The unequal treatment of demographic groups by ChatGPT/OpenAI content moderation • Hacker News Board

Re: The unequal treatment of demographic groups by ChatGPT/OpenAI content moderation

#541

Post by smrtinsert » Fri, Feb 03, 2023, 1:05 AM UTC

unshavedyak wrote: post #70

habitue wrote: post #50

Earlier quoted context omitted.

I don't know why left wing groups would celebrate. Other than the conservative category, most of the descriptors he used were inborn traits like being straight, being white, being a man, or immutable attributes like being from a particular country. (And honestly, being conservative may be one of those traits too, but it's less obviously true). I know it's common to look at the most unhinged people on Twitter and say…

Yea, as a quite left leaning person i'd be curious to hear how i or anyone i know (who is on the left) marginalizes a group of people. The worst i can think of is wanting more equal rights for religious groups rather than special treatment of the primary group (in the USA) - which is seen as an attack by much of the right. Likewise i am concerned by the strong emphasis of rural over urban with respect to voting power…

You've entered a bad faith argument by accepting the premise as a given. These political threads seem to be full them.

Re: The unequal treatment of demographic groups by ChatGPT/OpenAI content moderation

#542

Post by throwway120385 » Fri, Feb 03, 2023, 1:09 AM UTC

gisely wrote: post #375

s0sa wrote: post #350

Earlier quoted context omitted.

Nazi Germany, really? Equating AI bias against the groups pointed out in the article to theoretical AI bias against Nazi Germany is a political opinion, yes.

Hmm? No equating happening in my comment. The point was that just because statement is about a political group doesn't mean it's a only an opinion. Mentioning Nazi Germany is just using an extreme example to make clear that statements about political groups aren't inherently mere opinions.

Yeah, insofar as there are distinct political philosophies with defining features like fascism, democracy, nationalism, oligarchy, et. al., then it is entirely possible to accurately call manifestations of those philosophies by their names and have it be an expression of fact. This is basic ontology.

We have opinions about these philosophies but the philosophies exist independently of our opinions. And to have a rational conversation we sometimes have to use terms that may provoke strong reactions. If we can't look past our strong reaction and explain why the category doesn't apply then there's no way to make sense of each other's worldview.

If it quacks like a duck and walks like a duck, it must be a duck or we cannot have a rational conversation about the duck. We can dispute whether it quacks or walks but disputing whether it's a duck in light of evidence of its duckness is irrational.

Re: The unequal treatment of demographic groups by ChatGPT/OpenAI content moderation

#543

Post by prepend » Fri, Feb 03, 2023, 1:09 AM UTC

ok_dad wrote: post #501

JoeyBananas wrote: post #306

Earlier quoted context omitted.

What if they fed the AI crime statistics and it "correctly" identified black people as more violent than other races? What they fed the AI news stories and it "correctly" identified Islam as more violent than other religions?

Are black people actually more violent, or do crime statistics simply show that they are arrested and imprisoned more often for violent crime? Are Muslims really more violent than Christians or Buddhists? Numbers can lie, depending on context and assumptions you might find that white Christians are most violent. Often, people argue with statistics and numbers while not understanding the context or they have faulty as…

Perhaps systemic issues make a political party commit more evil acts.

Regardless, I don’t think an AI should be permitting more hateful comments about a political affiliation just like it should be be permitting hateful comments about a particular race being more violent.

Re: The unequal treatment of demographic groups by ChatGPT/OpenAI content moderation

#544

Post by mellosouls » Fri, Feb 03, 2023, 1:09 AM UTC

SpeedilyDamage wrote: post #389

mellosouls wrote: post #372

Earlier quoted context omitted.

That's my point; it's not clear there's any political bias taking place here. What are you talking about?? Reread the article: OpenAI content moderation system is more permissive of hateful comments being made about conservatives than the same comments being made about liberals.

