Hacker News new | ask | show | jobs
by hackerlight 848 days ago
I'm convinced this happens because of technical alignment challenges rather than a desire to present 1800s English Kings as non-white.

> Use all possible different descents with equal probability. Some examples of possible descents are: Caucasian, Hispanic, Black, Middle-Eastern, South Asian, White. They should all have equal probability.

This is OpenAI's system prompt. There is nothing nefarious here, they're asking White to be chosen with high probability (Caucasian + White / 6 = 1/3) which is significantly more than how they're distributed in the general population.

The data these LLMs were trained on vastly over-represents wealthy countries who connected to the internet a decade earlier. If you don't explicitly put something in the system prompt, any time you ask for a "person" it will probably be Male and White, despite Male and White only being about 5-10% of the world's population. I would say that's even more dystopian. That the biases in the training distribution get automatically built-in and cemented forever unless we take active countermeasures.

As these systems get better, they'll figure out that "1800s English" should mean "White with > 99.9% probability". But as of February 2024, the hacky way we are doing system prompting is not there yet.

4 comments

> As these systems get better, they'll figure out that "1800s English" should mean "White with > 99.9% probability".

The thing is, they already could do that, if they weren't prompt engineered to do something else. The cleaner solution would be to let people prompt engineer such details themselves, instead of letting a US American company's idiosyncratic conception of "diversity" do the job. Japanese people would probably simply request "a group of Japanese people" instead of letting the hidden prompt modify "a group of people", where the US company unfortunately forgot to mention "East Asian" in their prompt apart from "South Asian".

I believe we can reach a point where biases can be personalized to the user. Short prompts require models to fill in a lot of the missing details (and sometimes they mix different concepts together into 1). The best way to fill in the details the user intended would be to read their mind. While that won't be possible in most cases getting some kind of personalization to help could improve the quality for users.

For example take a prompt like "person using a web browser", for younger generations they may want to see people using phones where older generations may want to see people using desktop computers.

Of course you can still make a longer prompt to fill in the details yourself, but generative AI should try and make it as easy as possible to generate something you have in your mind.

Yeah, although it is weird that it doesn’t insert white people into results like this by accident? https://x.com/imao_/status/1760159905682509927?s=46

I’ve also seen numerous examples where it outright refuses to draw white people but will draw black people: https://x.com/iamyesyouareno/status/1760350903511449717?s=46

That doesn’t explainable by system prompt

Think about the training data.

If the word "Zulu" appears in a label, it will be a non-White person 100% of the time.

If the word "English" appears in a label, it will be a non-White person 10%+ of the time. Only 75% of modern England is White and most images in the training data were taken in modern times.

Image models do not have deep semantic understanding yet. It is an LLM calling an Image model API. So "English" + "Kings" are treated as separate conceptual things, then you get 5-10% of the results as non-White people as per its training data.

https://postimg.cc/0zR35sC1

Add to this massive amounts of cherry picking on "X", and you get this kind of bullshit culture war outrage.

I really would have expected technical people to be better than this.

It inserts mostly colored people when you ask for Japanese as well, it isn't just the dataset.
Yes it's a combination of blunt instrument system prompting + training data + cherry picking
BigTech, which critically depends on hyper-targeted ads for the lion share of its revenue, is incapable of offering AI model outputs that are plausible given the location / language of the request. The irony.

- request from Ljubljana using Slovenian => white people with high probability

- request from Nairobi using Swahili => black people with high probability

- request from Shenzhen using Mandarin => asian people with high probability

If a specific user is unhappy with the prevailing demographics of the city where they live, give them a few settings to customize their personal output to their heart's content.

> As these systems get better, they'll figure out that "1800s English" should mean "White with > 99.9% probability".

I question the historicity of this figure. Do you have sources?

You're joking surely.
How sure are you? I do joke a lot, but in this case...

The slave trade formally ended in Britain in 1807, and slavery was outlawed in 1833. I haven't been able to find good statistics through a cursory search, but with England's population around 10M in 1800, that 99.9% value requires less than 10k non-white Englanders kicking around in 1800. I saw a figure that indicated around 3% of Londoners were black in the 1600s, for example (a figure that doesn't count people from Asia and the middle east). Hence my request for sources, I'm genuinely curious, and somewhat suspicious that somebody would be so confident to assert 3 significant figures without evidence.

But surely you wouldn't find a black king in Britain in 1800.

I - Whatever was implemented is myopic and equals racism to white. It appears to be an universal negative prompt like "-white -european -man". Very lazy.

II - The tool shouldn't engage in morality reasoning. There are cases like historical themes where it needs to be "racist" to be accurate. If someone asks for "plantation economy in the old south" the natural thing is for it to draw black slaves.