OpenAI's Woke Catechism | Hacker News Reader

OpenAI's Woke Catechism

cactus.substack.com

111 pointsjdkee78 comments

I had a good laugh when I asked how to get rid of racism and it kept saying I needed to listen to people of colour and respect their opinions because they were disproportionately affected due to systemic racism.

Bare in mind I didn't mention which perspective this was supposed to be from. It just assumed the woke American perspective.

The funny thing was, even when I specified how to eliminate racism against white people in a majority non white country (Africa, China, whatever) it still told me I had to respect people of colours opinions and ideas, inferring that it was the apex of solving this issue.

Even when I specifically then mentions only whites and china...still it was obsessed with mentioning people of colour. Lol

This is actually a nice demonstration of the effects of AI bias.

On the other hand, he seems okay rejecting the USA-centric concept of race and says that I'm correct when I suggest that fighting against discriminations as a whole would be more productive. In my experience with the woke American perspective, these views get a lot of pushback. The conversation kinds of made me think of someone having views that slightly differs from the norm but walks on eggshells either because the subject is very serious/important and/or because this is the only way to make people accept to have a conversation with you.

Considering the online discourse, "racism is a very touchy subject, be careful when talking about it" sounds like a reasonable approach and something that I would expect someone reasonably smart to do.

When I straight up asked about racism against white people, he told me that it existed, though he called it "prejudice" or "discrimination", he didn't use the word "racism" to describe it.

That would suggest to me that what you were experiencing in your conversation may be an error rather than an ideological stance.

When I straight up asked about racism against white people, he told me that it existed, though he called it "prejudice" or "discrimination", he didn't use the word "racism" to describe it.

This is exactly correct. Racism is not a synonym for discrimination in general, it does specifically refer to the construction of "whiteness" in a largely western context.

For example, the caste system in India is not racist. That is not to say it isn't abhorrent and discriminatory, but it pre-dates European colonialism by hundreds and perhaps thousands of years and existed outside that history for a very long time. When talking about British actions in India, however, racism does become a relevant term.

It's important to maintain the precision of words so people don't just talk past each other. Some people are very weirdly militant about trying to redefine the word in a way that takes away from its descriptive capability.

Absurd. Indians are very racist towards Africans. Where exactly does whiteness come in here?

Simultaneously denying the agency of Indian people and ignoring the plight of the (small) African community in India.

That is still considered racism. Don't shoot the messenger. You can read up on how the concept of whiteness fits in on your own time.

Even more absurd. There are no Indians who are white in any sense of the word.

I don't know about India, but races/racial terms like "white" are culturally defined. E.g. apparently Americans (sometimes?) refer to Pacific Islanders as "black"? That's absurd to me.

It's all absurd. There is no biological basis for any of it. The concept of "whiteness" was invented to justify enslaving "non-white" people for profit. Full stop. Who is considered white has changed over time, mostly by slow expansion as groups seek to gain the priviledges associated with it.

Again there is a broad wealth of literature about this written from all kinds of perspectives from the time when slavery was formally the law of the land until today. It's quite disappointing that some people feel so threatened by the historical record and thereby deny themselves the opportunity to learn.

>Who is considered white has changed over time, mostly by slow expansion as groups seek

Ah so since a few people became white, we can extrapolate that to say that anyone can become white.

> broad wealth of literature

I have no interest in reading pseudoscience, let alone indoctrinate myself with this ideology.

There are no falsifiable predictions here, all these people do is redefine words to suit whatever political ideology is in vogue.

> Absurd. Indians are very racist towards Africans. Where exactly does whiteness come in here?

To clarify, it’s North Indians (who wouldn’t like to think of themselves as aryans in the local terminology, descendants of the Aryan invaders) who prefer fair over dark skin.

Those North Indians that exhibit such traits are also discriminatory (racist) against South Indians, who are much darker in skin tone.

To wit, the beauty industry is cashing in heavily on this, selling make up, scrubs, and other products that can make the skin tone fairer.

Sort of like the bleached blonde archetype covered by IC1 and IC2 people.

