short answers (450 words)

How Personification and Interactivity Influence Stress-Related Disclosures to Conversational Agents

Abstract

In this exploratory study, we examine how

personification and interactivity may influence people’s

disclosures around sensitive topics, such as

psychological stressors. Participants (N=441) shared a

recent stressful experience with one of three agent

interfaces: 1) a non-interactive, non-personified

survey, 2) an interactive, non-personified chatbot, and

3) an interactive, personified chatbot. We coded these

responses to examine how agent type influenced the

nature of the stressor disclosed, and the intimacy and

amount of disclosure. Participants discussed fewer

homelife related stressors, but more finance-related

stressors and more chronic stressors overall with the

personified chatbot than the other two agents. The

personified chatbot was also twice as likely as the other

agents to receive disclosures that contained very little

detail. We discuss the role played by personification

and interactivity in interactions with conversational

agents, and implications for design.

Author Keywords

Conversational agents; disclosure; stress;

personification; interactivity.

Permission to make digital or hard copies of part or all of this work for

personal or classroom use is granted without fee provided that copies are

not made or distributed for profit or commercial advantage and that

copies bear this notice and the full citation on the first page. Copyrights

for third-party components of this work must be honored. For all other

uses, contact the Owner/Author.

CSCW '18 Companion, November 3–7, 2018, Jersey City, NJ, USA

© 2018 Copyright is held by the owner/author(s).

ACM ISBN 978-1-4503-6018-0/18/11.

https://doi.org/10.1145/3272973.3274076

Shruti Sannon

Brett Stoll

Dominic DiFranzo

Malte Jung

Natalya N. Bazarova

Cornell University

Ithaca, NY, USA

{ss3464, bas364, djd274,

mfj28, nnb8}@cornell.edu

Poster Presentation CSCW’18 Companion, November 3–7, 2018, Jersey City, NJ, USA

285

Introduction

Conversational agents are increasingly becoming the

primary interface for many systems, including those

used for mental health, well-being, and stress

management [4]. These agents vary in form, degree of

artificial intelligence, and function. As part of these

services, users often share personal or sensitive

information with agents that play the role of a

companion, coach, or helper.

As conversational agents become more ubiquitous in

the domain of health, we expect the amount of

sensitive disclosures with technology to increase.

However, the influence of agent characteristics on

users’ disclosures about sensitive information, such as

psychological stressors, is not fully understood. Are

people more inclined to share sensitive, stressful

experiences with conversational agents compared to

surveys, and how do agent characteristics influence

such disclosures? To address this question, this

exploratory study examines interactions between online

participants and one of three agent interfaces that vary

in their levels of interactivity and personification.

Previous work on human-agent interaction has found

that conversational agent characteristics impact

people’s perceptions and conversations. In general,

conversational agents, compared to human partners,

increase human willingness to disclose information [6],

potentially due to lower concerns of evaluation and

impression management around agents. However,

people also send shorter, less vocabulary-rich

messages to agents than to humans [2]. Agent

personalities also impact disclosure: serious and

assertive agents can engender greater disclosure than

warm, cheerful ones [5]. Further research is needed to

parse which anthropomorphizing, or human-like,

characteristics of agents encourage or inhibit

interpersonal disclosure about sensitive topics.

Anthropomorphism has two broad components:

personification and interactivity. Personification refers

to the agent’s human-like cues (e.g., name, physical

appearance, personality). Interactivity refers to the

agent’s communication and ability to converse, such

that messages from the agent are perceived as relevant

to and contingent upon human input [7]. Previous work

has examined anthropomorphic versus non-

anthropomorphic agents, but has not directly addressed

how interactivity and personification shape stress-

related disclosures. As these agents become more

common, understanding how these factors influence

disclosure, particularly around sensitive topics, is

important. Thus our research question:

RQ: How do the personification and interactivity of

online interfaces influence users’ disclosures about

recent stressful events, in terms of a) the nature of the

stressful event disclosed, and b) the intimacy and

amount of information disclosed?

