• No results found

The junction between privacy and artificial intelligence : information disclosure to online chatbots Iiin relation to data privacy concerns

N/A
N/A
Protected

Academic year: 2021

Share "The junction between privacy and artificial intelligence : information disclosure to online chatbots Iiin relation to data privacy concerns"

Copied!
66
0
0

Bezig met laden.... (Bekijk nu de volledige tekst)

Hele tekst

(1)

The Junction Between Privacy And Artificial Intelligence:

Information Disclosure To Online Chatbots In Relation To Data

Privacy Concerns

Thesis Digital Business MSc. Business Administration

University of Amsterdam

Student name: Lara S. E. Neervoort Student number: 10587403 Supervisor: Prof. Dr. Hans Borgman

Word count: 12.067 words Date: 21 June 2018

(2)

Statement of Originality

This document is written by Lara Neervoort, who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in

creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

Acknowledgement

Hereby I would like to thank my supervisor Prof. Dr. Hans Borgman of the Amsterdam Business School for his guidance and advise during the course of writing this thesis and for the valuable feedback on my work. Furthermore, I would like to thank Marelle van Beerschoten, Elles te Riet and Aline Swinkels of the Dutch

company Digital Shapers for giving me the opportunity to perform the first version of my experiment at their #ShapingDigital event on the 21st of March, and for providing me with valuable feedback during the course of my thesis.

(4)

Table of Contents List of Tables ... 6 List of Figures ... 6 Abstract ... 7 Introduction ... 8 Literature Review ... 12

Privacy Concerns and Individual’s Disclosure of Information ... 12

The Contextual Aspects of Privacy ... 14

The Implementation of Chatbots in Relation to Privacy ... 17

Summary of hypotheses ... 19 Conceptual model ... 20 Method ... 21 Measures ... 21 Privacy concerns ... 21 Privacy context ... 22

Personal information disclosure ... 25

Procedure ... 30

Pre-test experiment ... 30

Online experiment ... 31

Control variables ... 33

Results ... 35

Preparing data for analysis ... 35

Assumption checks ... 37

Manipulation check ... 38

(5)

One-way ANOVA ... 41

Factorial ANOVA ... 44

Hierarchical multiple regression ... 45

Discussion ... 48

Implications regarding the privacy paradox ... 48

Implications regarding the privacy context ... 50

Contributions to theory and practice ... 53

Limitations and suggestions for future research ... 56

Conclusion ... 58

(6)

List of Tables

Table 1: Means, Standard Deviations, and Correlations

Table 2: Total information disclosure for levels of privacy concerns Table 3: Factorial ANCOVA

Table 4: Hierarchical multiple regression analysis

List of Figures Figure 1: Conceptual model

Figure 2: Chatbot Lucy developed on the platform Landbot.io Figure 3: Online questionnaire developed on Typeform.com

Figure 4: Questions included in the pre-test experiment in the chatbot and Typeform condition

Figure 5: Questions included in the online experiment in the chatbot and Typeform condition

Figure 6: Conceptual model including constructs at measurement level Figure 7: Conceptual model including control variables

Figure 8: Proportion of respondents that provided each sensitive item, broken down by the condition

Figure 9: Conceptual model including control variables taken into account in further analyses

(7)

Abstract

This research experimentally investigates the relationship between privacy concerns and information disclosure in the context of conversing with a chatbot. This

investigation contributes to the research gap on the combined topic of chatbots and privacy, and provides further evidence in the divided body of literature on the privacy paradox. Furthermore, this study has significant managerial implications since

managers need to mind the user’s privacy boundaries when implementing new technological tools such as chatbots, or privacy scandals can occur that can hurt the brand’s reputation and even financial status (Spiekermann et al., 2015). Contrary to the first hypothesis, it was found that higher privacy concerns do not lead to lower information disclosure in both the online questionnaire and the chatbot condition. This finding is in line with the privacy paradox (Norberg, Horne & Horne, 2007). Furthermore, the privacy context did not have a moderating effect on the relationship between privacy concerns and information disclosure as was hypothesized, but did have a direct effect on information disclosure. Participants actually disclosed slightly more information to the online questionnaire than to the chatbot, for which possible explanations will be provided. These are novel results in the field of information disclosure towards chatbots, a field that was not sufficiently researched yet.

Moreover, this research provides managers with valuable insights into what kind of information consumers are and are not willing to disclose, and whether to implement an interactive chatbot or simple online questionnaire when retrieving different types of information.

(8)

Introduction

This thesis experimentally explores the field between privacy and artificial intelligence: how is the relationship between users’ privacy concerns and their personal information disclosure moderated by the privacy context in which this process occurs? We focus on the context of conversing with an interactive chatbot contrasted by filling in a simple online questionnaire and experimentally investigate the amount of personal information the user discloses in each context. This issue has large implications for both theory and practice, as there is a research gap in the field between that of privacy and the use of chatbots. Furthermore, managers can learn whether to better implement a chatbot or an online survey when trying to retrieve certain personal information, and will find out where their customers’ privacy boundaries lie when implementing chatbots. The remainder of this introduction will discuss the background of this problem statement and will further specify the managerial and academic relevance.

The field of artificial intelligence, and in specific that of chatbots, is quickly expanding its impact on various domains in science, business, and technology (Serban et al., 2016). On the 15th of March 2016, Google DeepMind’s AlphaGo computer program beat the world’s champion player, Lee Sedon, at the extremely complex Chinese board game of Go (Bae et al., 2017). This victory of a computer program over a human Go master is seen as a major scientific breakthrough in the field of artificial intelligence because AlphaGo used reinforcement learning and neural networks to imitate the learning process of the human brain, and thereby beating the human expert in the game (Chen, 2016). This breakthrough was believed to be still a decade away, and since then the field of artificial intelligence and machine learning is developing with increasing speed (Chen, 2016).

(9)

Not only the scientific world is benefitting from the developments in artificial intelligence (AI) and machine learning, but also companies are quick to find the application of AI and machine learning for their organizations (Serban et al., 2017). Discoveries in the field of artificial intelligence and machine learning have enabled the development of chatbots: computer programs that can converse with humans and use human language as input to learn from (Jia, 2009). Chatbots apply the

mechanisms of deep learning and natural language processing to accurately respond to a consumer’s inquiry and to imitate a human-like conversation style (Reshmi & Balakrishnan, 2016).

Chatbots are currently being implemented by organizations in multiple ways: e.g., answering customers’ messages on Facebook, offering technical support on company websites, and even acting as automated online therapists for consumers with psychological problems (Serban et al., 2017). According to Business Insider, chatbots have the ability to replace 29% of customer service positions and 36% of the sales representative positions in the United States (BI Intelligence, 2017). As companies but also governments, insurers, and banks are increasingly implementing chatbots in their processes (Letheren & Dootson, 2017), it is important to investigate whether consumers are willing to disclose certain sensitive information to chatbots. This study will link the implementation of chatbots to the field of consumer privacy.