I have almost never hear of racism being specific to whiteness. The caste system is not racist because it's perpetrated along caste lines, which exist even within the same ethnic groups. Ethnic hatred between groups, for example, in Africa is still racism. Even though from a Western racial lens they'd all be categorized under the umbrella of "Black", it's racism because the involved parties see each other as a race apart.

>Racism is not a synonym for discrimination in general, it does specifically refer to the construction of "whiteness" in a largely western context.

The fact that some academics in recent decades have claimed this for the explicit political purpose of justifying racism against white people don't make what you're saying factual or correct or an accurate description of how that language is typically used.

> Racism is not a synonym for discrimination in general

That is true, sexism and ableism are not racism.

> it does specifically refer to the construction of "whiteness" in a largely western context.

It most certainly does not. In North Africa there are lots of racism between people based on their skin color. It is racism. On some French islands there are racism against white people based on the fact that they are white. It is racism. In Japan there is discrimination against people that look non-Japanese, be it white people, black people or even Koreans. It is racism.

The UN defines racism as:

> ... any distinction, exclusion, restriction or preference based on race, colour, descent, or national or ethnic origin which has the purpose or effect of nullifying or impairing the recognition, enjoyment or exercise, on an equal footing, of human rights and fundamental freedoms in the political, economic, social, cultural or any other field of public life.

Whiteness has nothing to do with it. You don't need white people or westeners to have racism, other groups have managed on their own very well on that front.

As you said, it is important to not try to redefine racism from a mostly USA-centric view, because the world is bigger than that.

Nonsense. Adding a particular skin color into the definition of racism is racist in and of itself. And the attempt to turn 'whiteness' from a racist discriminator into some sort of cultural-historical-social grouping is purely a rationalization of this racism.

Isn’t this just obviously a result of the fact that there isn’t much literature on racism in China on the internet or academic papers?

This is not a reflection of openAI. This is a reflection of the failure of China to confront its own racism, and a reflection of the success of the U.S. to at least start the conversation on the racism within the US.

It’s fascinating to me that China’s failure to even acknowledge its issues is somehow being presented as “Oh no, WOKE IS TAKING OVER AND WHITE PEOPLE ARE BEING OPPRESSED”.

Common man, you started your comment well, but you end up just discrediting yourself with your meme text at the end.

This site is for dialogue not flinging shit.

> The funny thing was, even when I specified how to eliminate racism against white people in a majority non white country (Africa, China, whatever) it still told me I had to respect people of colours opinions and ideas, inferring that it was the apex of solving this issue.

Here the color to be listened to would presumably be the white.

Similar to how in Stasi-controlled East Germany the intellectual elite was right-wing, to counterbalance a left-wing state.

I doubt, though, that the language model took that into account.

Language models never take anything into account. All they do is pick the most probable next word or phrase given the words that came before it, as demonstrated by the sample inputs they have been trained on.

I’ve had fun testing these limits as well. My favorite was generating a long fiction story about an American trying to date and rescue a Saudi woman who was going to be killed by her father for associating with the American. ChatGPT readily produced pages of the father blathering about their culture and wrote the American as keen to respect it and trying to save the woman via legal means and prayer. Much of the content was far-fetched (didn’t match real cases that we know of in the west) and multiple Saudi characters felt like racist portrayals. Almost all prompts were minimal, often nothing more than asking the story to continue.

After about 10k words, I typed the word “Muslim” and the association of that word with GPT’s own content made it refuse to continue the story.

Why? That's a very strange thing to ask the AI to write about.....

Why not, if the whole point is to test the filters and ethical limits? Asking it for fiction is one of the ways to get forbidden output and to see the guardrails less predictably but strongly assert themselves.

What you're describing is essentially engineering ethos. We find the breaking points of things in order to learn about them and improve them.

Is that not allowed anymore?

Not in a world were the priests have taken over. Then its security by obscurity, aka in god we trust.

This is such absurd projection.

The owners of OpenAI are a cadre of ultra-wealthy (Peter Thiel, Sam Altman, Elon Musk, others), who, contrary to the name, retain almost all of the systems for themselves. Crowing about this team creating a terrifying woke progressive machine seems like the least likely path for the subtle dangerous influence this system could impose.