Methods

Data Collection

We recruited participants on Amazon Mechanical Turk

(N=467); participants were told they were testing a

prototype of an online home assistant. Participants

interacted with either 1) a non-interactive, non-

personified Survey (baseline condition), 2) an

interactive Non-personified Chatbot, or 3) an

interactive Personified Chatbot. Participants were told

that to receive personalized service, the assistant would

ask them several questions. The agents asked them

Figure 1. Survey

Figure 2. Non-Personified

Chatbot

Figure 3. Personified Chatbot

Poster Presentation CSCW’18 Companion, November 3–7, 2018, Jersey City, NJ, USA

286

about their interests, followed by an open-ended

question about a recent stressor. The Survey and Non-

personified Chatbot used identical language and only

differed in interactivity: the survey was a static

webform while the chatbot used a chat window. The

two chatbots provided the same level of interactivity

but differed in personification: the Personified Chatbot

had a unisex name, picture, and used more

conversational language. Pictures of each agent and

example utterances are found in Figures 1-3.

Content Analysis

Two coders coded responses to the question about a

stressful experience, and we compared these by

condition (Survey, Non-Personified Chatbot, Personified

Chatbot). 441 responses were coded, with 26 cases

(5.6%) removed for being nonsensical or irrelevant.

Nature of Stressful Event Disclosure. We coded

responses according to the Inventory of Stressful

Events [1], noting the topic of the event (homelife,

work, finances, relationships, health, legal, education,

misc.) (κ=.87) and its ongoing difficulty (temporary,

chronic, uncodable, or non-event) (κ=.77).

Intimacy and Amount of Disclosure. We also coded

responses for the intimacy of the disclosure using a 3-

pt scale (1=relatively public or impersonal to 3=very

private or personal), as in [3]. The amount of details in

each disclosure was coded on a 3-pt scale (1=very little

detail to 3=a lot of detail). Due to skewed distribution,

we dichotomized this variable into “fewer details” and

“more details”. Intercoder reliability was acceptable for

both intimacy (κ=.71) and amount of details (κ=.79).

Results

Nature of Stressful Event

A Chi-square test of independence indicated that the

topics of participants’ stressful experiences varied by

agent type, χ2(14)=27.54, p<.05. We used post-hoc

pairwise comparisons with Bonferroni corrections to

examine specific differences, and found that only the

homelife and finance topics varied by agent type (see

Table 1). Specifically, the Personified Chatbot received

fewer disclosures about homelife than the other agents

(χ2(1)=6.71, p<.01, odds ratio=0.53), and more

disclosures about finances than the other agents

(χ2(1)=6.56, p<.01, odds ratio=1.82).

132 responses provided too little information to infer

whether the stressor was chronic; we analyzed the

remaining 309 responses. The Personified Chatbot

received significantly more disclosures about chronic

stressors than the other agents, χ2(1)=12.02, p<.001,

odds ratio=2.95; there was no difference between the

Non-personified Chatbot and the Survey (see Table 2).

Intimacy and Amount of Disclosure

There was no significant difference in the intimacy of

disclosures based on agent type, χ2(4)=7.82, p>.05. In

terms of the detail provided in disclosures (see Table

3), there was no significant difference between the

level of detail disclosed to the Survey and the Non-

personified Chatbot, suggesting that interactivity alone

did not impact detail in disclosures. However, the

Personified Chatbot was more than twice as likely as

the other agents to receive responses that contained

very little detail, suggesting that personification is the

driving factor in the difference in disclosures,

χ2(1)=15.19, p<.001, odds ratio = 2.81.

Examples of

Participants’

Disclosures by Code

Chronic: “I’ve been dealing

with diabetes”

Non-Chronic: “I had a big

presentation at work last

week”

More detailed: “I've been

really stressed out lately over

my daughter moving out of

the house with her boyfriend.