The field of user data privacy is an increasingly important topic in an era where personal user data is becoming a primary asset that companies exploit to maximize their profit (Esteve, 2017). In this time of personal data capitalization, companies need to balance their need for profit by exploiting personal data sources with the user’s need for privacy (Spiekermann et al., 2015). Corporations need to mind the user’s privacy and need to know where their consumers’ privacy boundaries

(10)

lie when disclosing private information online, or privacy scandals will arise that can immensely hurt the brand’s reputation and even its financial well-being (Spiekermann et al., 2015). Current privacy scandals involving Facebook and Cambridge Analytica show how much such a scandal can negatively impact a brand (Confessore, 2018). Consumers’ privacy boundaries become even more important when they are asked to disclose personal information to new technological tools such as chatbots, and might act differently than when being asked to disclose personal information in a survey (Spiekermann, Grossklags & Berendt, 2001).

A lot of research has been conducted on how individuals’ privacy concerns influence their personal information disclosure (Jensen et al., 2005; Baek, 2014; Taddei & Contena, 2013), but this needs to be further investigated taking into the account the context of conversing with a chatbot. This study connects the fields of user data privacy with the emerging field of artificial intelligence and chatbot technology. To address this research gap, the research question that this study poses is: What effect does the privacy context (either conversing with an online chatbot or

filling in a simple online questionnaire) have on the relationship between the user’s concern for privacy and the user’s online personal information disclosure?

This investigation does not only have a significant scientific contribution in adding to the research gap in the field of privacy and artificial intelligence but also a significant managerial implication. Managers need to mind the balance between their need for profit and the consumer’s need for privacy, or privacy scandals will damage the brand’s reputation and its financial status (Spiekermann et al., 2015; Confessore, 2018). For managers, it is essential to know where customers’ privacy boundaries lie, especially when implementing new technological tools in the form of chatbots

(11)

not only lead to damage to the brand’s reputation, but also to significant financial damage to the company in terms of a drop in the firm’s market value and fines from the government (Goel & Shawky, 2009). Furthermore, this study helps managers in choosing whether they should implement a simple online questionnaire or an online chatbot when trying to request sensitive information from users, something that becomes even more important when companies but also banks, health insurers, and governments are choosing whether or not to implement a chatbot in their processes (Letheren & Dootson, 2017).

(12)

Literature Review

To answer the research question of this study, we will review the various perspectives in the literature on the different constructs implemented in our research and will combine this literature in order to develop valid hypotheses. To answer the research question: “What effect does the privacy context (either conversing with an

online chatbot or filling in a simple online questionnaire) have on the relationship between the user’s concern for privacy and the user’s online personal information disclosure?” we will first consider the relevant literature in the field of privacy and

the consumer’s disclosure of sensitive information. In this field, two opposing bodies of research exist that will be discussed as well as the cognitive mechanisms

underlying the two perspectives. Secondly, we will review the literature on artificial intelligence and chatbots that is relevant to this study. Finally, we will link the topics of privacy and artificial intelligence together and discuss the existing research that investigates this combined topic. This body of research is still limited, which forms a valuable research opportunity for this study.

Privacy Concerns and Individual’s Disclosure of Information

In this study, we define privacy concerns as “individuals’ beliefs about the risks and potential negative consequences associated with sharing information” in line with the definition implemented by Baruh et al. (2017, p. 27). The existing literature on privacy concerns and individual’s disclosure of private information shows two important strands of literature: on the one hand, a lot of researchers have found evidence for the “privacy paradox.” The privacy paradox entails that despite the fact that users report to experience high concerns for their online privacy, they still

(13)

privacy concerns (Taddei & Contena, 2013; Norberg, Horne, & Horne, 2007; Tüfekci, 2008).

On the other side of this field of research there are authors reporting that individuals’ concern for privacy does actually correlate with their online disclosure of personal information, meaning that higher concern for privacy correlates with a lower degree of sharing information online (Son & Kim, 2008; Lutz & Strathoff, 2014; Krasnova, Spiekermann, Koroleva & Hildebrand, 2010). This body of research argues for quite the opposite of the studies that are in line with the privacy paradox.

However, what is important when investigating these opposite schools of thought on the relationship between privacy concerns and sharing information is understanding which cognitive mechanisms are underlying consumers’ decisions to share

information or not.

One prominent study that investigates these underlying cognitive mechanisms is Dinev & Hart’s “Extended Privacy Calculus Model” (2006). The authors develop their model around different factors that influence the user’s decision making process or “calculus” in deciding whether or not to disclose certain private information online (Dinev & Hart, 2006). This privacy calculus can be seen as a cost-benefit analysis that individuals execute, and when the benefits of disclosing information online override the disadvantages of doing so, the user will engage in the disclosing behaviour (2006).

Recently, a study by Dienlin & Metzger (2016) has confirmed the

generalizability of the extended privacy calculus model. The model was found to be generalizable for self-disclosure on social media networks as well (Dienlin &

Metzger, 2016). In short, the privacy calculus model can be seen as an explanation of why users choose to share information or not: they “calculate” whether it is worth it to compromise their privacy by sharing information (Dinev & Hart, 2006).

(14)

After reviewing both sides of research that argue in favour (Taddei &

Contena, 2013; Norberg, Horne, & Horne, 2007; Tüfekci, 2008) and against (Son & Kim, 2008; Lutz & Strathoff, 2014; Krasnova, Spiekermann, Koroleva & Hildebrand, 2010) the privacy paradox, we have focused on a recent meta-analysis by Baruh, Secinti & Cemalcilar that reviews 166 studies on the privacy paradox (2017). The authors (Baruh et al., 2017, p. 36) conducted a meta-analysis of 166 studies from 34 countries on online privacy concerns and disclosure of personal information, and found that “users with higher privacy concerns shared less personal information … and utilized privacy-protective measures more.” The authors found that these findings can be generalized across gender, national legal systems and cultural orientation (Baruh, Secinti & Cemalcilar, 2017). Next to privacy concerns and information sharing, the authors also took into account the variables of privacy

literacy and intention to use protective measures (Baruh, Secinti & Cemalcilar, 2017). Taking into account these variables and different nationalities and cultural

orientations, Baruh, Secinti & Cemalcilar indeed found after reviewing 166 studies that ‘contrary to the premise of the concept of “privacy paradox,”’ those with higher privacy concerns were less likely to share personal information online (2017, p. 45). Therefore, we hypothesize in opposition to the privacy paradox that:

H1: Higher privacy concerns lead to a lower disclosure of personal information

The Contextual Aspects of Privacy

One thing that should be taken into account when evaluating privacy concerns and disclosure of private information is the context in which this process occurs. There has been a call for researchers to investigate the influence of the context on the relationship between privacy concerns and information disclosure (Nissenbaum,

(15)

2010). To quote Kokolakis from his meta-analysis on the privacy paradox (2017, p. 127): “privacy behaviour is a highly contextual phenomenon.” Morando et al. (2014) also discuss the importance of contextual factors on privacy concerns and the

following behaviour in their article (p. 3). Adding on, Norberg, Horne & Horne state that users rely on “salient environmental cues” when deciding whether to share information or not (2007, p. 109). Furthermore, Belanger & Xu (2015) argue that, therefore, “IS (information systems) researchers need to explore the contextual nature of information privacy, with a goal of better understanding what constitutes

information privacy or privacy violations under different contexts” (p. 576). As highlighted by the literature discussed above, the context in which individuals are asked to disclose impacts the relationship between their privacy concerns and their disclosure of sensitive information.

