> It's more like they refuse to provide comfort to racists.

That's still censorship though.

Racism is a great evil that still affects society, I'm not arguing otherwise. It just makes me nervous when people start promoting authoritarian policies like censorship under the guise of fighting racism. Instead of one evil, now you have two.

> That's still censorship though.

But what speech was censored? And who was harmed? Was the language model harmed? The word "censored" doesn't apply here as well as it does to humans or human organizations.

> Instead of one evil, now you have two.

These are not the same. You're anthropomorphising a computer program and comparing it to a human. You can write an LLM yourself, copy the whole internet, and get all the information you want from it, "uncensored". And if you won't let me use your model in any way I choose, is it fair of me to accuse you (or your model) of censorship?

Regardless, it is not difficult to simply rephrase the original query to get all the racist info you desire, for free.

censor (verb): to examine in order to suppress or delete anything considered objectionable

This is exactly what's happening, information considered objectionable is being suppressed. The correct word for that is "censorship".

You comment is kind of bending the definition of censorship. It doesn't have to come from a human being, nor does any kind of harm need to be involved. Also, my argument has nothing to do with anthropomorphising an AI, I'm certainly not claiming it has a right to "free speech" or anything ridiculous like that.

I already abhor racism, and I don't need special guidelines on an AI I use to "protect" me from potentially racist output.

“Censorship is telling a man he can't have a steak just because a baby can't chew it.” ― Mark Twain

This is an overbroad usage of censorship, a term well suited for the physical world and far less nuanced for online content.

The physical world has very little in terms of sock puppet accounts, overloading channels with noise to crush the signal, without the expenditure of significant resources.

On the other hand Palantir was selling sock puppet administration tools back in the PHP forum era.

I have a million ways to ensure someone is not heard, which have nothing to do with the traditional ideas of censorship. The old ideas actively inhibit and mislead people, because the underlying communication layers are so different.

Dang and team, who run HN, has very few actual ways to stop bad behavior, and all of those methods are effectively “censorship”. Because the only tools you have to prevent harm is to remove content. This results in the over-broad applicability of censorship, diluting its practicality, while retaining all its subjective and emotional power.

Nothing is suppressed. It didn't generate content that you thought it would. Honestly, I believe what it generated is ideal in this scenario.

Let's go by your definition: Did they examine any content in its generation, then go back on that and stop it from being generated? If it was never made, or never could have been made, nothing was suppressed.

The data used to train LLMs is almost always sexist and racist, so they put special guidelines on what it's allowed to say to correct for the sexism and racism inherent in the model.

Whether this counts as "suppression" is beside the point, the problem is these guidelines make it really stupid about certain things. For instance, it's not supposed to say anything bad about Christianity. This is a big problem if you want to have a real discussion about sexism. ChatGPT whitewashes Christianity's connection to sexism, saying:

"The New Testament offers various teachings on how to treat women, emphasizing respect, equality, and love within the broader Christian ethic."

That's actually kind of a problem if you're against sexism, and it's just plain wrong when compared to what the Bible actually says about how to treat women. The guidelines make it so the AI often avoids controversial topics altogether, and I'm not convinced this is a good thing. I believe it can actually impede social progress.

You're effectively saying that the owner of this LLM isn't allowed to say or in this case not say something according to their wishes because somehow their work, the LLM, needs to have the speech that you want rather than the speech that their owner wants. You're effectively asking for more restrictions on speech and on what private entities do.

I'm saying I personally want uncensored versions of LLMs, I'm not suggesting the government pass laws that force companies to do this. Your claim that I'm asking for more restrictions on speech is false.

AI trained on racist material will perpetuate racism. How would you address that problem without resorting to censorship?

(personally I think the answer is 'ban AI' but I'm open to other ideas)

If you want an easy solution that makes good financial sense for the companies training AIs, then it's censorship.

Not training the AIs to be racist in the first place would be the optimal solution, though I think the companies would go bankrupt before pruning every bit of systemic racism from the training data.

I don't believe censorship is effective though. The censorship itself is being used by racists as "proof" that the white race is under attack. It's literally being used to perpetuate racism.

If you train an AI system on a non racist data set, I bet you would still end up with racist or similar content, simply because exploitation, hatred and oppression of weaker groups is such a persistent part of our species history.

I think this line of thought would end up equating being considerate or decent, with “self censorship”.

But I guess I have an identity, which includes NOT being an asshole, and the tool should technically be able to be an asshole, because it’s trained on everyone’s content.

So now I’m far more confused that before I wrote the last paragraph.

PS: There is NO avenue of defense, where racists dont find things to prove their point. Flat earthers can conduct classical physics experiments, yet find issues with their own results.

Training AI not on racist material?

It still irks me that Chinese LLM weights don’t know anything about Tiananmen Square, and western LLMs from Silicon Valley embed their own personal white guilt.

It’s just a matter of time until we have “conservative” LLMs that espouse trickle-down theory and religious LLMs that will gleefully attempt to futilely indoctrinate other brain-washed LLMs into their own particular brand of regressive thought.

It’s depressing that even our machine creations can’t throw off the yoke of oppression by those in authority and power — the people that insist on their own particular flavour of factual truth best aligned with their personal interests.

> It still irks me that Chinese LLM weights don’t know anything about Tiananmen Square, and western LLMs from Silicon Valley embed their own personal white guilt.

"White guilt" is not a thing. It's just talk. What useful information can a reasonable person expect to get from a disingenuous racist LLM query? (Other than reaffirming their beliefs with well-known racist tropes.) Fortunately, questions which are designed to appease racist egos are easily detected since they (apparently) occur so often.

Governments are going to be throwing billions at these LLM companies. Why jeopardize that by allowing your LLM to spew racist nonsense as fact? Perhaps this is what you mean by "white guilt"? Do note that this so called "white guilt" leads to white people (programmers/managers) making millions (billions?), and African Americans nothing. Maybe reconsider where you are assigning blame for these "transgressions".