. The DoD is demanding that the imperfect guardrails not be a potential impediment to their legal uses. Explicitly that is their demand. Yes, your framing is tribal and tortured.

[89th]

Didn't we have a Claude member here once? Maybe he can weigh in. If not, @klaus is as close as we get.

[Renauda]

@89th said in Hegseth to Anthropic: Nice company you got there…:

Didn't we have a Claude member here once? Maybe he can weigh in. If not, @klaus is as close as we get.

That’s going back awhile. Yeah, I think his complete handle was Claude Balls.

I just assumed it another one of the late Larry’s numerous fun sock puppets

[jon-nyc]

Ha. I remember that now.