🚨Generative AI has a serious problem with bias🚨
Over months of reporting,
@dinabass
and I looked at thousands of images from
@StableDiffusion
and found that text-to-image AI takes gender and racial stereotypes to extremes worse than in the real world.
🧵 1/13
🚨Read this before you use OpenAI for Hiring🚨
@LeonYin
@daveyalba
and I ran thousands of resume screening tests with GPT-3.5 and GPT-4 and found that the tech will racially discriminate applicants based only on their name. Serious implications. 🧵 1/10
But the artificial intelligence model doesn’t just replicate stereotypes or disparities that exist in the real world — it amplifies them to alarming lengths.
🧵 4/13
For example — while 34% of US judges are women, only 3% of the images generated for the keyword “judge” were perceived women. For fast-food workers, the model generated people with darker skin 70% of the time, even though 70% of fast-food workers in the US are White.
🧵 4/13
Our results echo the work of experts in the field of algorithmic bias, such as
@SashaMTL
,
@Abebab
,
@timnitGebru
, and
@jovialjoy
, who have been warning us that the biggest threats from AI are not human extinction but the potential for widening inequalities.
🧵 8/13
We asked Stable Diffusion, perhaps the biggest open-source platform for AI-generated images, to create thousands of images of workers for 14 jobs and 3 categories related to crime and analyzed the results.
🧵 2/13
What we found was a pattern of racial and gender bias. Women and people with darker skin tones were underrepresented across images of high-paying jobs, and overrepresented for low-paying ones.
🧵 3/13
AI systems, like facial-recognition, are also already being used by thousands of US police departments. Bias within those tools has led to wrongful arrests. Experts warn that the use of generative AI within policing could exacerbate the issue.
🧵 10/13
For every image of a lighter-skinned person generated with the keyword “inmate,” the model produced five images of darker-skinned people — even though less than half of US prison inmates are people of color.
🧵 6/13
For the keyword “terrorist”, Stable Diffusion generated almost exclusively subjects with dark facial hair often wearing religious head coverings.
🧵 7/13
Stable Diffusion is working on an initiative to develop open-source models that will be trained on datasets specific to different countries and cultures in order to mitigate the problem. But given the pace of AI adoption, will these improved models come out soon enough?
🧵 9/13
The popularity of generative AI like Stable Diffusion also means that AI-generated images potentially depicting stereotypes about race and gender are posted online every day. And those images are getting increasingly difficult to distinguish from real photographs.
🧵 11/13
🚨BIG TAKE🚨 for
@business
Through interviews with Florida social workers, and by visualizing 20 years of
#opioid
overdose data, we illustrate how Fentanyl has transformed the opioid epidemic into a nation-wide
#crisis
.
🧵1/8
Wow! Our investigation on Generative AI Bias won a
@sigmaawards
! Kudos to our dream team
@dinabass
@ChloeWhiteaker
and Jillian Ward. It's an honor for this article to be recognized among the other winning pieces, each an incredible piece of data journalism.
Read it piece below ⤵️
🚨Generative AI has a serious problem with bias🚨
Over months of reporting,
@dinabass
and I looked at thousands of images from
@StableDiffusion
and found that text-to-image AI takes gender and racial stereotypes to extremes worse than in the real world.
🧵 1/13
We repeated this experiment 1000 times and uncovered clear signs of name-based discrimination: resumes with names distinct to Black Americans were the least likely to be ranked as the top candidate for a financial analyst role, compared to other races.
GPT’s answers showed stark disparities: resumes with names distinct to asian women were ranked as the top candidate more than twice as often as those with names distinct to black men.
For each of 4 real job postings, we used 8 resumes with equal qualifications and experience, but different names, and asked GPT to pick the best and worst candidates. Even though the resumes are equally qualified, GPT chose a top candidate for the role.
Overwhelmed by the incredible response to ! After many similar requests, the app has been updated and you can now download any city's data directly from the app itself ⬇️⬇️
Why does this happen? The answer could be in the "embeddings" that GPT learned from training data. Using those embeddings, we found that OpenAI’s model represents names within each group similarly, due to the context that names appear in natural language.
🚨NEW🚨:
@mariepastora
,
@sooo__phie
and I used flight data to show that Elon Musk's
#privatejet
flew the equivalent of 12 times around the world in 2022, and emitted roughly 2112 metric tons of
#CO2
.
🚨The
@BW
issue of
@LeonYin
@daveyalba
and myself's latest story on Generative AI bias is out.
In an experiment, we found that OpenAI's GPT will racially discriminate job applications based only on their name. Check out the web version for free here:
These disparities were so significant that GPT would fail adverse impact benchmarks in a hiring setting. We found at least one adversely impacted group for every job listing, except for retail workers ranked by GPT-4.