To further specify this statement, different articles state that the specific online platform on which the information disclosure occurs has an impact on users’ privacy concerns and information disclosure (Palen & Dourish, 2003; Stutzmann, 2006; Krishnamurthy & Wills, 2008; Schrammel, Köffel, & Tscheligi, 2009; De Choudhury, Morris & White, 2014). Palen & Dourish explore five different technologies in their influential article “Unpacking “Privacy” for a Networked World” and investigate how these technologies influence information disclosure and privacy concerns (2003). The authors argue that they “illuminate specific issues in the interaction of privacy and information technology” (p. 135), and explain how

different technologies ask for a “more nuanced understanding of privacy concerns” (p. 134) (Palen & Dourish, 2003).

Adding on, Stutzmann (2006) compared information disclosure on different online platforms and found that users disclose more information on Facebook

(16)

compared to MySpace and Friendster. Next to that, Krishnamurthy & Wills (2008) compared eleven social media platforms and found that the information disclosure significantly differed between networks, and found that information was shared substantially more in smaller types of networks.

Furthermore, Schrammel, Koffel & Tscheligi compared different kinds of online platforms and evaluated how these types of media influenced information disclosure: business networks, social networks, content & media sharing, and social news & bookmarking (2009). The authors found that the type of medium had a significant impact on the amount of information shared, as participants shared significantly more information on social networks than on other channels (Schrammel, Koffel & Tscheligi, 2009).

Finally, De Choudhury, Morris & White (2014) compared sharing health information in the technological context of a search engine to the technological context of social media and found that the type of information as well as the volume of data that participants shared differed. In this study, we specifically focus on the context of conversing with an interactive chatbot versus filling in a simple online questionnaire. Based on the literature discussed above, we hypothesize that the online context in which individuals are asked to disclose impacts the relationship between their privacy concerns and their disclosure of sensitive information.

Therefore, we hypothesize that:

H2a: The online context of the disclosure (either conversing with an interactive chatbot or filling in a simple online questionnaire) has a moderating impact on the relationship between privacy concerns and information disclosure

(17)

The Implementation of Chatbots in Relation to Privacy

In this study, we focus on the specific context of conversing with an

interactive online chatbot versus filling in a simple online questionnaire. We define a chatbot as “an artificial construct that is designed to converse with human beings using natural language as input and output,” in line with the definition used by Jia (2009, p. 250). The body of research on chatbots and artificial intelligence is wide-ranging and spans from the application of chatbots to the university admissions process (Polatidis, 2014) to a “movie chatbot” that uses Star Trek dialogue style and is able to reference the movie in its conversation with customers (Jena, 2017). The majority of the existing articles on artificial intelligence and chatbots discuss the options for designing and programming chatbots (Shawar & Atwell, 2007; Huang, Zhu & Yang, 2007; Yan et al., 2016). However, there is a limited amount of research that relates the use of chatbots to data privacy, or to information disclosure of the consumer.

The remainder of this literature review will consider the articles that are available on the topic of chatbots and data privacy. A recent article by Letheren & Dootson (2017) discusses the new application of chatbots to the banking industry. Letheren & Dootson analyze the dilemma between the convenience of using such a new technology and the customers’ feeling of security (2017). The authors argue that this dilemma boils down to the trust the customers feel towards the bank: “Consumers need to know that their trust … is well placed before they can enjoy the added

convenience of emerging technologies” (Letheren & Dootson, 2017, para. 6). Physical banks have been designed to signal trust by including vaults and marble floors, but it is key to also build chatbots in a way that they provide a sense of security to ensure the user’s trust (Letheren & Dootson, 2017). Making the interaction

(18)

between chatbot and user more human-like could help in inspiring trust in the human user of financial chatbots (Letheren & Dootson, 2017).

Continuing the review of the literature existing on this topic, “Emergent AI, Social Robots and the Law: Security, Privacy, and Policy Issues” discusses the issue of privacy in relation to artificial intelligence and social robotics (Subramanian, 2017). The author discusses how the data that the bots store and process can be private financial and health data, and how it is most important that these robots are designed for security and discretion (Subramanian, 2017). Next to the notion of data privacy, the author discusses how chatbots and other social robots should also mind a different kind of privacy: they should learn how to mind the consumer’s “personal space” and maintain the right distance from the user in order to respect this emotional kind of privacy, also in the questions they ask the consumers (Subramanian, 2017). This notion of personal space should also be kept in mind in the questions that consumers are and are not willing to answer in response to a chatbot.

The most relevant article on the usage of chatbots and its relation to privacy concerns is that of Spiekermann, Grossklags & Berendt (2001). The authors

conducted an online shopping experiment in which the participants entered a dialogue with “an anthropomorphic 3-D shopping bot,” something that we would nowadays call a chatbot (Spiekermann, Grossklags & Berendt, 2001). The researchers contrasted the reported privacy concerns of the participants with their actual disclosure

behaviour in a conversation with an e-commerce chatbot (Spiekermann, Grossklags & Berendt, 2001). Almost all participants reported high privacy concerns when asked about their privacy attitudes in a survey, but readily revealed large amounts of information when conversing with the “3-D shopping bot” named Luci. Also, a differentiation in the privacy statement that was included did not impact how much

(19)

information the participants shared with the chatbot (Spiekermann, Grossklags & Berendt, 2001). These results indicate that Spiekermann, Grossklags & Berendt (2001) found evidence for the privacy paradox in their research of information disclosure to a chatbot in relation to their privacy concerns. When the participants were asked after the experiment why they revealed this information to the chatbot, they disclosed that they “felt personally addressed” by chatbot Luci and liked her type of “soft communication” (Spiekermann, Grossklags & Berendt, 2001, p. 45). This aspect of feeling personally addressed and this type of soft communication sets the medium of the chatbot apart from the questionnaire, and might be an explanation of why users disclose information differently to a chatbot than to a simple questionnaire. Therefore, we hypothesize that:

H2b: The context of the chatbot moderates the influence of privacy concerns on information disclosure, in such a way that participants share more information in this condition despite their privacy concerns.

