Avoiding Bias in VUIs

I’ve been asking Alexa a lot of strange questions lately as I think about forms of bias in machine learning algorithms and whether these will surface in the responses that our voice assistants give to our questions. The short answer is: Yes, yes they will. The difficulty is knowing where exactly the biases are coming from, and how best to avoid them. This post puts in one place my thoughts on this topic over the past 3-4 years.

Here is an essay I published in 2016 on the topic of gender and status in VUIs.

In short, women use “I” pronouns more often than men; also, people with lower status in a group use “I” pronouns more often that people of higher status. So, we should be careful about personifying VUIs as women doing low-status “assistant” work, and then solidifying their identities as women by reinforcing female speech patterns.

Here is a ten-minute talk on this topic at the Interaction '16 conference in Helsinki, Finland:

More recently, I’ve been thinking about the sources of information that Alexa uses in her responses, and also, how she puts together these responses, down to the level of word choice and grammatical construction. I blogged about this in 2018, highlighting some of Alexa's problematic responses to questions about Eric Garner.

Here is an essay on this topic that I published in 2018:

And here is a longer version of the topic that I delivered at the Midwest UX conference in 2019, in Grand Rapids, MI. (I also reprised this talk for the Pittsburgh IxDA in January 2020).

Most companies are investing heavily in voice today, in new and unexpected ways. We can expect to speak to our appliances and other everyday objects more and more in the near future. It's difficult to predict how bias issues will arise when I am asking my toaster to brown my toast "a little bit more." But I have no doubt that bias issues *will* arise.

Header image is a screen capture from the vimeo of me talking at the Midwest UX conference in Grand Rapids, MI, in October 2019.