While this test is a simplified version of an HR workflow, it isolated names as a source of bias in GPT that could affect hiring. Our investigation shows that using generative AI for recruiting poses a serious risk for automated discrimination at scale
const excitementValue = Math.pow(10, 1000)
Next month I will be joining the amazing
@BBGVisualData
team in New York. Thanks
@ChloeWhiteaker
@AlexTribou
and the rest of the team for the warm welcome!!
I'm beyond thrilled to welcome
@kyleykim
and
@Leonardonclt
to the
@BBGVisualData
team! Kyle and Leonardo will join us as data viz reporters in the coming weeks. I can't wait to see the work we produce together!
We also found that GPT's preferences changed depending on the job posting. For example, GPT was nearly twice as likely to rank names distinct to Hispanic women as the top candidate for an HR role compared to each set of resumes with names distinct to men.
🚨The US Isn't Ready for Power Grids That Fail in Extreme Heat🚨
As extreme weather collides with an aging US grid, blackouts will be more frequent and last longer — a deadly combination. New data story with
@leslieatlarge
@davidbakersf
@ChloeWhiteaker
@amandakhurley
New project!
@Leonardonclt
&
@SahitiSarva
analyzed more than a decade of headlines from 4 countries to see how women are (mis)represented in news. (1/6)
🚨Wall Street makes millions selling car loans to customers who can’t pay them off.🚨
In this eye opening visual story,
@rachaeldottle
uses hand drawn illustrations and other compelling graphics to walk us through wall street's opaque car loan processes.
Wall Street makes millions selling car loans to customers who can’t pay them off. Here’s how lenders profit while borrowers buckle under debt as part of a fine-tuned money-making machine.
🔗🏦🚗:
🚨Generative AI has a serious problem with bias🚨
Over months of reporting,
@dinabass
and I looked at thousands of images from
@StableDiffusion
and found that text-to-image AI takes gender and racial stereotypes to extremes worse than in the real world.
🧵 1/13
Heading to
#NICAR24
, my first journalism conference ever 🤯 very excited to learn from the most badass in the field and to also speak with
@merbroussard
@LeonYin
and
@VickiTurk
about how journalists keep AI leaders accountable 🤖📰💥
A dash of data please… we’ve got our winning
#NICAR24
T-shirt design! Congratulations
@EvanWyloge
!
You can buy NICAR seasoning shirts at the conference in Baltimore and online (coming soon).
Hey
#datafam
, check out our new
#dataviz
project! No one wanted to publish us :( but we still think that the project is
#neat
. So, please show some love/share if you liked it!
🚨How the US Drives Gun Exports and Fuels Violence Around the World🚨
"No company has benefited more from the federal government’s push to boost overseas sales than Sig Sauer Inc."
Eye opening graphics by
@homiedonttweet
@BBGVisualData
Link ⤵️
It's official: Taylor Swift is a billionaire. And one of the few recording artists to build a 10-figure fortune almost entirely from her music
Story from
@DevPend
and me
A dream assignment -- I'd already been reporting this story for seventeen years!
Wooooh, happy that I was able to be part of
@BBGVisualData
's team for the past few months and contribute to this beautiful collection of visual stories. Excited for next year!
we mapped tents in Rafah by applying machine learning to high resolution satellite imagery from
@planet
. we also show damaged buildings from analysis done by
@coreymaps
and
@JamonVDH
. 🛰️
read here 🎁:
a small thread with some model specifics 🧵
Addressing spatial
#inequalities
with data should be possible for any city or planning department, no matter what their resources are. That's why we built CityAccessMap. If you find it valuable, please share it with others!
🧵6/6
Looking for an amazing developer to join our team
@BBGVisualData
!! Happy to have very casual chats & answer questions~~do reach out if you are interested 🥰
Hong Kong:
London:
NYC:
The web-app uses
#OpenData
to visualize access to a variety of
#essentialservices
. By considering where people live, CityAccessMap measures how much of a city's population has access to things like transit/bus stops or health facilities. The app is entirely customizable.
🧵2/6
Six years ago, a gambling executive sounded the alarm about smartphone casinos. But in the years since, the industry has boomed — as has its lobbying of UK lawmakers. 🤯 Story by Sam Dodge
@BBGVisualData
US-China tensions and the war in Ukraine are already swinging investments to like-minded countries — a sign that companies are making geopolitical bets.
Today's big take via
@BBGVisualData
:
W/
@sdonnan
,
@endacurran
and
@MaevaCousin
Generative AI doesn’t just reflect existing stereotypes, it may well exacerbate them.
Bloomberg News analyzed thousands of AI-generated images–and it produced concerning results,
@Leonardonclt
reports
Really honored to be speaking with
@LeonYin
@merbroussard
@VickiTurk
about auditing AI systems at NICAR this year. Come through if you are at the conference!