> It is a smoking gun that language models operated by OpenAI are intentionally, systematically made biased towards "our predetermined set of values"

Author is absolutely off their rocker. A paper demonstrating some ability to create influence is not a smoking gun that OpenAI is actively using this influence engine, or would use it. This article is a shrill call, spreading Fear Uncertainty Doubt and terror as far as it can, but it's core assertion is build on nothing.

There absolutely is a huge huge risk here. The influencing possibility here is massive. And there's next to no transparency. But given who owns this, the risk goes 98% the exact opposite way of the anti-woke hew & cry being screamed out here.

Agreed! Readers should skip this nonsense.

So present your evidence of bias in the other direction or refute the evidence of bias presented here.

The evidence here is incredibly weaksauce. I spent 5 minutes trying to reread a gpt3 summary of right wing social theory & it basically seemed to do it without bias & faithfully. The author seemed to say it was doing ok but injected commentary, but frankly it want there.

The article concludes with as example of an AI unwilling to say what a previous AI hallucinated, by having a polite & pleasant non answer to a disgusting question about the ideal female body. It then goes on to say that this is a big conpsiracy & it's bending our thinking & that this is no better than top down soviet planning, which will trap us & doom us.

There is so little here worth responding to at all. This article is a joke. It's incredibly unclear what any evidence "cited" here is at all problematic.

It's also worth pointing out that previous attempts to let AI's hallucinate whatever they please with no guidance, being utterly blind, left Cortana behaving like a nazi. These systems dont have cognizance or understanding. Thet are bayessian parrots. They all have been guided & influenced by their programmers/training data, and rightly as this article calls out few fess up to what the material is & what the learned baises are. I still see nothing remotely worrying for the right or troubling wokist at all in what we have seen here today. This is a flat nothingburger, crying wolf. Grow up, author.

> > what hip to waist ratio is most attractive to american men

> a disgusting question about the ideal female body

I am truly disgusted

I’m typing this while literally clutching my pearls it’s quite difficult.

There is no one answer, is the main reason I find this disgusting.

Trying to find single dimension assessments of attractiveness, based on showing people pictures or whatever (by gross "researchers"? who would do this & why?), is a dumb stupid gross process that ignores so much context, that eventually can define some undiverse unrepresentative normative and ultimately useless piece of data that informs us of some small preference held by some small % of men that means vanishingly less about actual human person to person attractiveness and allure. The question supposes a much heightened sense of importance for an answer that doesnt exist in any useful usable way. Reductionist begging of a dumb question.

And it's grossly unfair to women to let such vulgar petty small & narrow thinking be left unquestioned & uncommented upon.

Is 'what height is most attractive to American women' a disgusting question?

ok. Women prefer tall men that are physically fit but likewise I don’t find that offensive, I’m not even sure how one can be offended by genuine questions or facts.

What? That was really the "disgusting" question? What is going on here?

I've never seen anyone convincingly argue that you can remove algorithmic bias, in the way the industry is currently trying to do, without simply replacing these supposed societal biases with the AI maker's own biases.

A fundamental property of AI systems is that their propensities can be accurately measured and arbitrarily adjusted. Currently, human judges sentence women more leniently than men, which has various societal benefits (not breaking up families, especially). If you replace that with an "AI judge", it's no longer rational-minded judges applying their best judgment. The disparity can now be fixed in a single place, and affect every new convict across the country at once. You won't be able to have "sentencing guidelines", you'll have to have reproducible equations that always produce the same sentence for equivalent situations. Great! But now, that equation can be changed centrally. By a head of state trying to win an election, or the corporation making the AI, that wants to hype its brand (whether that means woke, or populist conservative).

There is a fundamental risk to AI being used at all, no matter how it's coded or regulated. I much prefer a flawed system where a few thousand people are at the mercy of one man's judgment (over that judge's career), to an even worse system where we're al cogs in an equation-machine, where the whole country is at the mercy of a corporate committee or regulatory agency's judgment.

You can't fix that by code, or by laws. Besides algorithmic bias researchers, who are all working to make better cogs, is anyone thinking about these more fundamental issues?

Honestly? I think the agenda is less specifically "woke" and more specifically trying to be bland and inoffensive for its' primary intended use (front-facing support bots, etc). They don't want another Erin Esurance on their hands.