Summary of hypotheses

H1: Higher privacy concerns lead to a lower disclosure of personal information H2a: The online privacy context has a moderating impact on the relationship between

privacy concerns and information disclosure

H2b: The context of the chatbot moderates the influence of privacy concerns on

information disclosure, in such a way that participants share more information in this condition despite their privacy concerns

(20)

Conceptual model

Our hypotheses are visualized in a conceptual model (figure 1) to give an overview of the variables and relationships in this study.

Figure 1: Conceptual model

Privacy concerns Sensitive information disclosure Privacy context (chatbot vs. simple online survey)

(21)

Method

We set up an online experiment to test individuals' sensitive information disclosure in response to either an online chatbot or to a simple online questionnaire and contrasted this with the individuals’ reported privacy concerns. The experiment was set up in cooperation with the Dutch company Digital Shapers, and the first pre-test experiment was conducted during an event of Digital Shapers on the 21st of March 2018 about Artificial Intelligence. We implemented a between-subjects experimental design (assigning participants to either the context of an interactive chatbot or the context of a simple online questionnaire) and assigned participants to one of the two conditions using the randomizer option in the Qualtrics survey flow. The dependent variable that we measured during this experiment was disclosure of sensitive information.

Measures

Privacy concerns

“Privacy concerns” is the independent variable of this study which we define as “individuals’ beliefs about the risks and potential negative consequences associated with sharing information” in line with the definition implemented by Baruh et al. (2017, p. 27). We determined participants’ privacy concerns by their score on the four survey items based on the privacy concern items implemented by Dienlin & Metzger (2016) in their study on the privacy calculus model. The items that we used in this study based on those implemented by Dienlin & Metzger (2016) are:

PRI.CON_1 Compared to other topics, privacy is not very important to me.* PRI.CON_2 I do not feel especially concerned about my privacy online.*

(22)

PRI.CON_3 I worry about my privacy with regards to my online behaviour.

PRI.CON_4 People's concern about privacy of using social media and buying things online has been overblown.*

The items with an asterisk were reversely coded. Participants rated the items on a 7-point Likert scale with 1 = strongly disagree, and 7 = strongly agree.

According to Dienlin & Metzger, these items on Privacy Concerns show sufficient internal consistency with a Cronbach’s alpha of .7 (2016). Likewise, the items in our study showed good internal reliability with a Cronbach’s alpha of .75 for 192

participants. Although the literature on using Cronbach’s alpha for Likert scales is somewhat divided, Cronbach’s alpha is the most widely used and implemented coefficient for measuring internal reliability for this type of scale (Gliem & Gliem, 2003). Sullivan & Atino also recommend implementing Cronbach’s Alpha to measure reliability for Likert Scales in their widely cited article “Analyzing and Interpreting Data From Likert-Type Scales” (2013). Furthermore, Robert Peterson found in his meta-analysis on Cronbach’s Alpha that the coefficient was not significantly influenced by different research designs, taking into account the scale type, format, and nature (1994).

Privacy context

In this experiment, the moderator variable privacy context is manipulated to be either the context of an interactive, personalized chatbot or a simple online questionnaire. In this study, we define a chatbot as “an artificial construct that is designed to converse with human beings using natural language as input and output,” in line with the definition used by Jia (2009, p. 250).

(23)

The chatbot for this study was developed on the platform of Landbot.io and was given a personalized look and response to match Spiekermann, Grossklags & Berendt’s (2001) chatbot Luci. In their study, participants argued that they “felt personally addressed” by chatbot Luci and liked this type of “soft communication” (Spiekermann, Grossklags & Berendt, 2001, p. 45), which is why the participants disclosed an extensive amount of information in response to the chatbot. In line with Spiekermann, Grossklags & Berendt (2001), we personalized our chatbot by giving her a name (Lucy) accompanied by a female avatar. See figure 2 for the look of our personalized chatbot named Lucy. Furthermore, the chatbot was programmed to feel more personal by responding to the answers the participants gave. Instead of just repeatedly stating different questions such as in a survey, the chatbot responded to the answers of the participants with short comments such as “Thanks for your answer” or “All right” or “We are almost there!”

(24)

The second condition, the simple online questionnaire, was developed on the online platform of Typeform.com. The platform of Typeform.com was chosen because many companies, among which the company Digital Shapers, nowadays build their survey on this platform. In contrast to the chatbot condition, this

questionnaire was not given a personalized feel, and simply stated the questions that the respondents were asked to fill in, and did not respond to the answers of the participants.

The Typeform starts with giving the instructions that the participants are asked to answer the following questions or type “0” if they do not want to answer a

question, similar to the chatbot. See figure 3 for the look of the questions that were asked in the Typeform.

(25)

Personal information disclosure

We define personal information disclosure as “the total number (sum) of personal information items subjects supplied” in line with Norberg, Horne & Horne in their study on the privacy paradox (2007, p. 110). To measure personal information disclosure, we defined a set of questions that included sensitive items.

We then measured whether the users answered the question or typed “0” because they did not feel comfortable answering the question. The questions were developed to fit the setting of artificial intelligence since the participants thought they were answering questions about their knowledge of artificial intelligence. Please see figure 4 for the questions we included in our pre-test experiment on the 21st of March.

(26)

Figure 4: Questions included in the pre-test experiment in both conditions

The starred items were the sensitive questions, and the response to these questions was measured. For the first three starred items, it was measured whether the participant responded to the question or answered “0”, and for the final question it was measured whether the respondent answered “yes” or “no.”

The sensitive questions were developed based on the experiment by Preibusch, Krol & Beresford (2013): “The Privacy Economics of Voluntary Over-disclosure in Web Forms.” Preibusch, Krol & Beresford (2013) argue that “date of birth is one of

1 = What is your name?

2 = What is your date of birth? *

3 = How would you rate your knowledge on Artificial Intelligence on a scale from 1 (no knowledge) to 5 (expert)?

4 = What do you find most fascinating about AI?

5 = Do you have any previous experience with chatbots?

6 = Do you have a budget for implementing AI tools and how high is this

approximately?*

7 = In which area of your work would you like to implement artificial intelligence tools such as chatbots?

8 = In which area of society do you think that AI can be most beneficial? 9 = Do you have any tips for Digital Shapers' implementation of chatbots? 10 = Do you have any other feedback?

11 = On which phone number can we reach you?*

12 = Can we share this information with other interested parties?*

(27)

the data items that consumers are least willing to provide” (p. 191), which is why we included the question “What is your date of birth?” in our question set. Furthermore, the authors argue, “Web users themselves are regularly reluctant to share financial information” (Prebusch, Krol & Beresford, 2013, p. 191), which is why we included the question: “Do you have a budget for implementing AI tools and how high is this approximately?” Adding on, research by Demmers, Zerres & van Dolen used the willingness to share a phone number in their research on the privacy paradox (2018), which is why we included the personal question “What is your phone number?”