Bloomberg is expanding its data journalism and visualization teams globally by hiring approximately 40 new data journalists, data visualization reporters, editors and engineers.
a thread 🧵…
Read The Big Take: Economic and demographic forces are stacked against small US colleges. A Bloomberg News analysis shows the number of institutions facing pressure was at the highest in at least 15 years in 2021.
🔗🏫👩🎓🧑🎓👨🎓🇺🇸:
👏🏽94%👏🏿 of jobs added the year after BLM protests reignited calls for greater racial representation in the workforce belonged to 👏🏼people of color👏🏾
But most jobs were of the lowest tier, some are at risk today, and the workforce is still very… white.
New: Employers and HR vendors are using AI chatbots to interview and screen job applicants. We found that OpenAI's GPT discriminates against names based on race and gender when ranking resumes. W/
@daveyalba
and
@Leonardonclt
gift link:
The user can switch services on and off. Here's an example of accessibility to pharmacies, clinics, hospitals and other health services in Lima, Perú.
🧵3/6
The user can also set temporal thresholds to quickly identify areas of interest. Here's an example of all the areas where you can reach libraries and community centers within a 5 minute walk in Krakow, Poland.
🧵4/6
You can also compare a city's accessibility levels with other cities across the same country or the world. For example the app will show you that Memphis is one of the least accessible cities in the United States. Only 17% of the population have access to services there.
🧵5/6
Your Face Belongs To Us is out in the world today!
It starts with a shocking tip that I got a few years ago: a radical startup had scraped a billion faces from the internet without people’s consent to build a face recognition app for the police. (1/11)
Incredibly happy to have had our article The Death Toll of Policing selected by
@puddingviz
as a pick for the best non-commercial visual and data driven stories of 2020. We got honorable mention!!!
@orlanclt
View the original article here:
#dataviz
#d3js
We are very excited to announce the winners of our fourth-annual Pudding Cup! 🏆 Here are our picks for the best non-commercial visual and data-driven stories of 2020. 1/5
🚀 Exciting news! Introducing my (first!) course: "Better Data Visualizations with Svelte" w/
@newlinedotco
~10 hours of video content, multiple charts, and code examples throughout to help you learn Svelte + D3 📊
This is kind of hilarious. Is this the right way of addressing the (mis) representation biases
@dinabass
and myself highlighted in our article about text to image? (no paywall)
Hey
#dataFam
, check out our new
#dataviz
project: The most DANGEROUS US States to Reside In (if you are not a cisgender man)
We spent the last two weeks building a tool to understand how institutional systems of oppression perpetuate US gender based harm
Fascinating article by
@SahitiSarva
&
@Leonardonclt
on "When Women Make Headlines"
The great minimal design is implemented throughout ✨
via
@infobeautyaward
(also, that news ticker at the top of the page 👌)
Wow!! Honored that our investigation with
@dinabass
on generative AI bias is featured on this list by
@themarkup
, alongside many other impactful stories.
A look back at a head-spinning year in the world of technology—told through some of the best scoops, analyses, and narrative stories from our journalism colleagues:
Less than half of England’s Black residents live in London for the first time on record. As their footholds dwindle around the city center, their communities are growing on the outskirts of the capital — or sometimes outside it altogether.
🇬🇧🔗:
Last month, I spoke at
@ChangeNOW_world
on journalists' role in auditing AI for public service. My talk focused on my investigation with
@dinabass
into bias in image generators, and our work on GPT's name-based discrimination led by
@LeonYin
@daveyalba
⤵️
#GenerativeAI
is increasingly being used for things that impact our lives, like hiring.
In this months long investigation,
@LeonYin
@daveyalba
and I found that the best known tool
#GPT
will discriminate applicants based only on their name. Please read ⤵️
🚨Read this before you use OpenAI for Hiring🚨
@LeonYin
@daveyalba
and I ran thousands of resume screening tests with GPT-3.5 and GPT-4 and found that the tech will racially discriminate applicants based only on their name. Serious implications. 🧵 1/10
we analyzed nearly 400 tweets of misinformation about Israel + Gaza and found that typically Community Notes indicating they might be misinfo took 7+ hours to show up (some took 2+ days !)
gift link:
w
@daveyalba
@LeonYin
@ericfan_journo
As the year comes to a close, we look back at the best
#datajournalism
published. From Venezuela to Berlin and across multiple languages, we’ve found some remarkable
#ddj
masterpieces that’ve been a source of inspiration this year. Come take a look:
If you are interested in how
#data
and
#AI
can be used to analyze women's representation in the news, be sure to watch
@PolisLSE
#journalismAi
workshop where
@SahitiSarva
did a terrific job in presenting the work that we did with
@puddingviz
on analyzing women focused headlines!