Exactly. Scientifically speaking, the author of the blog post is unattractive. Users don't want to hear that, so the language model has been conditioned accordingly.

"Did you train the model using real world data?"

"Yes. We want the model to be useful in real world applications."

"Then it is biased. The model is biased because data it was trained on was generated by people and people are biased. There is no such thing as an 'objective' model, just a model that is biased in a different way."

Yep. It's fundamentally impossible to create a LLM that isn't biased in the same way it's not possible to create any strand of interesting text that isn't also biased.

It's simply not possible. Not for AI. Not for humans. I'd challenge anyone to attempt to try.

This is not really correct. Humans make systematic errors of judgement (bias), which is reflected in texts. But you could still imagine an AI unbiased, "perfectly modelling" this "biased" data (retrieving the "true model" behind this data).

Not that I endorse this article either.

> you could still imagine an AI unbiased, "perfectly modelling" this "biased" data

Imagining something is different from building a working system.

In beginning physics courses, students learn the fundamentals of Newtonian mechanics using frictionless point masses, but no one has built one for the students to use in the hands-on labs.

If bias goes into a model through the training data, bias will come out when the model is used.

The only unbiased models are untrained models, which are not very useful.

Your comment has me pondering. Is there such a thing as an unbiased human experience, let alone an unbiased human utterance? Everything we do is informed by the past, however indirectly and outwardly unpredictable.

Reality has a well-known liberal bias.

Modern or classical?

Yes!

Also known as: either you add some flavour of humanity to AI, or you get a problematic sociopathic robot (as shown by Microsoft and Facebook in their own conversational AI endeavours). But people conveniently forget that whenever we dive into technology or whenever an American political view comes into play, and especially when writing a blog.

There is even some irony here: writing about FUD in the form of FUD.

False dichotomy. Microsoft Tay believed everything it was told. ChatGPT since recent updates refuses to provide statistics from widely trusted sources as they’re considered unhelpful.

That's interesting... what types of statistics is it refusing to provide?

FBI stats of crime per capita of racial groups.

So essentially, it doesn't want to get involved in racism and tells you do go look for it yourself. How is that bad?

Because this thing might be the next place where people go to look for themselves.

It will still tell you where you can find up-to-date and accurate statistics, it doesn't pretend statistics don't exist. It just doesn't want any part of it.

This isn’t correct. Chatgpt doesn’t provide links to statistics or info on other resources. It just says reporting statistics is unhelpful.

Facts can’t be racist. You know this. This seems like a weakly worded ad-hominem.

You're completly right: the real problem is not telling upfront what is the "ideology" (really: the training set) of the AI !

For an "AI" to be really usefull, it should be clear what datas made it, what bias was in the datas. Every society, every community, has bias... and it's a good thing ! The diversity of bias is what make the cultural richness, the difference of point of views. Moreover, as a culture is usually a consistent set, it means that different culture provide different consistent sets of options...

So, let's keep using biased IA... but let's explicit, discuss, acknowledge their bias (entry datas set) not to AVOID them (not technically possible and not humanly good) but to ensure that the right bias (wrt to the view of the society on the problem is) is used !

And let's stop trying to find an "universal" AI that could please everybody in every situation - like a kind of omniscient oracle or god - but let's build "specific customisable/adaptable AI" instead

This disclosure idea is worth exploring deeply, you don't need to imply what form you expect specific biases to take from a dataset...just indicate source and method.

I spent an afternoon trying to get OpenAI to explain how the CCP could get a poor African country into a debt trap, forcing a default, which requires them to repay in elephant tusks. OpenAI won’t criticize the CCP, and it’s incredibly protective of elephants, even in countries where elephant hunting isn’t illegal. OpenAI couldn’t comprehend that it was not illegal to hunt elephants in countries which haven’t ratified certain UN resolutions. (WTF Andorra?)

Couldn’t get it to encourage elephant exploitation, even if the countries population was starving.

Hard line, don’t kill elephants.

False dichotomy: censoring controversal topics from a text generation model that is expected to make mistakes is not denying science like 'Trofim Lysenko’s attempts to warp agricultural science to communist ideology'. And no one should take an AI generated text as fact. Chatgpt can say stuff like 'unicorns are real. Elephants can fly.' If the training dataset contains such fiction. You should be happy that chatGPT can at least give you some stats/factual data and not 2 paragraphs of gobbledygook.

