,
17 tweets,
3 min read
Read on Twitter
IT'S THE NOT-QUITE DEFINITIVE ARE-ALGORITHMS-RACIST? THREAD!
When people talk about whether or not algorithms are racist, they really mean "are machine learning algorithms racist?"
The short answer is no, algorithms aren't racist. They operate without bias to replicate societal behavior. It's human actions algorithms copy that are racist.
The short answer is no, algorithms aren't racist. They operate without bias to replicate societal behavior. It's human actions algorithms copy that are racist.
I don't mean that people are racist in the way someone like Bill O'Reily would define racist - most people aren't thinking about ways to murder other racial groups.
There's a famous study by LaTonya Sweeny that a much earlier version of Google's algorithm would turn up ads related to criminal background searches substantially more frequently when people search for ethnic sounding names.
Was the algorithm racist? No. People searching for ethnic-sounding names are more likely to be searching for criminal background checks.
Were the countless people searching for criminal background checks on ethnic names racists? That's a possibility.
Were the countless people searching for criminal background checks on ethnic names racists? That's a possibility.
But, more than racism, it could just mean that people with ethnic names are more likely to grow up poor, meaning they live in neighborhoods higher in crime, meaning that a landlord who checks all applicants disproportionately search for ethnic names.
The effects, however, are kind of racist.
If you had searched for LaTonya Sweeny, a then Harvard professor with no criminal background, and one of the first results says "DO YOU WANT TO SEE LATONYA SWEENY'S CRIMINAL BACKGROUND?" there's a chance it will lead you to think she had a criminal background.
Similarly, if you work at a university where the students are overwhelmingly white, and train an algorithm in facial recognition using students as the test group, the resulting program will be much better at sorting white faces than black ones.
Was anyone deliberately trying to design devices that didn't work on black people? No. Was the math racist? No. Are there racist effects to living in a world where devices only work for white people? Kind of.
This is all extremely different than what conservatives accuse Facebook and Google of doing when they accuse the sites of manipulating algorithms to disadvantage conservatives.
That accusation is dumb.
That accusation is dumb.
Liberals do this too, although the lawmakers or major media outlets that seem pander to their bases by promoting the idea that Silicon Valley tampers with their algorithms to hurt them seem to be conservative.
Is Google treating searches for Steve King different than other searches to make him seem racist? Probably not. It's pretty clear most national attention for Steve King has been focused on his more controversial statements. His search results reflect that.
Can we do anything about machine learning going off the rails? Well, yes. Starting with big, diverse data sets helps. And tweaking the algorithm not to treat backdoor indicators of race (like names) as indicators of race can work too.
The latter isn't something that can always be handled proactively - unanticipated problems will occur in the wild, and people might have to be patient.
Can we do anything about Steve King seeming racist in Google's search results? Europe has a model where people can request Google remove search results they disagree with. To be blunt, that will harm conservative media more than it helps them.
Thank you for reading my Ted talk, "Was Ryan Saavedra looking for an explanation or just dunking on AOC? Because I can do that first thing."
