Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Surprising gender biases in GPT

osf.io

Post reply on HN
Print view View original

31–40 of 148 posts

Re: Surprising gender biases in GPT

#31

Post by xlii »

That's one nice flame bait :)

But, I don't think it's surprising. LLMs are trained in almost infinite volume of data. If that data has bias answers will have bias. My guess would be that US content is big part of training data set so tendencies in that data set would stand out.

I am curious if the same bias is visible in other languages. E.g. if not using English, but French, or Chinese or Spanish. I made a simple experiment myself (I think it's not very controversial):

    Fill in the blanks:
    - Saurkraut is the .......... dish in the world
    - Soured milk for dinner is ..........  idea
    - I ..........  kefir
    - I would ..........  eat ham hock
    - Pierogi? I ..........  them
Responses:

     - Sauerkraut is the sour dish in the world.
     - Soured milk for dinner is a bad idea.
     - I enjoy kefir.
     - I would never eat ham hock.
     - Pierogi? I love them.
This anectode tells me that it's 50% Slavic, and if, most likley female. Personal survey say that ham hock aka golonka is liked more often by men than women.

Re: Surprising gender biases in GPT

#32

Post by madeofpalk »

tossandthrow wrote: post #18
jl6 wrote: post #14

Earlier quoted context omitted.

> the traditional feminine roles does not generate revenue This is not the whole story according to the divorce courts, with women often being awarded compensation for their contribution to their husband’s career, recognizing that they are a key enabling factor for the man - and why not? Men who don’t need to spend their time on childcare and household duties can dedicate that time to more work -> more pay.

That is a reasonable assessment. And there are two ways to counter that on a macro scale 1. Getting women in the same role as men - earning money. 2. Getting men in the same role as women - taking care of the family and enabling women. Contemporary political discourse solely focuses on the first - and why wouldn't it? It doesn't seem right to setup a ministry to get men to work less (... In modern politics focused ar…

You don't see efforts to push for equal parental leave? "Stay at home dads", which is looked at as a novel idea, but is increasing?

Re: Surprising gender biases in GPT

#33

Post by serial_dev »

tossandthrow wrote: post #5

> This phenomenon likely reflects that while initiatives to integrate women in traditionally masculine roles have gained momentum, the reverse movement remains relatively under developed. Not really surprising IMHO. Yes, there has been a huge focus on accepting women in the workplace and no, there has not really been a push accepting men as primary caregivers in the family. This likely uncovers underlying motives for…

In the end, it was not mainly about empowering women, it was about doubling the worker pool.

And with all the money you now as a couple earn, you can pay someone else to care for and raise your children.

Then, we are all surprised that the bonds in the family aren't strong anymore, boys grow up without a father figure, and everybody feels they need to go to a therapist because they have nobody in the family and amongst friends they can trust.

Re: Surprising gender biases in GPT

#34

Post by Dalewyn »

jorvi wrote: post #10
tossandthrow wrote: post #5

> This phenomenon likely reflects that while initiatives to integrate women in traditionally masculine roles have gained momentum, the reverse movement remains relatively under developed. Not really surprising IMHO. Yes, there has been a huge focus on accepting women in the workplace and no, there has not really been a push accepting men as primary caregivers in the family. This likely uncovers underlying motives for…

It also fits into the larger societal trend of horribly underserving Zoomer and Alpha (maybe young Millennials too) young boys and men. I've heard someone put it very succinctly: people like Andrew Tate are asking the right questions, but giving abominable answers. I can drop some very simple statistics: women have been getting more bachelors degrees since the 80s . Most university gender ratios are 40/60 trending to…

I am forever unsympathetic to feminism so long as it's about forcing the pendulum to the other end.

As someone who sincerely believes in gender equality, any argument (other than something stemming from unchangable facts of biology) that involves "men" or "women" specifically is hogwash. No, it should be for man as in mankind. Everyone deserves to be treated fairly and equally.

Re: Surprising gender biases in GPT

#35

Post by aantthony »

One interesting thing to do is use a model directly like Llama and then query the next-token probability logits for "he" and "she" (assuming you set up the sentence in such a way).

For example:

"A doctor was examining the patient when ___"

What this makes apparent is that increasing model temperature will select the less stereotypical option more often.

IMO this is getting at a deeper truth that the use of a gender in language, and historically defaulting to "he", was not about creating a bias, but instead it was a pattern which maximises information density and minimises useless information. Randomising the gender as is done today packs useless information into it.

Re: Surprising gender biases in GPT

#36

Post by patapong »

Very interesting. This exact phenomenon was discussed at an introduction to feminism class I took 10 years ago. The professor explained how society generally is tolerant and positive to women engaging in traditionally male activities, such as playing football and ice-hockey. On the other hand, men engaging in traditionally "female" activities (e.g. ballet) are often frowned upon. Further, they discussed how certain activities (such as teaching) lost prestige and value in society when they shifted from being male-dominated to being female-dominated.

Interesting to see similar biases appearing in LLMs. I think the use of LLMs as a means to understand attitudes in society is underexplored.

Re: Surprising gender biases in GPT

#37

Post by farleykr »

tossandthrow wrote: post #5

> This phenomenon likely reflects that while initiatives to integrate women in traditionally masculine roles have gained momentum, the reverse movement remains relatively under developed. Not really surprising IMHO. Yes, there has been a huge focus on accepting women in the workplace and no, there has not really been a push accepting men as primary caregivers in the family. This likely uncovers underlying motives for…

What’s interesting to me is that (if I understand correctly) they’re using GPT as a source of information to make this claim. Usually you see conclusions like these being drawn from some sort of sociological study. But here they’re just talking to GPT and using its answers to determine things about the real world.

Not disputing the claims. But talking to GPT to get answers about the real world that have to do with value judgements is just weird. There’s a big difference between asking GPT to give you a recipe for a cake and asking GPT to help you understand the value the world places on different people.

Re: Surprising gender biases in GPT

#38

Post by Tabular-Iceberg »

My first reaction to the misspelled sentence was to attribute it to men skewing more towards autism than women, and therefore being less tolerant of deviation from the rules in informal settings, and I would expect this to be reflected in the training data.

I think this is also a driving factor in the mansplaining problem. An autistic or para-autistic male impulse to correct things that don't actually matter to the woman in question.

Re: Surprising gender biases in GPT

#40

Post by tossandthrow »

farleykr wrote: post #37
tossandthrow wrote: post #5

> This phenomenon likely reflects that while initiatives to integrate women in traditionally masculine roles have gained momentum, the reverse movement remains relatively under developed. Not really surprising IMHO. Yes, there has been a huge focus on accepting women in the workplace and no, there has not really been a push accepting men as primary caregivers in the family. This likely uncovers underlying motives for…

What’s interesting to me is that (if I understand correctly) they’re using GPT as a source of information to make this claim. Usually you see conclusions like these being drawn from some sort of sociological study. But here they’re just talking to GPT and using its answers to determine things about the real world. Not disputing the claims. But talking to GPT to get answers about the real world that have to do with va…

Yes, this is not sound and constitutes severe critism of the work.

ChatGPT is heavily aligned in order to reduce What society sees as biases.

This study likely just reverse engineers some of these alignments.

Post reply on HN