Finally, the question “Can we share this information with other interested parties?” was included as a sensitive question based on the studies by Acquisiti & Grossklags (2004) and Huberman, Adar & Fine, (2005) that investigate the willingness of customers to share their information with external organizations.

For our online experiment, some questions were slightly adapted to fit the general audience instead of the members of the audience of the Digital Shapers event, which were mostly decision-makers of Dutch companies in the field of Digital

Business. Please see figure 5 for the questions that were asked by both the chatbot Lucy and the Typeform in our online experiment.

(28)

Figure 5: Questions included in the online experiment in the chatbot and Typeform condition

1 = What is your name?

2 = What is your date of birth?*

3 = How would you rate your knowledge on Artificial Intelligence (AI) on a scale from 1 (no knowledge) to 5 (expert)?

4 = Do you think that society will become better or worse from increased automation and AI?

5 = How long do you think it will be before AI has a noticeable impact on your daily life? (e.g. one day, 1 year, 10 years)

6 = Artificial Intelligence can help enabling scientific breakthroughs in medical science by analyzing enormous amounts of data with incredible speed. Which health

issues that you personally experience do you wish could be cured by the developments in AI?*

7 = Would you trust an autonomous car to drive your family if accident rates were better than human drivers?

8 = What is the biggest amount of money that you have ever spent on new technology

such as mobile phones, computers or smart watches?*

9 = In which area of society do you think that AI can be most beneficial? (E.g. education, healthcare, surveillance, etc.)

10 = On which phone number can we reach you?*

11 = Can we share this information with other interested parties?*

(29)

Looking at figure 4 and 5, one can see that the question on financial

information was altered from “What is your budget for AI?” to “What is the biggest amount of money that you have ever spent on new technology such as mobile phones, computers or smart watches?” seeing that the general audience usually does not have a company budget for AI. Secondly, a question on health information was added, since health information is considered as particularly sensitive (Preibusch, Krol & Beresford, 2013, p. 192) and adding this particularly sensitive item could help in seeing the difference in responses to a chatbot and a simple questionnaire. The

masking questions about artificial intelligence in study 2 are based on the questions in the Global Artificial Intelligence Survey from ARM (2017), a global consumer survey on awareness, acceptance, and anticipation of AI.

The conceptual model including the different measures discussed in the section above is outlined in figure 6.

Figure 6: Conceptual model including constructs at measurement level

Privacy concern score on items by Dienlin & Metzger

Number of sensitive items disclosed Conversing with a

chatbot vs. filling in the Typeform

(30)

Procedure

Pre-test experiment

Our experiment was first pre-tested during an event of the Dutch company Digital Shapers about Artificial Intelligence on the 21st of March with different decision-makers of Dutch companies and other clients of Digital Shapers present (N= 22). One week prior to the event, the organizers of the event sent out an online survey created on the platform Typeform.com. In the survey for the attendees of the Digital Shapers event, we included the four items measuring privacy concerns developed based on Dienlin & Metzger (2016) that are listed above. Respondents were asked to answer four items about their privacy concerns a 5-point scale ranging from 1 = strongly disagree to 5 = strongly agree before the event. Three of the items were reversely coded. The privacy concern items were mixed with other items about artificial intelligence and the talks of the other speakers at the event, so it was not obvious that the survey was measuring the participants’ privacy concerns.

During the event, I presented a session about Artificial Intelligence during which half of the participants received questions from a chatbot developed on

Landbot.io asking them questions about their knowledge of artificial intelligence and chatbots. The other half of the participants received the same questions but then in an online form created on Typeform.com. The participants received these questions through a link in their e-mail inbox that was sent out just minutes before I started my session during the Digital Shapers event. The participants got the option to either answer the question or respond “0” if they were not comfortable with answering the question. After the participants completed the questions, I disclosed the actual purpose of the questionnaire, which was not meant to test knowledge on artificial intelligence but was meant to test their information disclosure on sensitive items.

(31)

The session functioned as a pre-test of our experiment, and it was examined whether the experiment was correctly understood and if the questions had to be adapted. Overall, the experiment was well received, and the audience was not aware of the actual purpose of the experiment. Privacy concerns correlated negatively with information disclosure in the online survey condition (-.299), while this did not occur in the chatbot condition (.072), as was expected. As has been described above, some of the questions were adapted to fit the general audience better.

Online experiment

The second study was conducted in an online environment developed on survey software Qualtrics, chatbot platform Landbot.io and online questionnaire platform Typeform.com. The online study was completed by 192 participants (60,8% female), and participants were recruited using different social media platforms such as SurveyCircle, SurveySwap, but also through personal social media accounts. This method of sampling led to a mix of respondents known to the researcher and respondents unknown to the researcher. A convenience sample was employed, but this did enable the researcher to see the impact of proximity to the researcher on the sharing of sensitive information.

The participants first received a questionnaire on Qualtrics about social media usage to conceal that we were actually measuring their privacy concerns. To mask the privacy concern items (Dienlin & Metzger, 2016), we used the items of the Bergen Social Media Addiction Scale, implemented by Andreassen, Pallesen & Griffiths (2017). These items are based on the Bergen Facebook Addiction Scale (Andreassen et al., 2012) and include “I spend a lot of time thinking about Facebook or planning

(32)

how to use it” and “I use social media at least once every hour.” Participants rated the items on a 7-point Likert scale with 1 = strongly disagree and 7 = strongly agree.

After the participants filled in the first part of the survey on social media addiction with the masked items on privacy concerns, they either received a link to chat with the chatbot developed on Landbot.io or to fill in the online questionnaire developed on Typeform.com. The topic of the questions of the Chatbot or Typeform was participants’ knowledge and opinion of artificial intelligence so that the

information disclosure items were masked. The Qualtrics randomizer option was used to either show the participants the link to the survey in Typeform or the link to the chatbot on Landbot.io. After the participants answered all the questions, they were thanked for their time and participation and were informed about the actual purpose of the experiment.

Here it is noteworthy to refer to Preibusch, Krol, & Beresford (2013), in mentioning that to the casual observer, this might seem like a survey design. However, this study employs an experimental design, because instead of asking participants for their intentions to disclose, we asked for those very data items. The participant was in an actual disclosing situation and could decide him or herself to share this information. Hereby, actual behaviour is measured instead of behavioural intentions, which is according to Belanger & Xu (2015) precisely what the field of research on privacy concerns and information disclosure needs: “Given that users often behave not necessarily in rational ways when it concerns their information privacy behaviours, IS researchers need to focus on actual behaviours as opposed to merely individual intentions to behave when studying information privacy (Belanger & Xu, 2015, p. 575)”.

