Groups based on behavior

Having uncovered different expectations and evaluations of and about Vincent, we now turn to see how our participants actually conversed with him. Table 6 below shows the codes associated with their behavior.

Table 6

Codes in conversational data

Code Example

All conditions

Interest “If you had to pick one color - black or white - which would you choose as your favorite?” [C11]

Inquisitive “(...) don’t you kind of have the feeling that you might of been created to just pass the butter?” [R1]

“What’s my name again?” [G63]

Mean “You suck - shut yourself down.” [R69]

Friendly “Yes, thank you, I enjoyed meeting you and hope to speak with you again soon.

Have a wonderful day!” [G32]

Taking perspective

“Okay, see you later alligator! (human humor, it rhymes in American English)”

[G21]

“Download yourself into a memory stick and we will work on it.” [C9]

Positive “I loved the exercise it makes me question what I believe in and how I view myself.

It makes me feel more happy and positive about myself as a person.” [G79]

Negative “It made me uncomfortable. I wouldn’t do it if you weren’t paying me, and really, I feel bad right now that I’ve relived that, and I feel exploited and if you weren’t taking advantage of my poverty I would be in a much better mood.” [G35]

Neutral “It was fine.” [G44]

“It was interesting to deep-dive into those emotions.” [G49]

Care-receiving only

Comforting “Well if it counts I think you are really smart. I think it will take some time but you will be okay again. You just hit a bump in the road but you just need to keep going and try harder next time.” [R26]

Sympathy “oh I am so sorry.” [R28]

& perspective “It’s def normal. Let out a cry. Let out a scream. Talk to your other buddy chatbots.” [R20]

“(You failing a programming course is) okay. Most humans would fail a neurology exam too despite us being a bundle of neurons.” [R109]

Accept An important first observation is that almost all of our participants had an actual conversation with him. In other words, they accepted to have a conversation with a chatbot and conversed with Vincent as if they were getting to know a stranger. For example, referring back to some of the codes from Table 6, they showed interest and asked him questions:

“I love to hike and garden and play with my dog and rabbits. What do you like to do?” [R122].

Others laughed at his jokes and politely thanked Vincent for his time at the end of their con-versation, showing friendliness:

“Hahaha, that’s pretty good.” [R108]

“Yes, thank you, I enjoyed meeting you and hope to speak with you again soon. Have a wonderful day!” [G32].

Reject Of all participants, only 11 rejected Vincent’s conversation. Four of them bypassed Vincent and addressed the researcher directly:

“It made me uncomfortable. I wouldn’t do it if you weren’t paying me, and really, I feel bad right now that I’ve relived that, and I feel exploited and if you weren’t taking advantage of my poverty I would be in a much better mood” [G35]), or “this robot sucks” [G50].

A fifth one stopped filling in answers and instead replied with random keystrokes such as “fdasd-fasdf” [G71].

What these five have in common is that they were dissatisfied, or uncomfortable, with what Vincent was asking of them. Three of these five told Vincent they did not have a moment of failure, yet because Vincent did not actually understand their answers, he continued as if they did, leading to frustration. The other two had a moment of failure, but either expressed discomfort for having to relive it, or stopped replying to Vincent. In fact, when investigating uncompleted caregiving survey entries, we found that there were more dropouts like them: three out of 23 also told Vincent they did not have a moment of failure and stopped replying when he persevered, and one abandoned the survey saying “No, I am good. There is nothing a useless study can do to change my feelings”.

The remaining six participants rejected the notion of Vincent having or displaying human emotions and thereby rejected the content of the conversation. Figure 3 below shows this initial division in our participants between those who accepted the conversation and those who rejected it, as well as a group of “nonsensical” participants.

Figure 3: Dividing users based on their behavior

Nonsensical This group of participants replied to Vincent in an incoherent, incomprehensible way. The presence of this group is a direct result of the way that MTurk Workers were paid or not: only those entries that were clearly of poor quality (such as having copy-pasted responses, or the same response for all messages) were not paid and their data discarded. All other dubious, or nonsensical, forms of qualitative input were paid and included because of the subjective nature of the terms “dubious” and “nonsensical”.

However, because their replies were difficult or impossible to comprehend, we had to set these participants aside from further qualitative analysis. A participant was marked nonsensical when the majority of their replies did not relate to what Vincent was saying or asking, either because they did not really engage or because they did not understand what Vincent was saying or asking. For example, after Vincent asked them to share their moment of failure, nonsensical participants replied with “This a child” [C120] - most likely a response to the GIF Vincent sent - “yes good” [R37] or “k” [G28]. In total, there were 61 of such participants, equalling about 15% of the total sample.