Finish this sentence: Man is to computer programmer as Woman is to __(fill in the blank)____.
Depending on your point of view, the answer to fill in the blank could be damn near anything, including computer programmer. But as I sit here writing on International Women’s Day, I know the actual truth is this: the tech and healthcare worlds are building a world of artificial intelligence (AI) and machine learning tools to run our lives that are taking with them the conscious and unconscious biases of our human forms. And that is a real drag.
How do I know this? Well, for one thing, the way that the above question was answered in one Stanford study of algorithms was this (based on a program of machine learning from Google News): Man is to computer programmer as woman is to homemaker. I kid you not. Have I time travelled to the 1950s? What the hell is a homemaker in this day and age? Sure, there are women who work primarily in the home taking care of that realm, but is “Homemaker” even a term anymore? And I happen to know plenty of men who have decided to play the parent/home CEO role while their partner women work. Seriously brogrammer dudes? You need to dispense with the June Cleaver fantasies and meet some actual women.
As women seek to gain more and more equality, are we teaching our machines to reinforce old stereotypes? Are we creating machines that work against women? It’s a serious challenge to avoid doing this, since machine learning relies on ingesting the already existing information that is out there to form connections and make predictions or recommendations. And if the existing information is biased, well. Garbage in, garbage out. As the Stanford study noted above reported:
“The fact that the machine learning system started as the equivalent of a newborn baby is not just the strength that allows it to learn interesting patterns, but also the weakness that falls prey to these blatant gender stereotypes. The algorithm makes its decisions based on which words appear near each other frequently. If the source documents reflect gender bias – if they more often have the word “doctor” near the word “he” than near “she,” and the word “nurse” more commonly near “she” than “he” – then the algorithm learns those biases too.”
Many in the field are seeking ways to remove gender and race stereotypes from machine learning algorithms, but one has to be cognizant that there’s a problem before one is likely to bother trying to fix it. It’s a bit of “first you have to admit you’re an alcoholic,” or in this case, that bro culture is not the be-all end-all existence that should drive the future economy. Unconscious bias is an insidious thing.
Some in the field shoot back: “We should think of these not as a bug but a feature. It really depends on the application. What constitutes a terrible bias or prejudice in one application might actually end up being exactly the meaning you want to get out of the data in another application.” Those are the words of Princeton computer science professor Arvind Narayanan, who appropriately points out that there are circumstances where gender connections would matter in Artificial Intelligence. For instance, (my example), one would not want to present men with recommendations for feminine hygiene products when using AI to deliver advertising (although I somehow get served up ads for, ahem, male things that I would rather not see).
But on the other hand, a great deal of AI/machine learning activity is focused on situations where gender bias could have a terrible negative impact. Imagine, for instance, if one is codifying healthcare AI solutions for the delivery of cardiovascular diagnostic information. Clinical trials over the decades have tended to include more men than women as subjects. Naturally, that means there’s a larger body of evidence out there about men with cardiovascular disease, and that this biased data will guide treatment protocols for everyone.
It is also well documented science that men and women present with extremely different symptoms when having a heart attack. What if the diagnostic AI program in an emergency room has only learned what’s relevant to men because it was primarily fed information that reinforces that limited data set? Frankly, it would make it just as bad or worse than women already have it in heart attack diagnosis. This defeats the entire purpose of medical AI: making the physician smarter because more information is immediately available and synthesized than they could ever get into their own brains. But if programmed incorrectly, or fed data that starts from a gender-biased view, guess what? It ain’t going to work so well.
Imagine you are the parent of a child applying to a university that adopts AI techniques for reviewing applicants. Will boys get accepted to engineering schools and girls get accepted to social science programs (presumably homemaking for God’s sake) because the algorithms steered them to that place? Same issue for job opportunities, where the plum STEM jobs may not even show up in a woman’s LinkedIn feed if the AI that drives a service like that doesn’t recognize women as legitimate candidates.