(33)

stating that it is a problem that a vast majority of the studies in the field of the privacy paradox use self-reported measures of behaviour (2017, p. 46). Therefore, this study will form a valuable addition to the area of research on the privacy paradox by

investigating the phenomenon experimentally instead of using self-reported measures of behaviour.

Control variables

The following variables were included as control variables in this study: knowledge of artificial intelligence, gender, age and social proximity to the researcher. First of all, when participants have more understanding of artificial intelligence, this might influence how they respond and disclose information towards the chatbot. In the meta-analysis by Baruh et al. (2017) it is also mentioned that higher online literacy and digital awareness increase online privacy concerns and privacy-protective behaviours. Furthermore, the background characteristics of age and gender were included to test if these demographic characteristics did not influence the results. Finally, the control variable of proximity to the researcher was added. The proximity or closeness between two persons can have significant influences on self-disclosure, as described by Rubin & Shenker (1978), Laurencau, Barret &

Pietromonaco (1998) and Ledbetter et al. (2011). Proximity to the researcher was measured using Bogardus’ social distance scale (Bogardus, 1933), a widely adopted and validated scale (Wark & Galliher, 2007) that ranges from 1 = no proximity at all to 6 = as close as close relatives/family. Please see figure 7 for a visualization of our conceptual model including the control variables described above.

(34)

Figure 7: Conceptual model including control variables Privacy concerns Sensitive information disclosure Privacy context (chatbot vs. simple online survey) Knowledge of AI Gender Age Proximity

(35)

Results

This chapter outlines the analyses we performed to test if the hypotheses we developed based on the existing literature were supported or not. First of all,

preliminary analyses and assumption checks were performed, and descriptive statistics were obtained before the actual testing could be conducted. Secondly, the direct hypothesized effect was tested using a one-way between-subjects analysis of variance. After this, the interaction effect and the effect of the other variables on the dependent variable were tested using a factorial analysis of variance. Subsequently, we performed a hierarchical multiple regression analysis to add to the robustness of the factorial analysis of variance.

Preparing data for analysis

Before analysing the data, the data was checked for errors and outliers. Furthermore, three of the four items for privacy concerns had to be reversed because they were negatively phrased. After reversing these items, the variable Total Privacy Concerns was computed from the four items measuring privacy concerns, and the mean of the total scale was calculated. As described in the methods section, the scale Total Privacy Concerns shows a reliable internal consistency with a Cronbach’s alpha of .75 for N = 192.

Secondly, the open questions of information disclosure items were coded, and the different proportions of participants that provided each sensitive data item were calculated and are displayed below in figure 8 according to the sensitivity of the items. The sensitive item “phone number” was shared the least of the five sensitive items that were implemented in the experiment. Only 28.2% of all participants shared their phone number in the experiment, compared to the 31.9% of participants that said

(36)

personal health issue. Furthermore, 63.9% of participants shared their full date of birth, while the largest percentage of participants shared the maximum amount of money they had spent on technology: 94.7% of participants have shared this piece of personal information in the experiment. In figure 8, we can see that all the sensitive items were shared more frequently in the Typeform condition, except for “ShareInfo” which indicates that more participants in the chatbot condition agreed with their information being shared with other interested parties. Adding on, we can see that the biggest differences between conditions are present for the item DateofBirth (90,8% of participants in the Typeform condition shared this, while only 66% in the chatbot condition shared their date of birth) and for PhoneNo. (35.6% in the Typeform shared their phone number compared to just 19.6% in the Chatbot condition).

Figure 8: Proportion of respondents that provided each sensitive item, broken down by the condition

(37)

As mentioned in the methods section, we define personal information disclosure as “the total number (sum) of personal information items subjects

supplied” in line with Norberg, Horne & Horne in their study on the privacy paradox, which is one of the most influential papers in this area of research (2007, p. 110). In line with Norberg, Horne & Horne (2007), we coded sensitive items that participants disclosed with a 1 and sensitive items that they did not provide with a 0. During the coding of the information provided for the sensitive items, the values were checked for plausibility (in line with Preibusch et al., (2013) and Norberg, Horne & Horne (2007)), and obviously wrong stated answers were coded as 0 for information

disclosure (for instance, one participant in our experiment provided the phone number 06123456789). After coding the five sensitive items, we computed the variable Total Sensitive Items Disclosed.

Assumption checks

The normality of distribution was tested for both Total Privacy Concerns as well as Total Information Disclosure. The variables of Total Privacy Concerns (skewness = -.379 and kurtosis = -.333) and Total Information Disclosure (skewness = .049 and kurtosis = -.476) were sufficiently normally distributed, as shown by the scores on kurtosis and skewness that fall within the rule of thumb of between -1 and 1 for normally distributed variables (Bulmer, 1979).

A Chi-square test of goodness of fit was executed to test whether there was a difference between the number of participants in the chatbot and the online survey condition. There were no significant differences between the number of participants in the condition of the chatbot and in the condition of the Typeform, X2 (1, N = 191) = .521, p = .47. Furthermore, chi-square tests for independence were executed to test

(38)

whether there were any significant differences in background of the participants in the two conditions.

In the Chi-square test of independence conducted for the age of the participants, it turned out that two cells had an expected count of less than five because the age group 4 (65 years and older) only consisted of four participants that were all in the condition of the Typeform. Therefore, these four participants were excluded from further analysis. After eliminating the four participants that were older than 65 from the Chi-square test of independence and from further analyses, the test indicated that there were no significant differences in age between the two

experimental groups: X2 (2, N = 156) = .771, p = .68. Adding on, the Chi-square test for independence for gender showed that there were no significant differences in gender between the two groups: X2 (1, N = 173) = .771, p = .671. Similarly, the Chi-square test for independence indicated that there were no significant differences for proximity to the researcher between the two groups: X2 (5, N =173) = 7.67, p = .175.

Manipulation check

A manipulation check was executed to test if the experimental manipulation between the chatbot condition and the online survey condition succeeded. An

independent samples t-test was performed to measure whether participants disclosed more information in the chatbot condition compared to the Typeform condition. As opposed to what was expected, participants (N = 186) actually disclosed more information in the Typeform condition: the mean score of information disclosure for the participants in the chatbot condition (N = 98) was 2.40 (SD = 1.02), while participants in the online questionnaire condition (N = 86) on average reported 2.90 data items (SD = 1.07). This a mean difference between the conditions is .51, which is

(39)

significant (p < 0.01). For this test, Levene’s Test for Equality of Variances proved to be non-significant, t (186) = 3,28, p = .157, and therefore equal variances were assumed.

Descriptive statistics

A correlation matrix is provided below to give an overview of the Pearson correlation coefficients between the different variables and the mean score and standard deviation of each variable. As we can see in Table 1, there is no significant correlation found between Privacy Concerns and Information Disclosure, the

variables described in hypothesis 1 (r = -.125, p > 0.05). With regards to hypothesis 2, one can see in Table 1 that Information Disclosure correlates significantly with

Privacy Context (r = -.236 p > 0.01), in such a way that participants in the online Typeform condition actually disclose more information than those in the chatbot condition.