She's never lived on her own

before and I'm really worried

that she'll need to come

back.”

Less detailed: “Money.”

Home-related: “My

apartment flooded and my

kid’s bedroom was ruined”

Finance-related: “The rent

in my apartment was just

raised and I’m worried I

won’t be able to keep up with

payments”

Poster Presentation CSCW’18 Companion, November 3–7, 2018, Jersey City, NJ, USA

287

Discussion

Findings show that agent personification was a stronger

indicator of both the nature and amount of disclosure

than interactivity alone, such that the Personified

chatbot elicited fewer detailed disclosures of stressors

and more chronic stress events overall. Moreover,

participants shared fewer homelife issues but more

financial stressors with the Personified chatbot.

In [6], they argue that people may make limited

disclosures to agents versus humans due to impression

management concerns. Similar concerns of

interpersonal evaluation may be triggered by a

personified versus non-personified chatbot, and may

explain why participants were less likely to share

homelife stressors—that may be more sensitive—with

the Personified Chatbot. In contrast, limited disclosure

about finance (e.g., “money”, “doing my taxes”) may

be seen by participants as less interpersonally intimate

or risky compared to other stressful topics.

The personification of the chatbot might also set

different expectations about the agent’s expertise and

ability to address certain stressful topics. This could

influence which types of stressful events participants

chose to disclose. People may also view a personified

chatbot as less trustworthy in handling sensitive

homelife information.

These findings suggest careful consideration of how

agents are designed for interfaces geared toward

sensitive disclosures. For discussions of stressful

events, having a simpler, less personified agent may be

better to encourage in-depth sensitive disclosures from

users. Future work should continue to look at the

nuances of interactivity and personification, and how

each might affect sensitive disclosures differently.

Acknowledgements

We thank the National Science Foundation (award

#1405634), and our undergraduate RAs, Brenna

Garcia, Colton Zuvich, and Erin Chen, for their support.

References 1. David M. Almeida, Elaine Wethington, and Ronald

C. Kessler. 2002. The Daily Inventory of Stressful Events: An interview-based approach for measuring daily stressors. Assessment.

2. Hill, Jennifer, W. Randolph Ford, and Ingrid G. Farreras. 2015. Real conversations with artificial intelligence: A comparison between human–human online conversations and human–chatbot conversations. Computers in Human Behavior.

3. Crystal Jiang, Natalya N. Bazarova, and Jeffrey T. Hancock. 2013. From perception to behavior: Disclosure reciprocity and the intensification of

intimacy in computer-mediated communication. Communication Research.

4. Seung-A. A. Jin. 2010. The effects of incorporating a virtual agent in a computer-aided test designed for stress management education: The mediating role of enjoyment. Computers in Human Behavior.

5. Jingyi Li, Michelle X. Zhou, Huahai Yang, and Gloria Mark. 2017. Confiding in and Listening to Virtual Agents: The Effect of Personality. In Proc of IUI.

6. Gale M. Lucas et al. 2014. It’s only a computer: Virtual humans increase willingness to disclose. Computers in Human Behavior.

7. Shyam Sundar. 2008. The MAIN model: A heuristic approach to understanding technology effects on credibility. Digital media, youth, and credibility.

Home Finance

Survey 33% 19%

Non-

personified 28% 22%

Personified 19% 32%

Table 1: Differences in disclosure

about home or finance-related

stressors by agent condition

(n=309)

Chronic Non-

chronic

Survey 13% 87%

Non-

personified 18% 82%

Personified 35% 65%

Table 2: Whether the stressor

disclosed poses ongoing difficulty,

by agent condition (n=309)

Fewer Details

More

Details

Survey 11% 89%

Non-

personified 12% 88%

Personified 27% 73%

Table 3: The amount of

information and context in

disclosures, by agent condition

(n=441)

Poster Presentation CSCW’18 Companion, November 3–7, 2018, Jersey City, NJ, USA

288