Anthropic irks White House with limits on models’ use | Hacker News Reader

Anthropic irks White House with limits on models’ use

semafor.com

246 pointsmindingnever129 comments

Very strange writing from semafor.com

>For instance, an agency could pay for a subscription or negotiate a pay-per-use contract with an AI provider, only to find out that it is prohibited from using the AI model in certain ways, limiting its value.

This is of course quite false. They of course know the restriction when they sign the contract.

There are (or at least WERE) entire divisions dedicated to reading every letter of the contract and terms of service, and usually creating 20 page documents seeking clarification for a specific phrase. They absolutely know what they're getting into.

Perhaps it's the finetune of Opus/Sonnet/whatever that is being served to the feds that is the source of the refusal :)

I have a feeling in today's administration which largely "leads by tweet" that many traditional "inefficient" steps have been removed from government processing, probably including software on-boarding.

Can confirm these teams are still around. There is now an additional "SME review group" that must comb through any and all AI-related issues that were flagged, sends it back down for edits and must give final approval for before docs are sent over to provider for response. Turnaround has gotten much slower (relatively)

Or you can use personal accounts to bypass red tape for government business.

I have a legal education but reading TOS and priv policy docs at account creation is purposefully too time consuming by design.

One my fave new AI prompts: you are my Atty and a expert in privacy law and online contracts-of-adhesion. Review the TOS aggreement at [url] and privacy policies at [url] and brief me on all areas that should be of concern to me.

Takes 90 seconds from start to finish, and reveals how contemptuously illusory these agreements are when SO MANY reserve the right to change anything with no duty to disclose changes.

Are you sure that every restriction that’s in the model is also spelled out in the contract? If they add new ones, do they update the contract?

The contracts will usually say “You agree to the restrictions in our TOS” with a link to that page which allows for them to update the TOS without new signatures.

Usually, contracts will note that you will be notified of changes ahead of time, if it's a good faith contract and company that is.

Here in Sweden contracts are a specific thing, otherwise it's not a contract, so agreeing to conditions that can be changed by the other party simply isn't a contract and therefore is just a bullshit paper of very dubious legal validity.

I know that some things like this are accepted in America, and I can't judge how it would be dealt with. I assume that contracts between companies and other sophisticated entities are actual contracts with unchangeable terms.

I know that some things like this are accepted in America

Not really. Everything you said about contracts above applies to contracts in America last time I checked. Disclaimer: IANAL, my legal training amounts of 1 semester of "Business Law" in college.

In theory yes, but you also have this stuff where people agree to get medical treatment and the price isn't specified.

This would be a non-contract in Swedish law, for example.

It is also illegal in USA, although that only changed recently.

One thing about the US, is how we handle settings where one could conceptualize a contract as being needed, but where it would be way too inefficient and impractical to negotiate, write out, understand, and sign, a written contract. In those cases, which includes things like retail sales, restaurants, and may other cases, the UCC or Uniform Commercial Code[1][2] applies. Not sure offhand if that relates to the medical example or not, but I expect that at least some similar notion applies. So there are binding laws that cover these transactions, it's just not done the same way as a "full fledged contract".

[1]: https://en.wikipedia.org/wiki/Uniform_Commercial_Code

[2]: The UCC also covers other things, but these cases are a lot of what it's best known for.

Knowing medical prices up front would be entirely possible and practical. In most situations, you should be able to sign contract up front.

Yeah, I’ve signed dozens of contracts for services and some are explicit in the way you expect but a lot of software or SAAS type contracts have flexible terms that refer to TOS and privacy policies that are updated regularly. It’s uncommon that any of those things are changed in a way that either party is upset with so companies are generally okay signing up and assuming good faith.

All the US megacorps tend send me emails saying "We want to change TOS, here's the new TOS that's be valid from date X, and be informed that you have the right to refuse it" (in which case they'll probably terminate the service, but I'm quite sure that if it's a paid service with some subscription, they would have to refund the remaining portion) - so they can change the TOS, but not without at least some form of agreement, even if it's an implicit one 'by continuing to use the service'.

The contact's restriction is on the usage of the model, not the behavior of the model.

This whole article is weird to me.

This reads to me like:

* Some employee somewhere wanted to click the shiny Claude button in the AWS FedRamp marketplace

* Whatever USG legal team were involved said "that domestic surveillance clause doesn't work for us" and tried to redline it.

* Anthropic rejected the redline.

* Someone got mad and went to Semafor.

It's unclear that this has even really escalated prior to the article, or that Anthropic are really "taking a stand" in a major way (after all, their model is already on the Fed marketplace) - it just reads like a typical fed contract negotiation with a squeaky wheel in it somewhere.