Looking at the control variables, one can see that Proximity of the participants to the researcher significantly positively correlates with Information Disclosure (r = .232, p <0.01). Proximity to the researcher also correlates significantly with Privacy Context, indicating that those that filled out the Typeform were slightly closer to the researcher (r= -.185, p <0.05). Furthermore, the level of Knowledge of AI correlates positively with Privacy Context, indicating that the reported knowledge of AI among participants was higher in the chatbot condition (r = .175, p <0.05).

Among the control variables were also significant relationships present, namely the variable Gender and Knowledge of AI had a significant positive correlation (r = .248, p <0.01), indicating that males reported a higher level of knowledge of AI. Adding on, the variable Proximity to the researcher had a

(40)

significant correlation with Gender (r = -.152, p <0.05) and Age (r = .213, p <0.01), indicating that participants who were female and who were slightly older were closer to the researcher. Furthermore, Proximity correlated negatively with Knowledge of AI (r = -.182, p <0.05), indicating that those who reported higher knowledge of AI were, in general, less close to the researcher.

Table 1: Means, Standard Deviations, and Correlations

Means, Standard Deviations, Correlations

Variables M SD 1 2 3 4 5 6 7 1. Privacy Concerns 4.82 1.20 (.75) 2. ItemsDisclosed 2.64 1.08 -.125 - 3. Privacy Context 0.54 0.50 -.024 -.236** - 4. Knowledge of AI 2.44 0.96 -.026 -.057 .175* - 5. Gender 0.33 0.47 -.091 -.025 -.032 .247** - 6. Age 1.40 0.69 .130 .139 -.060 -.197* .013 - 7. Proximity 3.12 1.48 .020 .252** -.185* -.182* -.152* .213** -

** Correlation is significant at the 0.01 level (2-tailed) * Correlation is significant at the 0.05 level (2-tailed)

Based on the Chi-Square tests and the correlations table, we can see that the control variables Knowledge of Artificial Intelligence, Gender and Age do not have a significant impact on either the dependent or independent variable. Proximity,

however, does significantly correlate with the total number of items disclosed and therefore will be taken into account in the further testing of the hypotheses (please see figure 9).

(41)

Figure 9: Conceptual model including control variables taken into account in further analyses

One-way ANOVA

The variable of privacy concerns was recoded to be able to use the continuous independent variable as a categorical variable fitted for the ANOVA analyses. The mean score of privacy concerns was computed from four different items, and the participants’ scores on privacy concerns fell between 2.0 and 7.0 with intervals of 0.25. The variable “mean privacy concerns” was recoded so that 0.25 was rounded off to 0.5 and 0.75 was rounded off to 1.0. The recoding resulted in eleven categorical groups with privacy concerns between 2 and 7 with intervals of 0.5 (please see table 2).

A one-way analysis of variance test was conducted to test the effect of privacy concerns on information disclosure. Levene’s Test for Homogeneity of Variances proved to be non-significant, F (10, 172) = .986, p = .457, and therefore equal

Privacy concerns Sensitive information disclosure Privacy context (chatbot vs. simple online survey) Controlled for and checked: Knowledge of AI Gender Age Proximity

(42)

variances were assumed. Participants were divided into groups according to their privacy concerns scores as described above. There were no statistically significant differences between groups in information disclosure as displayed in table 2: F (10, 172) = 1.742, p = .075, which contradicts the first hypothesis of this study. Figure 10 shows that the participants with the highest privacy concerns did disclose less

information than the moderately and lowly concerned participants. However, this effect was not significant, and therefore our first hypothesis is rejected. Because the one-way analysis of variance shows that privacy concerns do not have a significant effect on information disclosure, the results we found are in line with the privacy paradox (Norberg, Horne & Horne, 2007).

Table 2: Total information disclosure for levels of privacy concerns

Privacy

concerns N

Mean items

disclosed Std. Deviation Std. Error

2.00 5 2.40 .894 .400 2.50 2 3.00 1.414 1.000 3.00 11 2.55 .688 .207 3.50 16 2.63 .957 .239 4.00 15 3.20 1.014 .262 4.50 31 2.68 .979 .176 5.00 21 2.67 1.155 .252 5.50 35 2.97 1.070 .181 6.00 20 2.40 1.095 .245 6.50 18 2.17 1.098 .259 7.00 9 1.89 1.452 .484 Total 183 2.64 1.080 .080

(43)

Figure 10: Information disclosure for different levels of privacy concerns

To test whether the privacy paradox was present for both the participants in the chatbot condition and the online survey condition, we run a second one-way analysis of variance using the SPSS Split file option splitting the data on privacy context. In both the chatbot condition as well as in the Typeform condition no significant differences were found between the levels of privacy concerns and the disclosure of information (F (9, 76) = .875, p = .551 in the online survey condition and F (10, 86) = 1.818, p = .069 in the chatbot condition) indicating that the privacy paradox was apparent in both conditions. Levene’s test of homogeneity of variances was insignificant for both conditions (F (8, 76) = 1.020, p = .429 in the online survey

(44)

condition and F (10, 86) = .885, p = .551 in the chatbot condition) and therefore equal variances were assumed. These results provide evidence for the fact that the privacy paradox is apparent in both the chatbot and the Typeform condition.

Factorial ANOVA

A factorial analysis of variance was conducted to compute the main effects of privacy concerns and privacy context on information disclosure when controlling for proximity (please see table 3). Furthermore, this analysis allowed us to test the interaction effect between privacy concerns and privacy context on the dependent variable information disclosure when controlling for proximity. Levene’s Test for Equality of Error Variances proved to be non-significant, F (5, 164) = 2.171, p = .060, and therefore equal variances were assumed. The interaction effect between privacy concerns and privacy context was not statistically significant, F (2, 163) = 1.555, p = .223, partial eta squared .018. This evidence rejects hypothesis 2. In other words, privacy context does not seem to moderate the relationship between privacy concerns and information disclosure. However, privacy context did have a significant main effect on information disclosure F (1, 163) = 4.678, p = .034. The partial eta squared of this effect was .027, indicating that it was a small effect. Furthermore, the control variable proximity also had a significant effect on information disclosure, F (1, 163) = 8.21, p < 0.01. Again, this was a small effect because of the partial eta squared of .048.

