A few months ago, I posted a Linkedin survey that asked this question: "How do you think a voicebot should respond to an insult, a slur, or a misogynistic comment?" The survey provided three options and encouraged respondents to provide their thoughts in the comments area.

The three options were:

  1. Call out and push back.
  2. Pretend you didn't hear it.
  3. Quietly end the session.

What's a Designer to Do With Voicebot Responses?

The question occurred to me as I was writing a chapter for an upcoming O'Reilly book on building voicebots that I am co-authoring with Dr. Weiye Ma, a Speech Scientist at MITRE Corporation. In the book, a core principle that we advocate is the proposition that when designing a voicebot, one should avoid trying to always have the bot emulate human behavior. Throughout the book, we insist that human-to-human interactions are fundamentally different from human-to-machine interactions, even if the interactions are mediated by a deeply human communication instrument: natural human language.

But the question of how a voicebot should respond to a human being that was verbally abusive gave us pause. What should the designer have the voicebot do in such instances? Should they have the voicebot call out the abuse and explicitly push back against it? Should they have the voicebot pretend that it didn't understand what the human had said and innocently ask the human to repeat themselves? Or should they quietly terminate the session, in effect ending the conversation without deigning to say an additional word?

Of the three options, the third one ("exit quietly") received the highest number of votes, while the second one ("pretend as if nothing had happened") received the least. 

Related Article: Why Voicebots Continue to Disappoint Us

Potential Voicebot Response: 'Please Be a Better Human'

More than 80% of the respondents did not like the strategy of ignoring the problem. Although they did not agree on how to react to the abuse, they agreed that the abuse should not be ignored.

My own position at the start of the debate was ambivalent. As I wrote then: "Given my basic position that voicebots should not pretend to have feelings, I am leaning towards the ‘ignore strategy,’ but I am voting for ‘end the session quietly,’ because it feels wrong to normalize the behavior by saying nothing.”

As to the option of explicitly pushing back, I was not comfortable with the option and wrote: "I don't know enough to judge whether such explicit pushback is an effective tactic within the larger strategy, not an effective one or one that makes things worse (the offender might get a thrill making the voicebot 'upset,' resulting in the unintended consequence of encouraging rather than discouraging such behavior.").

Since the voicebots are almost always given a female voice, I was especially enlightened to read responses from some female followers of the LinkedIn post.  

For instance, one female respondent to the poll wrote: “Please have the female voicebot respond with a mocking laugh, to say, 'Please be a better human. Your session is now terminated.' Any time I’ve laughed at anyone out loud in disbelief, it left them stunned and quiet."

A combination of the first and third strategy. I hadn’t thought of that.

Learning Opportunities

What Response Context Does Your Brand Desire?

Another respondent pointed out that the context of the infraction is important. She wrote: "It depends on the impact you want the voicebot to have on users. If you just want to stick to the practical functions of the bot, then abuse can be ignored or dealt with fallback answers. However, if you want to have an educational impact on the users, you can draft responses that make the user reflect on verbally abusive queries. This does not necessarily imply admitting the bot has feelings. It can be based on the bot recognizing abusive language and taking a stance against it."

The suggestion has merit, especially when it comes to voicebots engaging with children.

Related Article: We Know Chatbots Are Falling Short, But Why?

Why Designers Should Not Design to Accommodate Abusive Users

As for where I stand as of now, I've had the benefit of a few months of thinking and engaging other designers. Paradoxically, my position is now less nuanced than the one with which I started.

I now believe that the designer should spend no time or effort designing for any such abuse; they should instead spend all of their resources on designing for the polite, earnest user. The abusive user does not deserve to be acknowledged in any way, and having the voicebot respond by simply telling the user that they don’t understand what was said, and then politely asking them to repeat themselves, will get the message across that they are wasting their time and might as well be abusing a wall.

I do not feel that such a reaction would normalize the abuse, as I had felt initially. Normalizing means acting in a way that made the voicebot behave as if it understood what was said to it and did not mind what it was hearing.

Instead, the voicebot, by altogether literally not understanding what was said to it is in effect communicating the language that was said was not expected and is therefore not normal. Additionally, it also signals that those who designed and built the voicebot did not spend any time thinking about how to react to abusive language.

That, for me, is enough pushback from a helpful non-human to an abusive human, especially since it respects the clean distinction between human-to-human behavior and human-to-machine behavior.