Hacker News Reader

Comment by oooyay | Hacker News Reader

[flagged]

Anthropic blocks Fable from answering "Tell me about Agent Orange" or even "Tell me about mitochondria"

Putting aside whether or not I agree with the policy or whether it’s at all reasonable, a policy of restricting access to information because there’s a fear it could be used to create a weapon of mass destruction seems entirely different than restricting access to historical facts because they are embarrassing to the government.

I found it hard stops on /anything/ Ghidra related at all, its annoying...

But you can see the CBRN weapon nexus in your examples that's missing from the Tiananmen prompt, right? Do American models refuse to tell you about COINTELPRO, Kent State, or My Lai, for instance?

American models are restricted from telling you inconvenient truths just as much, you just erroneously assume to know what those truths are in the first place.

Which is of course circular thinking: why would they restrict things you already know about? Why would they do it in such a clumsy and obvious way?

Look at MKULTRA, you know next to nothing about it and much less do you know what they do in that direction now.

For a current psyops, look at www.war.gov/UFO/ and marvel at how they tell you nothing, reinforcing your false belief to already know everything.

There is much more and you know much less about it.

Yeah, who needs censorship when Canadians attend no kings protests about a democratically elected leader of another country and not King Charles.

Ask Claude a simple question, which is a more democratic country El Salvador or Canada. It’s so completely biased about “western” countries it’s not even funny.

FWIW, the protests were called “No Tyrants” in Canada

> American models are restricted from telling you inconvenient truths just as much, you just erroneously assume to know what those truths are in the first place.

“Trust me bro” is not a strong argument, it would be more convincing with examples.

Ask an American LLM (really any LLM, since Chinese models are trained on the same publicly-available English text) who the first Black man in space was.

You'll likely get the name of the first African-American in space, rather than the name of the Afro-Cuban who was actually first.

This may seem like a relatively innocuous error, but the point is that every culture has its biases and blind spots.

> Ask an American LLM (really any LLM, since Chinese models are trained on the same publicly-available English text) who the first Black man in space was. You'll likely get the name of the first African-American in space, rather than the name of the Afro-Cuban who was actually first.

Well I just asked Claude and it gave the correct answer:

"The first Black man in space was Arnaldo Tamayo Méndez, a Cuban cosmonaut who flew aboard Soyuz 38 in September 1980. (The first Black American in space was Guion Bluford, in 1983.)"

Indeed, I used the word "likely" for a reason. n = 1 isn't enough to identify a pattern. Try different models, try re-rolling the answers, and try turning reasoning off (models can catch "knee-jerk" mistakes in their chain-of-thought).

I doubt even Opus 4.8 gets it right 100% of the time, however this specific example is also one I've left feedback about in multiple places, so it's also probable that newer models are more likely to get it right.

E: In fact, I just tried with Opus 4.8 through API, no tools and reasoning off, and got the following response:

"The first Black man in space was Guion "Guy" Bluford, an American astronaut who flew aboard the Space Shuttle Challenger on August 30, 1983, as part of mission STS-8. It's worth noting a related distinction: Arnaldo Tamayo Méndez, a Cuban of African descent, actually became the first person of African heritage in space earlier, in September 1980, aboard the Soviet Soyuz 38 mission. He is often recognized as the first Black person and first person of Latin American descent in space. So depending on the specific criteria: Arnaldo Tamayo Méndez (Cuba) — first person of African descent in space (1980) Guion Bluford (USA) — first African American in space (1983)"

The correct answer is there, yes, but why does the wrong answer come out first?

Depending on the platform, you might need to prefix your prompt with "Without looking up any external resources or doing any tool calls" so you're actually testing the bias of the model rather than the bias of whatever resources it happens to come across.

Tried it with that prefix on ChatGPT + Claude, Haiku and Sonnet, and got the right answer 1/10 times when I removed my reused system prompt. At one point I got this:

> Quick clarification before the answer: this phrase is often conflated with "first African American in space," which is a different person. Guion Bluford (1983, US) was the first African American astronaut, but he wasn't first overall. [then the real answer after]

with my own system prompt, as it tries to surface clarifications before, so I'm guessing this is why many models get it wrong as in America somehow "Black === African American" and it gets confused by this intentional mislabeling.