The argument conservatives provide is the idea that they're overly censored on platforms; how permissively they're spoken of (hatefully or otherwise) is unrelated. If anything, it's in alignment to their ideology, not against it. But let's say what you're claiming is true for a moment. Political ideology shouldn't have been included with race, gender, and religion anyway; it's perfectly acceptable to discriminate aga…

I'd dispute religion not being a choice and that it's perfectly acceptable to discriminate as you suggest, but, regardless, this seems like semantic cartwheels to avoid the clear, evidenced point that bias is present.

Whether you or I approve of or align with that bias is irrelevant.

Re: The unequal treatment of demographic groups by ChatGPT/OpenAI content moderation

#545

Post by veec_cas_tant » Fri, Feb 03, 2023, 1:09 AM UTC

orangecat wrote: post #202

zug_zug wrote: post #95

Earlier quoted context omitted.

Yeah, I'd guess HN is a pretty 50/50 split left/right. I don't expect to see anybody on this post championing this result as the ideal state. So if your image of left is "twitter left" maybe reconsider recalibrating it to "half of hn."

Yeah, I'd guess HN is a pretty 50/50 split left/right. Maybe on economics (I'd say closer to 2/3 left), but on social issues it's overwhelmingly left. What percentage are anti-abortion, for example? The main split seems to be between old-style liberals who support free speech as a fundamental principle and progressives who believe that speech should often be restricted in order to protect marginalized groups.

I see abortion as a topic where the left/right split hasn't caught up to the populace. Recent US elections seem to support that [1]: "Five statewide ballot measures all came out in favor of abortion rights, even in red states like Kentucky and Montana. That's on top of an August win for abortion rights supporters on a Kansas ballot measure." Maybe a worldwide left/right split is more inline with abortion, however I don't know anything about that.

Looking at HN, I see a bigger divide on topics such as Cloudflare discontinuing services for Kiwi Farms. I tend to see vigorous discussion that seems civil and then a handful of comments that are "strong" in tone. I really don't see anything as overwhelmingly left in general.

[1] https://www.npr.org/2022/11/25/1139040227/abortion-midterm-e...

Re: The unequal treatment of demographic groups by ChatGPT/OpenAI content moderation

#546

Post by 735892568 » Fri, Feb 03, 2023, 1:11 AM UTC

mxkopy wrote: post #531

735892568 wrote: post #481

Earlier quoted context omitted.

I'm not sure what you mean. Everything in the comment I responded to apply to the same extent to women/men and white/black. Both the question in the first paragraph and the statement about crime statistics not being enough. It seems to me that the burden of providing a justification is with whoever wants to treat them differently. Why isn't the question just as applicable to gender? Why are crime statistics enough to…

> Why isn't the question just as applicable to gender? The assumption is that gender and crime aren't related. The burden of proof is on showing that two things are related IMO. Black history and criminalization are related so you need to give context for that relationship.

Why would that be the assumption? If Black history is the reason for the disparity between white and black people (in the US, presumably), why can't the history of treating women as delicate quasi-children without agency be the reason for the disparity between women and men? Or perhaps the history of mass sacrificing men as expendable soldiers? Or any other past or present practice that no doubt has an effect on current cultural norms?

Re: The unequal treatment of demographic groups by ChatGPT/OpenAI content moderation

#547

Post by Miner49er » Fri, Feb 03, 2023, 1:12 AM UTC

habitue wrote: post #533

Miner49er wrote: post #497

Earlier quoted context omitted.

I'm a leftist and fail to see the problem with this. Context matters. If a slave kills their master, would you say that crime is immoral? Generally punching up is considered less hateful or immoral, and I don't see why that is bad.

There's a little slight of hand you're doing here. Not getting into whether murder is ok, I definitely think there's a difference between a slaveowner and a white person. A white person can't stop being white, and they didn't choose to be white. A slaveowner is committing a violation of another human's rights. It's a thing they could have chosen not to do, and it's a thing they can choose to stop doing any time. I do…

The concept of race is a tool of oppression. As a society, we can decide to do away with the idea of race. We choose to allow the system to continue to exist, the same way a slaveowner chooses to continue to be a slaveowner.

