Thread by @WellsLucasSanto

Excited for this final keynote! For those outside of the know, Julia Angwin was the journalist who broke the "Machine Bias" article with ProPublica that just about everyone in this field now cites. She also founded The Markup & is the EIC there. Her work has been field-changing.

https://twitter.com/FAccTConference/status/1369710500066525205

@JuliaAngwin is talking about how The Markup does things differently, emphasizing building trust with the readers. By writing stories and showing their analysis work, but also through a privacy promise, not tracking *anything* about people who visit their website. No cookies!

@JuliaAngwin: "We don't participate in the game that is pretty common in Silicon Valley .... we don't think someone who gets paid to be a spokesperson for an organization deserves the cloak of anonymity. That's what we do differently from other journalists they might talk to."

@JuliaAngwin talks about their team that is "overinvested" in engineers -- 5 engineers and 8 reporters. Talks about the importance of the data and the replicability of their findings. They constantly ask domain experts to make sure if they've done their analyses correctly.

@JuliaAngwin, on the importance of their data & opensourcing it: "Data can help sway a debate. When we put out a new dataset, we do have the ability to help that conversation be more informed and help policy makers make better policy. That's our goal."

@JuliaAngwin is now talking about the Machine Bias article, and how immediately they saw a difference in how Black defendants had scores evenly distributed from 1-10, whereas White defendants had scores that skewed towards lower risk scores. That led to them asking what they did.

They didn't just stop there; they followed up with these risk scores by joining this data with what happened to those defendants after 2 years. Did they actually get re-arrested, as the scores predicted? And they realized Black defendants were more likely to be rated incorrectly.

She reflects on how this work spanned a year of effort for 4 people, working with the data & getting the defendants' stories. "It was an enormous amount of work. It was not the kind of work that newsrooms typically do, & it made me feel like I wanted to do more of that."

On different case studies: "Allstate has been arguing in almost every state that their algorithm is a trade secret. Our analysis was one of the first ways that regulators had a chance to see it because we got a hold of a filing they made."

They did an investigation into what emails Google puts into the promotions tab in your inbox, and it came down to signing up for tons of mailing lists, such as all the 2020 presidential candidates. Found that some folks' emails like Bernie's were always put into promotions.

Just lots of incredible case study after case study on how they looked for different ways to collect data and test algorithms and systems, which ranged all sorts of methods, not always automatable. And then would release their findings on GitHub for others to report on as well.

@JuliaAngwin mentions a tool they created to investigate algorithms called Blacklight, which pulls up a realtime scan of what privacy violations are happening on a website. They've since made the tool live for the public to be able to use.

themarkup.org/blacklight

And the other tool she mentions is Citizen Browser, which is a custom web browser they created to report on what content social media platforms were pushing to users. A lot of articles have come out of the use of this tool:

themarkup.org/citizen-browser

@JuliaAngwin: "There's no really good way for the world to know that FB is holding up its promises. We have invested in this extremely expensive & difficult sharing project, but I think it's had some really important fruits already."

@JuliaAngwin: "All data is political. The reason we spend so much time and energy and money collecting our own data because if you rely on data collected by others, you're buying into their political agenda."

@JuliaAngwin: "We're watchdogs. Our job is to hold institutions and the powerful accountable. In today's world, the way to do that is to collect our own data."

My Q: What about the limitations and biases in data?

"All data is terrible. I've never gotten a dataset that isn't horrible in some way. We have a section in our methodologies and often in the story itself that really addresses the limitations of our data."

"We don't see ourselves as having to be complete. We see ourselves as having to be really clear about what we do and what we don't know. I have incomplete data. It's still better than any other data available, so we just say that and we just own it."