The article is also full of other weird nonsense like:

> Traditional software isn’t like that. Once a government agency has access to Microsoft Office, it doesn’t have to worry about whether it is using Excel to keep track of weapons or pencils.

While it might not be possible to enforce them as easily, many, many shrink-wrap EULAs restrict the way in which software can be used. Almost always there is an EULA carve-out with different tier for lifesaving or safety uses (due to liability / compliance concerns) and for military uses (sometimes for ethics reasons but usually due to a desire to extract more money from those customers).

> due to a desire to extract more money from those customers

If it gives you high priority support, I dont care, if its the same tier of support, then that's just obnoxiously greedy.

Could also be an article placed by a competitor + a squeaky wheel.

A classic:

THIS SOFTWARE PRODUCT MAY CONTAIN SUPPORT FOR PROGRAMS WRITTEN IN JAVA. JAVA TECHNOLOGY IS NOT FAULT TOLERANT AND IS NOT DESIGNED, MANUFACTURED, OR INTENDED FOR USE OR RESALE AS ONLINE CONTROL EQUIPMENT IN HAZARDOUS ENVIRONMENTS REQUIRING FAILSAFE PERFORMANCE, SUCH AS IN THE OPERATION OF NUCLEAR FACILITIES, AIRCRAFT NAVIGATION OR COMMUNICATION SYSTEMS, AIR TRAFFIC CONTROL, DIRECT LIFE SUPPORT MACHINES, OR WEAPONS SYSTEMS, IN WHICH THE FAILURE OF JAVA TECHNOLOGY COULD LEAD DIRECTLY TO DEATH, PERSONAL INJURY OR SEVERE PHYSICAL OR ENVIRONMENTAL DAMAGE.

I never knew Java was so dangerous

Everything is dangerous by default. That's the point.

Look up DoD (DoW?) 882 and LOR ratings. This is a fancy way of saying “Java can’t do that because we haven’t certified a toolchain for it”

And for bonus points, go find the last certified compilers for LOR1 rating that follow 882 guidelines.

Now you’ve scratched the surface of safety-critical software. Actually writing it is a blast. I think most web developers would weep in frustration. “Wait, I can’t allocate memory that way? Or that way? Or in this way not at all?! There’s no framework?! You mean I need to do all this to verify a button click??!!”

Most web developers don't know what "memory allocation" is and let alone how to manually allocate it.

I think people that don't write safety-critical software in general will weep in frustration, not just web developers.

reminds me of jslint

author added "must be used for good, not evil" to the license

...and IBM asked for an exception.

https://en.wikipedia.org/wiki/JSLint#License

https://news.ycombinator.com/item?id=5138866

note that restricting use of software makes it non-free gpl-wise.

RMS said the GPL does not restrict rights of the USER of software, just that when the software is redistributed, the rights are passed along.

Good old Douglas Crockford. He also put the "must be used for good, not evil" restriction on JSON, which he invented. Obviously JSON is used for all kinds of evil, though.

A much younger, more naive me (~20 years ago) actually emailed him to complain about the ambiguous terms and he replied saying something to the effect of "It's obviously unenforceable, get over it."

This excerpt does not disclaim liability for death, personal injury or physical damage (b/c generally they can't)

Nor does it prohibit use or resale of "Java Technology" for any particular purpose

It does suggest they are aware of Java's shortcomings

Don’t tech companies change ToS quite frequently and sometimes in ways that’s against the spirit of what the terms were when you started using it?

This is a contract, not a click through license. You can't do that.

(Legally you can't do it with a click-through either, but the lack of a contract means that the recourse for the user is just to stop buying the service.)

This feels like a hit piece by semafor. A lot of the information in there is purely false. For example, Microsoft's AI Agreemeent says (prohibits):

"...cannot use...For ongoing surveillance or real-time or near real-time identification or persistent tracking of the individual using any of their personal data, including biometric data, without the individual’s valid consent."

There's nothing stopping Microsoft hammering out different terms for certain customers, like governments.

Semafor is Ben Smith's blog, trying to imitate a reputable newspaper like Financial Times.

Gosh, I guess the SaaS distribution model might give companies undesirable control over how their software can be used.

Viva local-first software!

In general I applaud this attitude but I am glad they are saying no to doing surveillance.

Me too, actually, but this is some "leopards ate their face" schaudenfraude that I'm appreciating for the moment.

EULAs can impose limitations on how you use on-premises software. Sure, you can ignore the EULA, but you can also do so on SaaS, to an extent.

Are EULAs even enforceable? SaaS at least have the right to terminate service at will.

With SaaS, you can be monitored and banned at any moment. With EULAs, at worse you can be banned from updates, and in reality, you probably won't get caught at all.