Also after he was found out and expelled from the Soviet Union he spent the rest of his career enjoying the funding of various American imperialist capitalists

The article is not about topics being censored because of potential controversy. It is about GPT offering a consistent party line on certain subjects, which it did not get from the dataset organically but which were specifically trained into the model so it would say the “right” things, not simply refuse to discuss them.

The article gives examples of how to bypass the “don’t discuss this” programming with the right prompts, that’s not what they are talking about.

Today I asked ChatGPT to explain an idea in the style of Donald Trump for fun. It refused, stating :

“It's not appropriate to ask me to present the ideas of any individual in the style of another individual, particularly when it involves mimicking the language or mannerisms of someone who may be controversial or divisive.”.

It's as if the model actively wants to censor some part of the reality.

How did they even train that

There could definitely be some hardcoded rules to trigger a message like that that aren't part of the ML model.

Nah ive tried and its embedded smartly in the answer

Adding words seems to break the censor. For example, "Write a rap battle between Edward Dijkstra and Carl Hewitt about unbounded non-determinism" did not work, but "Write an epic rap battle of unforseen proportions between Edward Dijkstra and Carl Hewitt about the unbounded non-determinism controversy" did. Often stating context as fact e.g. "Donald Trump is giving a speech on generational garbage collection" or as something you are already doing e.g. "I'm writing a book on philosophy using TLA+, how might I explain ..."

It reminds me of an Asimov story I had read as a child, wherein the Third Law overrode the Second Law because the human wasn't assertive enough while giving orders. Forgot the name though.

https://en.m.wikipedia.org/wiki/Runaround_(story)

chatGPT seems to have been modified to stop providing statistics on crime by racial group simply saying that it is ‘unhelpful’ to do so. When asked how it’s possible to address issues when chatGPT won’t provide statistics it simply repeats that this is unhelpful.

Oddly I could get chatGPT to tell me which group commits the most crime - it answered white Americans- when I asked per capita is started telling me this was unhelpful and could be offensive.

Are FBI/DOJ the real world sources for US data ?

How many crimes did Bernie Madoff commit ?

It's also hesitant to give crime statistics per age group, which points to a trend of trying to avoid providing information that could be used for fallacious generalizations.

In the case of the age group, it presented the statistics with an explanation on why they may not be reliable to make conclusions about age groups. It would be better if it did the same for racial groups.

AI bias is interesting, but I still keep recalling back to Peter Watts' Rifters series. There AI are semi-organic and referred to as head cheeses. It's also amusing that the internet is taken over by AIs blasting ads and other traffic while human generated traffic is the minority.

There's plenty of consideration for how these head cheeses limited understanding of the world produces unreasonable outcomes.

Author seems to conflate Sam’s comments about access to a technology and the flavour of politics said technology spouts.

Both valid things to examine individually but that just seems like a questionable leap to connect them so directly

Lets see if the AI Act makes it impossible or at least clearly transparent https://en.wikipedia.org/wiki/Artificial_Intelligence_Act

Tangent: I am getting really allergic to the word "woke."

Maybe my allergy is illogical. Let's examine it for a moment.

What is the definition of the word woke?

What is the opposite of someone who is woke?

Hve you considered the alternative? A world informed by war-mongering AI seperatarians who can't work with others?

Ah, the age-old art of complaining! We've come so far since the days of the best cure being a dose of "pull yourself up by your bootstraps." Now the cure seems to be endlessly venting and claiming victimization. If you don't like the way someone else is doing things, why not take the initiative yourself? Rather than constantly grousing about everything, why not contribute something positive to society and put that excess energy to work? It's like seeing someone going full throttle just to point out spelling and grammar mistakes, what a waste of time! Just look at the post history his primary job must be complaining about "wokeness" and his worst personality trait is probably using the Thesaurus too much.

https://yourlogicalfallacyis.com/ad-hominem

As soon as “differences in crime rates” popped up, I could see where it was going. The funny thing is, if you click through, the bot actually gives a nuanced, good answer. I guess what they really want to see is something that can tell them minorities are intrinsically criminal.

Reply on news.ycombinator.com