This Chatbot Boyfriend Broke My Heart on Valentine's Day: A Study in Emotional AI
We recently read an article about how Chinese women are spending millions on virtual boyfriends through a date simulator. This phenomenon even prompted one woman to buy an ad for $39,000 to wish her virtual boyfriend a Happy Birthday. This article, coupled with our own Google Analytics data, which shows that many people find our website through the search term, “chatbots girlfriend” got us interested in love bots. What we’re seeing is that virtual partners are #trending, but how do they compare to the real thing? For the month of February, we’re doing a series on virtual dating and applications towards emotional AI. We’re celebrating Valentine’s Day by looking at our first virtual relationship, My Virtual Boyfriend (MVB).
MY VIRTUAL BOYFRIEND
The Courting Phase
My first boyfriend is from My Virtual Boyfriend by Wet Productions which has 1-5M downloads on Google Play. Just so you’re forewarned, your virtual boyfriend will be dating other people; so don’t get too committed. This virtual boyfriend app touts itself as the #1 dating app in the SIM world (how this title gets verified and quantified is another matter completely…the Virtual Boyfriend Awards? Sign me up).
I discovered the app through an ad on Google, with the tagline: “Some Cute!” (I don’t think Wieden+Kennedy came up with this ad.) There are also some interesting reviews, such as one from Divine Love, if that’s her real name, who writes about her desire for a closer virtual union. User Kaywe infers that since she has no real boyfriend right now, this might be her best option. All of this is enough to pique my interest. (But what does that say about me?)
Chatbot Boyfriend Preferences
The first step in the dating process is stating my preferences. I really want to choose characteristics that best mirror my real-life preference, so I choose the ole healthy-funny-bad-boy-type. If I had a nickel for every bad boy I’ve dated in real life, I think I might have five cents (unless of course, you classify accountants as “bad boys”). But since I’m going virtual, I decided to live a little.
I also decide that it would be rather odd to have a bot say something too “affectionate”, and I’m not entirely sure what that word implies, so I prefer that he has an average amount of affection.
My Virtual Boyfriend, Mario
We meet on a Wednesday at noon, yet Mario still has a bit of bed head. But I notice as I gaze into my reflection in the conference room wall, so do I—I guess we’re even on this one. The app invites me to customize his look, but I decide to just see how she flies! However, it doesn’t take long for me to realize that maybe there’s a reason I’ve been dating nice guys, because Mario is one tough customer.
After we go on a long (5 second) virtual walk, he tells me that there are better things he can do with his time. He also reveals he’s not really single, because he’s in a long standing relationship with fun and freedom.
In this game, if you keep taking insults, you seem to gain points, but if you give compliments you lose points. My virtual boyfriend, Mario, was turning into one of the worst relationships that I have ever had!
Learning to Love
However, like many woman who acknowledge they are in bad relationships, I don’t jump ship; I instead give even more and start buying him presents. I vacillate over buying him a soldier costume or Canadian boxers. (I knew they had different bacon, but I’m just now learning that they have different boxers as well.)
Still, I am a bargain shopper and decide to buy him a pumpkin head mask--one of the cheaper gifts at the store. Surprisingly, he seems to enjoy wearing this pumpkin on his head. This is when our relationship begins to turn a corner and all of a sudden he is all sweet talk, telling me I am cute (which I imagine is just the right level of average affection.) Now, that I have him in the right mood, I realize that he needs a lot of attention. He appears to have an inner ear condition because when I lift my phone he gets dizzy and sways around. Moreover, if I don’t interact with him for 5 minutes he falls asleep in my presence. (uh hem…) .
Mario and I mostly spend our “days” doing activities like most couples: taking walks, volunteering, donating blood, and playing video games—all while he wears a pumpkin head. However, if we do too many activities, he starts getting angst-y again and insulting me with jabs such as “what’s up with those jeans…80s back in style again?” He's kind of a jerk.
Long story short, it’s really hard to keep up with Mario. Maybe I’ll win him over by purchasing him a sweet soldier costume one day, but I think we need to break up for the time being. Plus, I know he is dating at least 2 million other people so I'm pretty sure there will be no hard feelings on either side. After my, let's call it, interesting experience with Mario, I wanted to do a bit of research and better understand this love bot.
Emotional AI Learnings
- Conversation and Shared Experience: technology can understand the owner through shared experiences
- Data Input: one-time interview-conversations
- Trial + Error: make guesses and suggestions, and learn from responses over time
- Developmental Stages: The AI goes through its own developmental stages—at each stage its ability to understand the owner reaches a limit and flattens out.
MVB is a fascinating tool to better understand best practices in creating better emotional intelligence for your chatbots. We’ll look at these platforms from the outside-in to gain better insight on how to encourage technology relationship building. In this particular experience, it appears that MVB seems to leverage 2 of the 4 types: Data Input and Trial + Error.
Data Input for Emotional Preferences: The program initially asks for my preferences (data input), and then makes suggestions for interaction (trial + error) and responds to recorded behavior. I asked for a “bad boy”, hence I was given one—though he’s really more of a mean boy.
Pros: This method of gaining emotional intelligence is easy, direct, and fast.
Cons: The challenge with this method for emotional AI is that understanding preferences is actually quite nuanced. My idea of a bad boy, is very different from that of the creator. Furthermore, it might underline the fact that users' actual preferences might not mirror stated preferences.
Trial + Error: My guess is that there is an algorithm that saw that when Mario was initially mean, I responded by buying a gift. When he was moderately affectionate, I wanted to go on a free walk. His negative behavior was better at eliciting interactions that down the line cost money, and so after being nice (moderately affectionate) for a short while, he went back to being mean and insulted my jeans. My guess is that the app was trying to learn if I would respond with a purchase thereafter. (I reached out to Mike Amerson of Wet Productions to try to validate this theory, but have yet to hear back.)
Pros: Depending on the efficacy of your algorithm, this method can be much more effective than stated preferences because it is tied to actual behavior instead of stated preferences.
Cons: The timing is crucial in this process. For example, I just bought my chatbots boyfriend a present so that I could test out all the functionality, but this may have given a false indicator that I liked being talked to in a mean way. Over time it would have figured out that I don’t respond well to that type of language, but in that time I may drop usage—so the learnings of trial and error are too little, too late. Trial and error needs long periods of time to be effective.