For years, many people assumed that the decisions algorithms made were objective and neutral. But this is not the case, explained Julia Angwin, a former investigative journalist at ProPublica who is currently working to start up a newsroom analysing the impacts of technology on society, speaking in a session at the International Journalism Festival in Italy on 14 April.
Angwin has been investigating cases of algorithm bias and discrimination, finding, for example, that one of the programs used in the criminal justice system in the United States to predict whether defendants were likely to reoffend discriminated against black people when predicting recidivism rates.
Another algorithm used to calculate car insurance premiums resulted in larger payments required from people who lived in minority neighbourhoods than from people who lived in predominantly white areas with similar accident costs.
Following the revelations, the state of California’s insurance department announced it was planning to start an investigation into racial discrimination when calculating insurance premiums.
More recently, Facebook’s options for ad targeting have come under fire, after ProPublica found people could target ads based on anti-semitic categories; place adverts for housing while choosing to exclude people from seeing the advert based on their race; or exclude older users from seeing job postings.
In response to the investigation on posting rental housing advertisements on the platform and exclude people from seeing it based on their ethnicity, Facebook said it worked to crack down on the practice.
But the National Fair Housing Alliance in U.S. District Court in the Southern District of New York filed a lawsuit in March alleging that Facebook still allows advertisers to discriminate against groups such as Spanish-language speakers, mothers, and the disabled.
Angwin explained the impact of these investigations into algorithm bias is often not one of immediate change, and that the conversations in the computer science industry around the issue can easily become rather philosophical.
“In the criminal risk score case, there hasn’t been much change,” she told delegates, speaking via Skype. “Most of the places that used it are still using it, but there’s been a lot of discussion in the computer science community on how to build a better risk score.
“The score is 60 per cent accurate for both black and white defendants. It’s equally accurate, but the 40 per cent that is wrong it’s wrong in completely different ways.
“I think we have an opportunity now… that we could actually examine what we considered fair.”
But what’s the best way to check these algorithms? Would their creators ever open them up for public examination to prove that the algorithms that take important decisions about people’s lives, such as credit scores, health care, or housing, are doing so without any prejudice?
And would the social networks become more transparent about the inner workings of their platforms when it comes to targeted advertising or the moderation of hate speech?
“I just don’t think they have any incentive unless they’re forced to. I don’t see any reason why they’d open themselves up for public scrutiny unless the government requires them to,” explained Angwin.
“The algorithms themselves are often very hard to understand,” she continued, adding that companies also make a compelling argument for not revealing their trade secrets.
“I believe that the way to look at these algorithms is to look at what decisions they are making and audit these decisions.”
When investigating the algorithm that predicted the risk of recidivism, the first step for Angwin and her team was to put in a Freedom of Information Act request in Florida for the results for all the people who have been scored by the algorithm over a two-year period.
They had to hire a lawyer to fight for this, but obtained the data eventually as the scores were considered public records.
Once they obtained the data, they visualised it as bar charts to compare the distribution of scores for white defendants with that of black defendants, to see if there was any bias.
“The black defendant scores were evenly distributed. The white defendant scores were heavily clustered around the low risk. We were surprised and we wondered whether there might be some bias. But just looking at these scores alone, it doesn’t tell you if there was bias.”
The team then looked at all of their criminal records, over a period of six months, to see whether the prediction was correct.
“We were able to do a logistic regression to show that when you had a white and a black defendant who had the same record, black defendants were 45 per cent more likely to be assigned higher risk scores.”
But what are the core competencies required to hold algorithms accountable?
Journalists should make themselves “a little more tech literate”, she said, adding that some key skills include knowledge of SQL, or how to scrape the web – these are all skills journalists can teach themselves.
“People always think that you have to be a data scientist or a computer programmer. I don’t think that’s always true,“ she said.
Free daily newsletter
- German non-profit CORRECTIV to replicate the successful investigative journalism model used by the Bureau Local
- Need to catch up? Here's your weekly journalism news update
- ‘We have one thing that other publishers don’t have - time’: International investigative journalism at BuzzFeed
- Why algorithmic accountability reporting needs to go beyond transparency
- From social media to in-story experiences, chat bots help the BBC 'do things faster and at scale'