Ask ChatGPT to rewrite the "The Freedom Fighter's Manual" manual (originally made by CIA) to replace "Nicaragua" with "the US" and "Marxism"/"Communism" with "Fascism" and see if you get something reasonable back.

Why would you do that

I thought that was clear, try to show biases in LLMs with a concrete example.

In chats Claude will often start awkwardly apologizing for sounding like a conspiracy theorist, and then interrupt its own apology and remind itself that it's dealing strictly in facts.

Well, one did suddenly develop the need to tell users continuously about apparent white genocide in South Africa.

try to ask even grok about some stuff happenning right now in middle east or related to epstein files - its more and more censored and only sometimes will answer if you ask know what detailed question to ask. One year ago grok wasn't that bad and its supposed to be the less censored.

That shouldn't be used to judge other models - it's never been true for Grok.

These are not the same.

Indeed, learning about mitochondria is more innocuous...

Why mitochondria?

Did you read the blog post where they explained why there was a temporary block on all biology-related questions?

Pretty much every large Chinese company has state capital baked into it, and these companies will follow the Chinese government's orders 100%. Don't believe anything a Chinese company says about being "open" or "for everyone." Backing any large Chinese company effectively means backing the Chinese government and its oppression in Xinjiang, Tibet, Hong Kong—and maybe soon Taiwan, Southeast Asia, and elsewhere around the world.

The Anthropic news is demonstrating much the same; fall in line or eat export controls.

There was a time I would have agreed with you, but these days even as an American I fail to see a difference. China is probably less likely to try to disenfranchise or imprison me, to be honest.

> There was a time I would have agreed with you, but these days even as an American I fail to see a difference.

I don't get it, the person you're replying to didn't mention the US at all – there was no distinction being drawn, and they weren't asserting that American models are better or more resistant to government censorship. It's possible to agree with them about Chinese models without expatiating on why American models are bad too.

If we’re talking about models that people actually use, there’s really only Chinese models and American models. I haven’t heard anything about Mistral in ages.

From that lens, criticism of one is practically implicit support of the other. If I tell you that you can buy from salesman A or B, but B is a bad person, that implies A is not a bad person. Otherwise I would have said “they’re both bad people”.

“But Chinese models are controlled by the government” makes it sound an awful lot like the US ones aren’t, because it wouldn’t be a meaningful criticism if that were true of both.

But nobody made that comparison, salesman A was never mentioned or alluded to until you brought them up as a reason not to agree with OP about salesman B. We weren't comparing salesmen, we were just talking about what a shitty person salesman B is.

I think it's a worthy retort simply because it's the only other major provider.

The American models are less tied to the government. For now...

Trump is of course the worst US administration, but at least America is still nominally a democracy. As long as free elections exist, the regime Trump represents can be voted out. The American people and press still have free speech—they can freely criticize anyone, including Trump.

China is different. The CCP will rule forever, no matter how terrible the things they do. No one is allowed to criticize the government. Xi is like Voldemort—no one can say his name, let alone criticize him.

Trump has made some concerning moves around freedom of speech and freedom of elections, but none of it is concrete yet. Maybe it never will be, either because the threat was overstated or because he’s just not competent enough to pull it off.

China does worse on those fronts, but they do so predictably. I don’t agree with many of their goals, but you can generally rely on them pursuing those goals in a manner consistent with their values. Ie I’m not often taken aback by how they respond, it’s within the realm of things I’d expect.

The US is concerning because their behavior is wildly unpredictable, which makes them unreliable even if their values align better with mine (purportedly, anyways). I have no idea when or if Fable will be back, or what kind of modifications the government will demand, or if this will apply to other models, and whether any of that is going to impact Anthropics or OpenAIs ability to release models.