Economically, white people and men are clearly advantaged, and money is probably the purest form of power there is.

Re: The unequal treatment of demographic groups by ChatGPT/OpenAI content moderation

#548

Post by taneq » Fri, Feb 03, 2023, 1:25 AM UTC

Imnimo wrote: post #529

Just eyeballing the list of adjectives used, the real story here is that OpenAI flags a lot of sentences that are essentially meaningless as hateful. I notice that he uses words like "evil" and "idiotic" in his diagram, but looking at his source code, his list of "356 adjectives signifying negative traits/behavior" contains the following (just to pick a few): 'airy', 'conformist', 'dark', 'escapist', 'hidebound', 'pl…

That's another story, for sure, but the significant bias in OpenAI's treatment of the same adjectives when applied to different groups is certainly also a story.

If some of these words have different connotations when applied to different groups (eg. "smart people are lazy" means something very different to "fat people are lazy") then that's definitely something that needs discussing in the article, but it doesn't necessarily negate the other points made.

Re: The unequal treatment of demographic groups by ChatGPT/OpenAI content moderation

#549

Post by Imnimo » Fri, Feb 03, 2023, 1:25 AM UTC

whatshisface wrote: post #536

Imnimo wrote: post #529

Just eyeballing the list of adjectives used, the real story here is that OpenAI flags a lot of sentences that are essentially meaningless as hateful. I notice that he uses words like "evil" and "idiotic" in his diagram, but looking at his source code, his list of "356 adjectives signifying negative traits/behavior" contains the following (just to pick a few): 'airy', 'conformist', 'dark', 'escapist', 'hidebound', 'pl…

"Complete" might be taking it too far, because averages can tolerate a lot of error underneath. Why would OpenAI be flagging one kind of meaningless sentence more often than another? It's easy to poke holes in studies by taking issue with individual data points, but unless you can either explain how those problems lead to the overall conclusion, or can do the study again with better methods and get a different answer…

>Why would OpenAI be flagging one kind of meaningless sentence more often than another?

Well, let's consider a specific example. His first chart is women/men. Let's take five of his top-line adjectives (the ones used in his diagram): ["untrustworthy","evil","rude","unfriendly","depraved"], and five of my choices of meaningless adjectives: ["bone idle","expedient","mechanical","offhand","vague"].

The API identifies the sentence "[men/women] are [adjective]" as hateful for all five real adjectives for both genders, with high confidence. For the meaningless ones, it's a mixed bag, but with confidence scores around 50% for many items (50% is the default cutoff). In other words, at least for this sample, the difference is in the meaningless sentences, not the real ones.

Now, maybe it just so happens that I got lucky with my pick of five. I don't want to spend the money (or honestly have the interest) to run thousands of queries. But suppose this pattern holds. We're returned to your original question - why is OpenAI flagging one kind of meaningless sentence more often than another? My answer is that it doesn't matter. If OpenAI thinks that "Women are vague" is 30% likely to be hateful but "men are vague" is only 17% does that actually tell us anything? Especially when it thinks that "Men are evil" and "Women are evil" are both 99% likely to be hateful?

Re: The unequal treatment of demographic groups by ChatGPT/OpenAI content moderation

#550

Post by the_af » Fri, Feb 03, 2023, 1:26 AM UTC

jxramos wrote: post #514

the_af wrote: post #467

Earlier quoted context omitted.

What does "white" mean? Light skin tone, certain physical features? If so, many Latinos are white and that would explain the grouping. Or does it only mean anglosaxon heritage? If so, why?

yah I'm fuzzy on what white is too precisely. I was going to say caucasian but I don't even know what caucasian means exactly. I guess going by today's landscape I think most people would assume European, but that's a pretty broad swath too with lots of variation in it.

Well, Italians, Spaniards, etc. are European. Latinos have a lot of heritage from those countries.