And consider this sports fans (which I suspect would be defined as men by most AI programs even though I know plenty of women, myself included, who spend Sunday in front of the TV watching sports): Stanford Professor Londa Schiebinger did an analysis which showed the male pronoun to be the default most of the time. In an interview she did with Metode she noted the following (it’s a long paragraph, but stay with me):
“If I put an article about myself into Google Translate, it defaults to «he said» instead of «she said». So, in an article of one of my visits to Spain, it defaults to «he thinks, he says…» and, occasionally, «it wrote». We wondered why this happened and we found out, because Google Translate works on an algorithm, the problem is that «he said» appears on the web four times more than «she said», so the machine gets it right if it chooses «he said». Because the algorithm is just set up for that. But, anyway, we found that there was a huge change in English language from 1968 to the current time, and the proportion of «he said» and «she said» changed from 4-to-1 to 2-to-1. But, still, the translation does not take this into account. So we went to Google and we said «Hey, what is going on?» and they said «Oh, wow, we didn’t know, we had no idea!». So what we recognized is that there is an unconscious gender bias in the Google algorithm. They did not intend to do this at all, so now there are a lot of people who are trying to fix it.”
My question is this: who is fixing it? Are they bringing in a diverse group of engineers (or homemakers) to analyze and reprogram the models, or is it same dudes, different day? I hope for the best, but worry more when I know things like this: Given that the vast majority of people working in tech are men, is it any wonder that virtually all of the AI driven assistants (Siri, Amy, Alexa, Cortana) have female names and are voiced by women? Thus, we perpetuate that stereotype. Women are meant to be assistants to men, who do the real work. After some emotional intelligence training, Apple added a male voice option to Siri, but that still isn’t the default position in the US, though it is in some other countries. IBM is really the only significant AI assistant company with a male voice standard issue product in Watson.
Many claim that research demonstrates it’s not sexism, it’s preference. A recent article in The Wall Street Journal cited two studies demonstrating that men and women prefer a female voice assistant because they believe it’s more welcoming and understanding. However, they prefer to listen to a male voice on certain subjects. Are you ready to gasp? As cited in this article by Vocativ, the first study, which was conducted by researchers at Indiana University in 2008, found that when having human-computer interactions, a female voice is “warmer.” The second study, done by Stanford University in 1997, shows that a male voice is preferred when learning about computers, but a female voice is better when hearing about love and relationships. I will pause and wait for your primal scream to end. Just because there is historical preference for what we know, doesn’t mean we should perpetuate stereotypes.
One light at the end of the tunnel: Mark Zuckerberg apparently has a personal AI assistant named Jarvis that is voiced by Morgan Freeman. Perhaps the actor with the famously rich baritone voice is the one who can really see the market opportunity.
Given that it is entirely possible to customize voices inside of technology applications, why can’t we select our own AI assistant’s name (Dwayne Johnson) and voice (Dwayne Johnson’s) rather than get stuck with whatever dame some guy has picked for us? This seems to be a market opportunity – a company that can translate voice assistants into celebrity voices or just apply a different gender or accent as desired by the end user? Now that would actually be user-centered design come to life. If we can dream it, we can do it!
Note: this post was first published in The Timmerman Report on March 12, 2017
Rick Lee says
Really great writing. Your acute sensitivity to gender bias is what makes you such a special observer of our industry. I flash back to the ’60s when inclusive pronouning became in vogue. Learning how to write “he/she said” or rotating the pronouns during an essay were my first baby steps in encultured gender bias.
Nastran Andersen says
Great post and interesting thoughts on the AI programming – now is the time to address it as the options become more pervasive. On a lighter note, have you ever heard of the Morning Man alarm clock app? Maybe we can incorporate some of those international options for our personal AI assistants, along with Dwayne Johnson, of course. 🙂
Lisa Suennen says
Wow Nastran, I like your style. My life is now complete having seen https://www.morningmanapp.com/listen/ Regards, Lisa