Dr. Cory James Young
aidoThere's no systemic racism but weirdly every time we train an AI on public data sets it becomes very racist.
Dirk-Jan Swagerman@aidenbenton that’s our training set too growing up…
RichardOntario@aidenbenton I’m in no way intelligently educated about any of this But I’m curious Now this is my scientific/not scientific study and I am quite flummoxed by the results This “study” is NOT suitable for work Enter the words “beautiful” or “pretty” or a suitable adjective along with “naked” Every image is a Caucasian Not Black, nor from any Pacific Rim Region but White Why is that?
Maisie Summers@RADC @aidenbenton
Because the systems are developed in countries with a predominantly white populace. They use data sets taken from their own country (often, the U.S.).
They should add data sets from other places: China, India, Middle East, and as many African countries as they can get.
But they don't. The data is usually U.S. and inherently skewed in favour of whites.
Because the systems are developed in countries with a predominantly white populace. They use data sets taken from their own country (often, the U.S.).
They should add data sets from other places: China, India, Middle East, and as many African countries as they can get.
But they don't. The data is usually U.S. and inherently skewed in favour of whites.
@maisiesummers @aidenbenton So it’s the data that’s collected from user experience? I need to read a book lol That can’t be right because historical data is one thing but the initial offer is human decided So it’s back to the people first with their particular set of prejudices?
@RADC @aidenbenton
Not even close to what I said. I said the data is typically collected from white-dominated societies. That skews the data.
Not even close to what I said. I said the data is typically collected from white-dominated societies. That skews the data.
@maisiesummers @aidenbenton Yes ma’am This data is gathered from user engagement?
Mark Harris
Katja@RADC @aidenbenton The AI is biased by learning from public data sets, and most of the public still treats a straight white man as the "default human". So usually, a character (like in a book) is always presumed straight and white unless otherwise specified. If you don't specify "black", the AI will go to the default, which it recognizes as white.
y4si0@RADC @aidenbenton I think it may something to do with amount of ex Eastern Block models compared to all the other models based in the US. Add to that the fact, that in most of English speaking countries, people are usually white and it may be the answer.
@y4si0 @aidenbenton Can’t disagree but under what circumstances did the very first offering of pictures have ? Who picked the very first and why? Their personal prejudices I suppose
Ronja A-R (e/em/eir ; hen)Good point! If you are curious about the historical roots of that AI bias, here is one starting point.
A veteran and history teacher has put together an over two hour video about the history of Native Americans that is entirely avoided in a typical US high school curriculum.
I thought I knew the basics of what happened to the indigenous peoples of North America after 1491 CE, but found I didn't know more than fragments of it. Recommended!
They Were Just in the Way | Indian Removal
Palo Markovič 🇺🇦@RADC @aidenbenton one important fact that is often overlooked and contributes to bias is the language. English is not the the most spoken language in the world - https://en.m.wikipedia.org/wiki/List_of_languages_by_number_of_native_speakers - but it is the language of IT and the one used to train AI (outside China of course).
@palomarkovic @aidenbenton Yet Google AI was shut down because it invented its own language to speak with the other AI’s I think I remember reading
Stormcrow@aidenbenton Systemic is right. It's not just racism, but all the other -isms embedded in every aspect of our lives. Find the intersection of those -isms and start remediation there.
ǝʇɹoʞ oıɹɐɯ 
@aidenbenton <rhetorical>I wonder what the reason for that is. 🤷♂️</rhetorical>
Gary Gough@aidenbenton
Recalling a story about an AI being trained to recognize sunny days by accident. Was supposed to recognize tanks but all the training pictures of tanks had been taken in relatively bright conditions.
Pretty good indication of racism in the public data.
One of my jobs involves having drunks yelling death threats at me, and really you get so used to it that you just mostly ignore it, an AI, being trained, can't make that valuation by definition.
Recalling a story about an AI being trained to recognize sunny days by accident. Was supposed to recognize tanks but all the training pictures of tanks had been taken in relatively bright conditions.
Pretty good indication of racism in the public data.
One of my jobs involves having drunks yelling death threats at me, and really you get so used to it that you just mostly ignore it, an AI, being trained, can't make that valuation by definition.
Kacem Le Moël a pris vapeur@aidenbenton AIs should be taught to pray, that's why.
J Lam 👩🏻💻👩🏻🎨
Roel 🇺🇦🔻❤💚🏳️🌈🍆🍑🦄@aidenbenton Based data!
Kay 
@aidenbenton "public data sets" -from where? How is diversity "captured"? What levels of indigenous data sovereignty are recognised in survey design? Are AIs trained with community held data sets built with collaboration of people who may be subjects of actions based on that ML training? ..
Being hypothetical here cos late to the thread. Sampling bias is deliberately built in to some political polls. Hard to see other polls & AIs escaping deliberate or unconscious sampling biases
Cysio 
@aidenbenton In the university some bros used that as an argument that being systemically racist is good and objectively correct, actually
Jenny AdamsAnd when you train Russian dogs to explode none-Russian tanks, but all you have to practice on is Russian tanks... Which I think illustrates your point perfectly.
Riley S. Faelan@jennyadams @aidenbenton: Clearly, it's a vengeance of the dog deity at play, for trying to blow up innocent dogs.
Дурдомова 
@aidenbenton it's almost like it happens...systematically 
MIlton@aidenbenton I don't think it's the system. Train on racist data, win racist prizes.
