So, Should Chatbots Be Female?
Our monthly analysis on machine learning trends.
With all of the recent technical advances in machine learning, it’s easy to overlook how much our broader culture sneaks into the products we build. Even if we imagine programs writing themselves in some nebulous world of cloud servers and graphical processing units (GPUs), the reality is that humans are still at the helm. Our team thinks about this relentlessly: We work with behavioral and transactional data from retail banks, retailers, and telcos, and always bear in mind that these blips of transactions in time are people — traces of people doing their best to live a meaningful life.
And reflections of our culture don’t just exist in the data. They also seep into the design choices we make when we imbue systems with human traits. So in this newsletter, we thought we’d ask why are so many chatbots female? And is that a net positive or negative for society?
This is actually a question that’s generated some scholarly interest, and rightly so. It’s one of those questions that seems simple until you realize all the things, both implicit and explicit, that are being encoded by our conceptions of gender and technology. And sure, it would be easy enough to pass this all off as the result of a long history of evil male robots — from Hal 9000 in 2001: A Space Odyssey to all those early Terminators — not interested in human preservation. But we’re going to do our best to avoid any simple explanations here (and anyway, there’s been a far share of pretty destructive female robots, too).
A Chatbot by Another Name
While there are certainly chatbots and virtual assistants with male characteristics, and a handful that are gender-neutral, for the most part they remain the exception. Apple has Siri. Amazon has Alexa. Microsoft has Cortana. Is this skew just the perpetuation of outdated gender roles, specifically a stereotype of the female secretary/assistant? Possibly.
It’s worth noting that in the U.K., Apple’s original virtual assistant was actually male, perhaps reflecting the archetype of the English butler/valet, particularly visible in British fiction. Then again, maybe this is all just the result of research that suggests that in general, people have more positive and pleasant associations with female voices. Or, more specifically, women appear to respond better to female voices while men don’t have a strong preference either way.
There are a lot of possible explanations for the skew toward female names and voices for our chatbots, and there’s probably a mix of causes rather than some neat narrative of straightforward cause and effect. There’s user testing and feedback that goes into these decisions after all. Then there’s the fact that the feminized virtual assistant has become a recognizable and expected entity, something to be copied and reproduced. Plus there’s the historical precedent that the first truly operational chatbot, designed to emulate the style of a Rogerian psychotherapist, was named ELIZA.
Of course, having female voices at the helm of these services doesn’t necessarily have to be purely negative. Sure there’s the constant threat of bot sexual harassment, and there’s good reason to feel uncomfortable with the dominance of female characteristics in these products while women themselves continue to be deeply underrepresented as engineers. But you could also just as easily point to some potential positives.
Virtual assistants are “assistants” in one sense, but they’re also sources of knowledge, modern oracles through which we learn not only mundane facts about sports and celebrities, but also essential information about the state of our world. And, yes, they’re imperfect, and sometimesoffend or misunderstand things in seemingly absurd ways, but there’s also a power being invested in these voices and personalities that’s by no means trivial. In fact, to reinforce an attitude of respect, companies are now rewarding better social etiquette by having their voice assistants provide equally polite responses. There’s also been some effort put into responding to verbal harassment (often of a sexual nature) more thoughtfully, though the results aren’t consistent across the board.
All of this brings up broader questions. For instance, how do we currently imagine AI both as an interface and as a psychological extension of ourselves? Should we adopt the viewpoints of Andy Clark and think about technology as extending our minds? How much of a role do we want the broader culture to play in these decisions?
We know that word embeddings, the first step in almost every natural language processing model these days, reflect deep-seated cultural gender biases, and that it takes careful and creative engineering to reinstate proper neutrality. We also know that in the context of the older technology of the book, the voices of women characters in literature became increasingly scarceover the course of most of the twentieth century. It’s important, in other words, to ensure that female voices don’t simply disappear from technology, that they remain verbal embodiments of the various different forms of knowledge being channeled and filtered through our devices. But it’s necessary too to keep in mind how our culture continues to perpetuate bias and inequality, not only in the way we act but also in the way we speak.
Both Apple and Google now allow you to select the gender of the their virtual assistant voices, so that these decisions feel more interactive, more under the control of users. But the precedent of misrepresentation set by Google Duplex’s initial unveiling still hangs over the increasing ambiguous line separating human and machine communication. Meanwhile, as algorithms get better at detecting what we say and how we feel when we’re saying it, it’s worth considering the new array of stereotypes that risk sneaking into products that are busy not only trying to communicate with us but also to make sense of who we are.
Better Engineering Doesn’t Necessarily Mean Better Outcomes
For companies, transparency and responsibility will always play big and necessary roles in these contexts. But all of this also raises complicated questions that don’t have simple or straightforward answers. While trying to serve customers better by building more satisfying and “human” interfaces, companies also run the risk of undermining trust. It’s hard, after all, to serve customers without also defining their expectations, just as it’s difficult to emulate human interactions in ways that don’t end up feeling unnatural or manipulative.
As recent ML products have demonstrated, there’s still a lot of work to be done on this front. But the more general research on fairness and transparency has also shown that there’s enormous room for thoughtfulness and creativity here.
Arguably, it’s at the intersection of engineering and more humanistic modes of inquiry that we’ll find solutions that feel flexible enough to attend to the multiplicity of human experience. Technology and culture are both imperfect mediums. It’s by rethinking these imperfections, rather than just reproducing them, that we’re likely to end up with better and more equitable products.