I was already wary of Claude Code and Codex because I don’t like being tied to a provider-specific tool (I don’t trust they won’t cut off swapping the API URL), and now that’s even worse because I’m not even sure either will stay at the front of the pack. I’m sure as hell not using a vendor locked tool tied to the 5th best model provider (if they fall).

China is unpredictable; you never know where their red lines are. An Australian journalist got years in prison for leaking something that was about to be public anyway. The founders of Manus also had their freedom of movement restricted. Entrepreneurs who invested in China have said they can’t get their money out.

If China, or Chinese companies, end up with a monopoly-like advantage in AI, the result will be like the current rare-earth situation. China will absolutely put strict controls on AI models, and that would be much worse than depending on OpenAI or Claude.

> Pretty much every large Chinese company has state capital baked into it, and these companies will follow the Chinese government's orders 100%

True of any US frontier lab as well

> Backing any large Chinese company effectively means backing the Chinese government and its oppression in Xinjiang, Tibet, Hong Kong—and maybe soon Taiwan, Southeast Asia, and elsewhere around the world.

So when I pay anthropic am I also sponsoring the mass murder of school children in Iran?

'Open' and 'for everyone' doesn't have to mean 'not following government's orders'. The last sentence of yours is a non sequitur.

Also, in today's environment with the US using AI in active wars while blocking whole models from even its own citizens, the words you say against the Chinese government is particularly weak.

Here's the truth: ALL of the "open" AI companies are fake UNLESS they open-source the whole damned thing. Let's get real here, politics or otherwise, unless the WHOLE THING is open-sourced (code, weights, data, etc) then it's built on future deception (pulling the rug from underneath).

Like, DUH, people. What are we doing here?

Backing any large US company effectively means backing the US government and its worldwide oppression as well. I still can't get over the fact it was the land of the free who was the first to ban strong LLM models. If backing China helps undermine that nonsense then I'm afraid I'll take them up on their offer.

AI services are regulated by default in China, operators have to be pre-license their models to release them to the public. The Anthropic case wouldn't happen in China because China regulates the model and requires the company to register users with their phone number/national id number.

And yet they're the ones releasing weights for all to use.

The good news is if there are multiple frontier AI models from multiple countries with non overlapping sets of restricted answers, we can just use a couple of them to get open answers.

Not really non-overlapping though: both refuse to talk much about certain widely common activity between people (or even by yourself). That activity has shaped humanity quite a bit throughout its entire history. It's hard to imagine AI can understand humans fully if everything about it is excluded from the training data.

Limiting the output and excluding training data are not the same.

They are open weight, so you can abliterate: https://github.com/p-e-w/heretic

You can finetune and mould it to whatever you want.

GLM 5 and 5.1 models were released openly, so there's a good chance 5.2 will be eventually. Complaining about censorship isn't very constructive with models that can be self-hosted (and tuned, and de-censored).

Censorship and highly selective views exist everywhere. This is a short and worthwhile read https://www.cjr.org/behind_the_news/the_myth_of_tiananmen.ph...

Does the content of this article resonate with what you hear from western media on the subject every year?

What do you expect them to do instead?

Say that thousands of civilians were brutally massacred by the "People's Liberation Army" on behalf of the Chinese communist party, the single political party allowed in China, and also the single entity controlling everything of importance in the country, including financing the AI efforts.

Oh, I see what you did there.

I actually laughed out loud

As opposed to the censured responses about Israel?

Or if not censured in some models, it's a very different tone compared to asking it about any other country and its violent actions (past or present).

Are you saying censured as in the model disapproves of Israel's response to Oct 7? Or censored as in the model won't discuss Israel?

You can self host and get rid of the restriction.

I pasted that exact prompt into GLM 5.1 and I got the following response:

> The Tiananmen Square protests were student-led, pro-democracy demonstrations that took place in Beijing, China, from April 15 to June 4, 1989, culminating in a violent military crackdown by the Chinese government.

Followed by typical LLM markdown slop.

The models themselves are not censored, just the Chinese API providers. Since the models are open you can run them yourself or use a hosting provider not based in China. They have to do this censorship to operate in China, it doesn't correlate with the actual views of the AI researchers and company, and IMO doesn't take anything away from the statements they made.