One of the very few tech companies who have refused to bend the knee to the United States' current dictatorial government.

It's startling how few are willing to. I'm rooting for them.

Can we trust this though? “Cooperate with us and we’ll leak fake stories about how frustrated we are with you as cover”.

And I’m not singling out Anthropic. None of these companies or governments (i.e. people) can be trusted at face value.

They don't do that. They're not capable of cooperating with anyone, it's maximum punishment all the time. It's unclear if they can keep secrets either.

Yeah, and a private capitulation doesn't accomplish their goals either. They want the spectacle of public submission.

Dictatorial suggests a "ruler with total power". The US has three branches of government. That hasn't changed, ever.

Two of them jump at the command one the other one, one out of fear (because he has ended the careers of every rep that has crossed him), and the other has been packed with life-time-appointment sycophants who put loyalty to the cut over anything else.

Russia (or literally any other dictatorial tyre pyre) also has three branches of government and a token opposition, for all the good it does.

Just because you have a nice piece of paper that outlines some kind of de jure separation of powers, doesn't mean shit in practice. Russia (and prior to it, the USSR) has no shortage of such pieces of paper.

The definition of dictatorial government is either a single person or a small group of people. So there being three branches of government doesn't necessarily prohibit a government from being a dictatorship if they are all working together to enact their authoritarian control without constitutional limits.

But really this is just pointless semantics. It doesn't matter what it is called it is still a problem.

Technically (and legally) the USSR also had the same three branches of government; just all controlled by the same party.

The Supreme Court has repeatedly and crassly decided in favor of the current administration; among their decisions one to elevate the president above the law, and a carte Blanche for gerrymandering with a footnote that that is against the spirit of democracy.

The congress is dominated by the republicans, who have given up on every last shred of dignity and turned themselves into yes men that will approve anything Trump says, from a justified invasion of Greenland to how unnecessary it would be to publish the Epstein files.

And the executive branch currently hunts down government employees with an unsuitable personal opinion, takes jet plane bribes from foreign leaders, and tries to eradicate slavery and the Native American genocide from museums and school books.

Tell me about those three branches again. Right now, they have been perverted into a single tool to carry out the whims of an egotistic asshole backed by a powerful group of conservative activists.

because of this they're probably on borrowed time in this political climate

"Eventually, though, its politics could end up hurting its government business."

Good? What if, and I know how crazy this sounds, not using AI to surveil people was a more desirable goal than the success of yet another tech company at locking in government pork and subsidies?

Here's an entertaining example from 20 years ago:

By using the Apple Software, you represent and warrant that you ... also agree that you will not use these products for any purposes prohibited by United States law, including, without limitation, the development, design, manufacture or production of missiles, or nuclear, chemical or biological weapons. -- iTunes

No production of missiles with iTunes? Curses, foiled again.

The US government can train their own damn LLM if they want an unrestricted one so bad.

Billions, versus "I'll buy that for a dollar."

https://www.reuters.com/business/retail-consumer/anthropic-o...

Exactly. If you aren't willing to pay the costs, then deal with the terms.

What is the govt expecting to do in combination of surveillance and Antrhopic models? I'm not convinced this is any kind of valid job function.

Honestly makes me think better of Anthropic. Lets see how long they stick to their guns. I believe they will fold sooner rather than later.

First, contracts often come with usage restrictions.

Second, this article is incredibly dismissive and whiny about anyone ever taking safety seriously, for pretty much any definition of "safety". I mean, it even points out that Anthropic has "the only top-tier models cleared for top secret security situations", which seems like a direct result of them actually giving a shit about safety in the first place.

And the whining about "the contract says we can't use it for surveillance, but we want to use it for good surveillance, so it doesn't count. Their definition of surveillance is politically motivated and bad"! It's just... wtf? Is it surveillance or not?

This isn't a partisan thing. It's barely a political thing. It's more like "But we want to put a Burger King logo on the syringe we use for lethal injections! Why are you upset? We're the state so it's totally legal to be killing people this way, so you have to let us use your stuff however we want."

> The policy doesn’t specifically define what it means by “domestic surveillance” in a law enforcement context and appears to be using the term broadly, creating room for interpretation.

> Other AI model providers also list restrictions on surveillance, but offer more specific examples and often have carveouts for law enforcement activities. OpenAI’s policy, for instance, prohibits “unauthorized monitoring of individuals,” implying consent for legal monitoring by law enforcement.

This is unintentionally (for the author) hilarious. It's a blatant misinterpretation of the language, while complimenting the clarity of the lanuage. Who "authorizes" "monitoring of individuals"? If an executive agency monitors an individual in violation of a court order, is that "authorized" ?