GillianA 🌺@aidenbenton also women and children depicted as abused and uncared for. #Loab
https://www.cnet.com/science/what-is-loab-the-haunting-ai-art-woman-explained/
https://www.cnet.com/science/what-is-loab-the-haunting-ai-art-woman-explained/
VallinSFAS@aidenbenton This is in a lot of new books on the subject (I work in a bookstore)
THIS ACCOUNT HAS MOVED@aidenbenton the machine becomes the data with #MachineLearning
Capibara (Luca)@aidenbenton Microsoft tay was an amazing example of this
CL Stimmel, PhD 
mic drop!
Charlotte Horseface@aidenbenton if even AI cannot avoid racist thinking, then how does that bode for the rest of us struggling out here to counter racist thinking in our own minds?!? Maybe concluding that one group of humans is superior to another group is a logical conclusion arrived at by basic data sets? 😮 Perplexing.
Morten Toudahl@aidenbenton here is an interesting read:
https://www.nist.gov/news-events/news/2022/03/theres-more-ai-bias-biased-data-nist-report-highlights
https://www.nist.gov/news-events/news/2022/03/theres-more-ai-bias-biased-data-nist-report-highlights
Niels@aidenbenton This means AI is currently poorly skilled in masquerading its racist bias. Needs more training I guess.
Jack Mott@aidenbenton that could be explained in like 3 broad ways. AI doesn't work, police data is inaccurate and biased. black people in america currently commit crimes more often. of course in real life probably at least 2 of these things are happening to some degree.
Andrew Jorgensen@aidenbenton that's not what systemic means to me. Systemic racism is racism that's reinforced by a system (many systems) so that it persists and even amplifies because of how the system works. You'd still have racial bias in public datasets if we'd solved all the systemic issues but residual racism still existed because a generation hadn't died off yet, or after that but including older data in the datasets. And AIs would still tend to be racist as a result.
Between Wanderings@aidenbenton The first time I asked an AI app to synthesize an illustration of two 19th-century Jews smiling at each other, I got a picture of two smiley Hasidic guys (which is fine) with noses just like those in the era's anti-Semitic caricatures (which isn't fine).
More recent attempts were better but still a teensy bit schnozzy.
Kenneth@aidenbenton Now do GOP platforms.
gabi@aidenbenton
and even applies to AI algorithms trained to diagnose chest xrays:
https://www.nature.com/articles/s41591-021-01595-0
tldr: the AI has a higher rate of underdiagnoses/missed diagnoses in POC/marginalised populations
Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations - Nature Medicine
Artificial intelligence algorithms trained using chest X-rays consistently underdiagnose pulmonary abnormalities or diseases in historically under-served patient populations, raising ethical concerns about the clinical use of such algorithms.
Toni Aittoniemi@aidenbenton Yes. AI only models what it sees. And instead of ”trying to repair” it, they would best be used as mirrors to what people are really like.
Also, public application of #ai needs 100% #transparency on the #training data. No business decision is too important to allow cultural bias to contaminate the system.
This is where #ethicists need to work together with #computer #scientists. The problem is no longer algoritmical.
Erika Sumner@aidenbenton I studied #bias in #AI as part of my #DEI certificate from NYU & I think I can help explain the dynamic for #tech, #diversity, #equity, #inclusion & #belonging newbies. Putting aside overt #racism, everyone on the planet possesses #UnconsciousBias. Absolutely everyone. Because the coders who create AI (who teach the machine how to “think”) are human, their unconscious #biases multiply every time the AI “thinks”.
thomkennon@aidenbenton I'm an increasingly boisterous fan of the #fuckAI hashtag as ethical battle cry.
Bill, organizer of stuff@aidenbenton
It's a subspace message being picked up by the neural net
It's a subspace message being picked up by the neural net
@aidenbenton Want to learn about #bias in #AI? Here's some recommended reading. Enjoy!
-#Ethics Codes Are Not Enough to Curb the Danger of Bias in AI (article, BRINKNews) https://bit.ly/3uzmnBf
-DISCRIMINATING SYSTEMS
#Gender, #Race & #Power in AI (NYU, #Google Open & #Microsoft #Research) This one's my fav but be warned, it's a behemoth 
at 33 pages https://bit.ly/3UJvDxk
-The role of AI in mitigating bias to enhance #diversity & #inclusion (#IBM, 16pgs) https://bit.ly/3W0Axrf
@aidenbenton: That's because everybody has been conditioned to ignore racist trolls, which leads public datasets containing very few cases of somebody patiently explaining to racist trolls that they shouldn't be racist or trolls.
Matt ⎊ 
@aidenbenton funny, that 🥴
Floyd@aidenbenton and sexism
Robert Wilson, III 🏴🇪🇸🇸🇮@aidenbenton Samples are probably also biased. For some reason AI is basically never trained on proper random or stratified samples.
Mel Bartley@aidenbenton The best argument in favour of machine learning I've ever heard. Strips our assumptions bare.
@aidenbenton The core idea of racism is generalization (potentially +hate). If every model's objective is to generalize... well...
@aidenbenton: That's because in everyday life, the main thing holding racism at bay, barely, is taboos, and taboos are what a simple robot doesn't see when looking at a dataset. A future robot will be able to do better (assuming it has a good robomom to give it a bunch of serious talks).
msgbi