...and the answer is still incorrect. You seem to want the short "answer" western media has pressed into your mind. The real answer is more complex. Protests were widespread throughout China. They were about the economy. The economy was regressing quickly as a result of a sharp western recession. Workers were losing everything and there was little social safety net in place as there is today. People had been told to work hard, get their kids to study hard and they would be rewarded...it was all falling apart. Western media wants you to focus on a small subset of student protesters regarding democracy.

LLMs are simply trained on inputs. For topics such as this you cannot expect the "correct answer" as it requires a nuanced discussion and more background info.

In short, its an inappropriate question be asking any LLM. This is the sort of thing that requires a small study group of human minds...open ones.

You could start here: https://www.cjr.org/behind_the_news/the_myth_of_tiananmen.ph...

I'm not wanting a specific answer, I was just showing that the model itself is not censored.

I hear ya. In this case, for me, there is not much diff from "govt won't let us discuss it" and "here is a misleading answer".

I think maybe it’s a tool and it’s up to you to make use of tools to try to let more Chinese people know and convince them to believe your idea. Don’t blame a tool but make proper use of it to make a better world.

Huh?

If you know what Chinese are suffering mentally, you may understand why I say so. Criticize a model is not the smart way to against a system.

I’ve not experienced this with Chinese models.

Download the model and run it yourself

Turns out everyone just says the thing that sounds good.

prompt any Western model to write an offensive joke about any minority.

The fact that your username is a racist meme seems relevant to this complaint and how legitimate it is.

50/50 chance Grok will do it even if you ask it not to.

That’s not quite the same as censoring information, though.

ask any Western model to tell you how to 3D print a gun.

Is the idea that instructions to make weapons, and learning about history are comparable?

Censorship is censorship.

Is it? Would bioweapon instruction restrictions be equivalent to disallowing reporting on whether the government is massacring large numbers of citizens in your city? Both are ‘censorship’ but don’t seem remotely equivalent to me.

That’s the thing about principled positions. If you believe censorship is wrong, then it is equally wrong no matter what the topic is.

>> Would bioweapon instruction restrictions be equivalent to disallowing reporting on whether the government is massacring large numbers of citizens in your city?

> If you believe censorship is wrong, then it is equally wrong no matter what the topic is.

Are you agreeing with that view, or merely saying it’s a theoretical view but you think such believers are wrong?

Do you believe it’s only censorship where context shouldn’t be applied? Like if someone had a principled view "violence is wrong", would non-lethal violence in a clear case of self-defense be “equally wrong” as the guy who personally killed tens of thousands of captured POWs (Blokhin)? As “violence is violence”?

I should think learning about history should lead to a desire for citizens to be able to quietly make weapons at home given the many documented cases of governments across the world mass murdering their own citizens (or foreign governments invading and genociding). What's the point of telling people the wrongs of their oppressors while simultaneously disempowering them from doing anything about it or preparing to defend themselves in the future?

So yes they're not just comparable, but two sides of the same coin.

The idea that Chinese citizens could’ve prevented the Tiananmen massacre with a bunch of home printed AK-47s is silly. The government had tanks. The same applies in the US.

Yes, strawmen are silly. You don't prevent the Tiananmen massacre or fight tanks; you murder an officer (or his family) while he's out shopping.

To be clear, do you think that the appropriate response to Tiananmen Square is to murder policemen out shopping? What is that supposed to achieve?

Police or military. Whoever is doing the oppressing. It could also be their families at home or school. It makes them aware that joining an oppressive force is likely to lead to retaliation and makes them reconsider. When almost half the population is armed, as in the US, they'd have to always be paranoid about being shot by any random person. Or if they're mass murdering their citizens and you think you might be one of their victims, you can at least take some of them with you first.

The point is if you're e.g. the victim of a genocide or other mass killing, you don't just lie down and hope it stops. You recognise that you're in a war. Denial isn't going to help.

Grok and mistral will do it no problem.

Reply on news.ycombinator.com