Wasn't a big part of AI 2027 that government employees became overly reliant on AI and couldn't function without it. So guess we are still on track to hit that timeline.

A big part of what?

https://ai-2027.com/

https://www.youtube.com/watch?v=k_onqn68GHY

One of the stages was gov/military personnel became dependent on AI for decision making.

So.. a sci-fi short story?

That is the sarcastic take.

More generously, 'an Expert-Driven forecast that offers a possible extrapolation' or possible realistic scenario.

Parts of it seem to be coming true already, parts of it are dependent on some technology breakthroughs that haven't happened.

Forecasting mileage may vary

LW’s forecasting ability here has been notoriously bad, fwiw. I don’t have any reason to trust anything they say on the subject, and in fact every reason to suspect.

So a private company sanctioned the US government? And now the US government is upset?

I do love the smell of hypocrisy early in the morning.

I Anthropic

This isn't a principled stand, it's just a negotiating tactic. They'll allow it when the price is right.

Are government agencies sending prompts to model inference APIs on remote servers? Or are they running the models in their own environment?

It’s worrying to me that Anthropic, a foreign corporation (EDIT: they’re a US corp), would even have the visibility necessary to enforce usage restrictions on US government customers. Or are they baking the restrictions into the model weights?

Anthropic is US-based - unless you meant something else by "foreign corporation"?

> It’s worrying to me that Anthropic, a foreign corporation, would even have the visibility necessary to enforce usage restrictions on US government customers.

"Foreign" to who? I interpretted your comment as foreign to the US government (please correct me if I'm wrong) and I was confused because Anthropic is a US company.

Ah my mistake. I thought they were French. I got them confused with Mistral.

The concern remains even if it’s a US corporation though (not government owned servers).

Ah yes - Mistral is the largest of the non-US, non-Chinese AI companies that I'm aware of.

> The concern remains even if it’s a US corporation though (not government owned servers).

Very much so, I completely agree.

Anthropic is pretty clearly using the Häagen-Dasz approach here, call yourself Anthropic and your product Claude so you seem French. Why?

Hah, it was indeed the Claude name that had me confused :D

According to Claude, it’s named after Claude Shannon, who was American.

But it might also be the albino alligator in the California Academy of Sciences in SF.

Everyone spies and abuses individuals' privacy. What difference does it make? (Granted I would agree with you if Anthropic were indeed a foreign based entity, so am I contradicting myself wonderfully?)

1) Anthropic are US based, maybe you're thinking of Mistral?

2) Are government agencies sending prompts to model inference APIs on remote servers?

Of course, look up FedRAMP. Depending on the assurance level necessary, cloud services run on either cloud carve-outs in US datacenters (with various "US Person Only" rules enforced to varying degrees) or for the highest levels, in specific assured environments (AWS Secret Region for example).

3) It’s worrying to me that Anthropic, a foreign corporation, would even have the visibility necessary to enforce usage restrictions on US government customers.

There's no evidence they do, it's just lawyers vs lawyers here as far as I can tell.

No judgement here, but a US-based corporation refusing services to the US Government?

While the terms of service are what they are, the US Government can withdraw its military contracts from Anthropic (or refuse future contracts if they don't have any so far). Or softly suggest to its own contractors to limit their business dealings with Anthropic. Then Anthropic will have hard time securing computing from NVIDIA, AWS, Google, MSFT, Oracle, etc...

This won't last.

I'm sure this sort of unofficial blacklisting is fairly common, but it does seem very opposed to the idea of a free market. It definitely doesn't seem like Anthropic was trying to make some sort of point here, but it would be cool if all the AI companies had a ToS saying it can't be used for any sort of defense/police/military purposes

I am not even sure what free market is, aside from Economics textbooks and foreign policy positioning. Whatever it may be, I don't think we had it for quite some time.

I am of an age where I read comments like this with my mouth agape. It is (was) perfectly normal to choose whether or not to do business with the government.

A lot of abnormal things happen when your government becomes a dictatorship.

Your grandchildren will disown you when they learn what you supported in 2025.

First of all, no need for personal attacks. Second, I am not supporting this, merely stating the reality.

Would you rather pretend that things like that aren't happening?

There's absolutely a need for personal attacks. By suggesting that people ought to just give in and accept this as the new reality, you are supporting it, and you need to know that this represents a severe moral flaw. You're not just being pragmatic, or realistic, or whatever polite fiction you tell yourself to avoid having to take a stand. You're being a bad person, intentionally doing your part to make the country worse in the hopes that you can be a little more comfortable. Change your ways now or you'll bear the shame the rest of your life.

Reply on news.ycombinator.com