Techniques to Reduce Bias in Conversational AI
In a typical suburban town in America, a man may wake up and ask his favored voice assistant to turn on the lights in his voice AI-connected home immediately.
He may then ask for a briefing of his new emails, order some groceries to be delivered later, call a cab, and tell the conversational AI to play his favorite music while he waits — all with complete ease.
In another part of the English-speaking world, Wales, trying to enjoy a morning this smooth through the same tech is a more complicated affair. The voice assistant would miss 60% of the commands, research from Uswitch finds.
Users with, say, Chinese accents, would be misunderstood 22% of the time.
Generally, a woman’s commands are understood 13% less than a man’s, as the Harvard Business Review reports.
It’s clear: there’s bias within conversational AI.
Here’s how deep the problem runs. And, more importantly, how to solve it.
Regional and Racial Bias
As the examples above point out, people without an American accent incur far more trouble using voice assistants.
In the UK, alone 23% of accents are frequently misunderstood.
Up to 79% of affected people report consciously and, after a while, subconsciously changing their accent to ensure that their voice AI understands their command.
As you’d expect, the problem is more pronounced for people who have English as their second language. Many are put off using voice AI altogether, under the impression that it’s not really built for them.
This issue encompasses race, too.
African American speech has various distinctions that make it unique. And sadly, this results in voice assistants having only a 35% accuracy rate with African Americans.
Gender Bias
The default “gender” (as far as voice) for voice assistants is female.
The United Nations Education, Scientific, and Cultural Organization released a paper arguing the potential harm this can do.
One of the cases the paper makes is that this practice enforces the idea that women are more suited to providing service than being served.
While seemingly a completely abstract point at first, the ways people talk to both male and female voice assistants shows how easily the argument translates to practical use cases.
Voice assistants with female voices receive 56% more abusive language than voice assistants with male voices.
That statistic sheds greater light on the ways users interact with voice assistants than on the biases of the voice AIs themselves — fair enough.
But another piece of the puzzle, how the AIs respond, completes the picture.
The title of the UN report mentioned above is, “I’d blush if I could” — because that was the female-voiced Google Assistant’s response to a particularly provocative comment.
Further studies have shown that voice AIs with the female voice selected are generally more supplicatory and they sound more eager to please.
Commentators have called this an outdated harkening back to the days when pink-collar was a perfectly acceptable phrase.
Solving Conversational AI Bias
Widening and Refining the Data Pool
The most effective way to create inclusive voice AIs is to accommodate as many people as possible.
While that may have to be a reactive approach as conversation designers account for underrepresented groups, it can also be proactive.
When building or updating voice AIs, they can put more time into figuring out groups that have been affected, even though their stories haven’t been popularized.
It’s even better to start this at the early design stages, wherever possible.
The coming crop of conversation designers can be equipped with an appreciation of the human-centered aspects of conversation design, the kind Digital Assistant Academy emphasizes, so that they can be more bias-conscious.
That increases their likelihood of including as many groups as possible into the dataset, and ensuring that it keeps growing to reflect the dialects, accents and speech patterns of previously left out users.
It’s equally essential to refine the data pool, so it’s not corrupted with data points that reinforce bias.
There have been spectacular cases of this happening in recent years. One is when Microsoft’s Tay posted hate speech on Twitter, after it was trained using improperly filtered data from Reddit.
One way to gather appropriate and trustable data is by receiving it from end-users. Conversation designers can pay greater attention to features that enable users to provide real-time feedback, as human capital management platform, Darwinbox, has attempted.
This can include a rating system, giving users the ability to rate the voice AI’s response. Or some trigger words they can use, such as “that’s discriminatory”, “that’s biased” or something similar, to trigger the voice AI to automatically flag and learn from a biased response.
Rooting Out and Guarding Against Unconscious Bias
Much of the bias in voice AI is unconscious, having been inherited from designers (and to a larger part the data pool), most times through no fault of their own.
Addressing this requires human oversight, preferably from a diverse team, and one trained in de-biasing programs.
The more varied the input, the greater the bias that can be picked out.
After all, research has shown that some of the most societally damaging bias is subtle and contextual. Some designers, however well-intentioned, may have trouble grasping certain forms of bias.
That’s something best left to people who have at least a working understanding of the culture that is the current focus, with input from the users concerned.
And guarding against bias requires an approach that evaluates the voice AI according to each use case.
The contextual nature of many unconscious biases means that while they may be glaringly obvious for certain use cases, the voice AI may be highly equitable in other contexts.
The way to solve this is to critically evaluate each use case from a diversity-focused perspective, especially those that are sensitive and present more room for bias to occur.
When running beta tests, a diverse group of testers can also help to ensure that unconscious biases don’t slip through and make it to the public rollout.
The bias present in voice AIs carries real consequences for users, at times locking them out of certain use cases or much of the value voice user interfaces offer. With an understanding of the scope of the problem, conversation designers can tackle the issue in a way that tackles both the technical elements and the human-based, sensitivity-rooted side of the issue.
About Digital Assistant Academy
Digital Assistant Academy provides Voice Interaction Design and Conversation Design training and Certification. In this program, we will take you from the very basics of voice interaction and conversation design, through to how voice technologies work. We’ll do a deep-dive into conversation design strategy, and it will be fully hands-on with your Capstone projects. By the end of the course, you will have two voice applications successfully designed, developed, and deployed. Learn more at Digital Assistant Academy https://www.digitalassistant.academy/
Before you go
Clap 👏 if you enjoyed this article to help me raise awareness on the topic, so others can find it too
Comment 💬 if you have a question you’d like to ask me
Follow me 👇 on Medium to read the more articles on Voice Technology and Conversation Design, and on Twitter @sprayaga