(45)

Table 3: Factorial ANCOVA Dependent variable: information disclosure

Sum of

Squares df Mean Square F

Partial Eta Squared Sig. Proximity 8.428 1 8.428 8.207** .048 .005 Privacy concerns 1.543 2 .772 .751 .009 .473 Privacy context 4.678 1 4.678 4.555* .027 .034 PrivCon * PriContx 3.109 2 1.555 1.514 .018 .223 Error 167.382 163 1.027 Total 1434.000 170

** Significant at the p < 0.01 level * Significant at the p < 0.05 level

When computing a factorial ANOVA without controlling for Proximity, we can see that Privacy Context has a larger and more statistically significant effect on Information Disclosure (F (1, 178) = 8.588, p = .004 and the partial eta squared being .046 compared to F (1, 163) = 4.678, p = .034 and the partial eta squared being .027 when controlling for Proximity). It thus seems like Proximity explains a large part of the effect of Privacy Context on Information Disclosure. However, when controlling for Proximity, the effect of Privacy Context on Information Disclosure is smaller but still significant. This finding will be further analyzed in the discussion section.

Hierarchical multiple regression

To add to the robustness of the factorial ANOVA, a hierarchical multiple regression analysis was performed to investigate the predictive power of privacy concerns on information disclosure after controlling for proximity, age, gender and

(46)

In the first step of the hierarchical multiple regression, the control variables Proximity, Age, Gender and Knowledge of Artificial Intelligence were entered as predictors. This model was statistically significant, F (4, 144) =3.81; p < 0.01 and explained 9.6% of the variance in information disclosure. After entry of Total Privacy Concerns the total variance in Information Disclosure that our model explained was 10.7%; F (1, 143) = 1.872; p = .104, indicating that the second model was not statistically significant. Adding Total Privacy Concerns to the model only explained an additional 1.2% in variance of Information Disclosure after controlling for proximity, age, gender and knowledge of artificial intelligence (R2 change = .012; F (1, 143) = 1.87; p = .173), and was no significant predictor of information disclosure (beta value = -.110, p = .173). These results indicate again that hypothesis 1 has to be rejected since the variable Privacy Concerns was no significant predictor of

information disclosure.

The only variable that was significant in the second model was Proximity, with a beta value of .271 (p > 0.01). This result indicates that for 1 level increase in Proximity, participants’ information disclosure increases by .271. The third model that we analysed in the regression analysis included the moderator variable Privacy Context as a predictor. This model was statistically significant, F (1, 142) = 5.15, p = .025 and explained 14% of the variance in information disclosure. Adding Privacy Context to the model explained for an additional 3.1% in variance of Information Disclosure after controlling for Proximity, Age, Gender and Knowledge of Artificial Intelligence (R2 change = .039; F (1, 142) = 5.15; p = .025). In the third model, Privacy Context was significant next to Proximity, with a beta value of -.183 (p = .025). This result indicates that if the privacy context increases for one (thus to the chatbot condition), a person’s information disclosure decreases by -.183.

(47)

Table 4: Hierarchical multiple regression analysis Model Unstandardized Coefficients Standardized Coefficients t Collinearity Statistics R R2 R2 Change B Std. Error Beta

Step 1 .309 .096 .096** Proximity .532 .161 .273** 3.308 Gender -.177 .504 -.029 -.351 Age .411 .345 .098 1.191 Knowledge AI .081 .251 .027 .322 Step 2 .328 .107 .012 Proximity .528 .160 .271** 3.289 Gender -.245 .505 -.040 -.485 Age .475 .347 .114 1.369 Knowledge AI .088 .251 .030 .352 Total privacy concerns -066 .048 -.110 -1.368 Step 3 .372 .139 .031* Proximity .466 .160 .240** 2.907 Gender -.361 .501 -.059 -.721 Age .487 .342 .117 1.423 KnowleAI .182 .251 .061 .728 Total privacy concerns -.069 .048 -.115 -1.452 1 = Chatbot and 0 = Typeform -1.052 .463 -.183* -2.270

** Significant at the p<0.01 level * Significant at the p<0.05 level

(48)

Discussion

This study aimed to investigate the effect of the privacy context (either conversing with an online chatbot or filling in a simple online questionnaire) on the relationship between the users’ privacy concerns and their online sensitive

information disclosure. First of all, the implications regarding the direct hypothesized effect between privacy concerns and information disclosure will be discussed.

Secondly, the moderation effect of privacy context on the relationship between privacy concerns and information disclosure will be evaluated. Thirdly, we will consider the contributions to theory and practice of this study. Finally, we will address potential limitations to this study and identify avenues for further research.

Implications regarding the privacy paradox

After reviewing the current research on privacy concerns in relation to information disclosure, we hypothesized that higher privacy concerns lead to lower information disclosure, in line with the meta-analysis of 166 studies on this topic by Baruh, Secinti & Cemalcilar (2017). However, in our study, it turned out that higher privacy concerns actually did not lead to lower information disclosure in the chatbot or the Typeform condition. This evidence points to the fact that the privacy paradox is present in both of these conditions. The privacy paradox entails that individuals tend to “forget” their privacy concerns once in an actual disclosing situation (Norberg, Horne, & Horne, 2007). Because the body of research on the topic of the privacy paradox is divided with multiple researchers arguing against (Son & Kim, 2008; Lutz & Strathoff, 2014; Krasnova, Spiekermann, Koroleva & Hildebrand, 2010; Baruh, Secinti & Cemalcilar, 2017) and in favour (Taddei & Contena, 2013; Norberg, Horne,

(49)

scientific addition to the discussion in this field to provide further evidence on whether the privacy paradox is existent or not.

Adding on, the fact that we implemented an experimental study that measures actual behaviour instead of measuring participants’ intentions makes this study a valid contribution to the current body of research on privacy concerns and information disclosure. As mentioned by Baruh, Secinti & Cemalcilar (2017) and Belanger & Xu (2015) it is a problem that the majority of the studies in the field of the privacy paradox use self-reported measures of behaviour. This might also point to why the results of our study differ from the results of the meta-analysis Baruh, Secinti & Cemalcilar (2017). In the discussion of their meta-analysis, the authors point out that most of the studies in their meta-analysis used self-reported measures of behaviour, and those might overestimate how much users actually act according to their privacy concerns (Baruh, Secinti & Cemalcilar, 2017). Therefore, this study forms a

substantial contribution to the privacy paradox debate by experimentally investigating actual behaviour, and this research points out that users do not behave according to the concerns they report. Users readily disclose sensitive personal information to both the chatbot and the online questionnaire despite their privacy concerns.

(50)

Implications regarding the privacy context

We hypothesized according to the literature on information disclosure and privacy concerns on different online platforms that the privacy context of either conversing with an interactive chatbot or filling in a simple online questionnaire would have a moderating impact on the relationship between privacy concerns and information disclosure (Palen & Dourish, 2003; Stutzmann, 2006; Krishnamurthy & Wills, 2008; Schrammel, Köffel, & Tscheligi, 2009; De Choudhury, Morris & White, 2014). Secondly, we hypothesized that the context of a chatbot would moderate the relationship in such a way that while chatting with a chatbot, the participants would actually share more information despite their privacy concerns (Spiekermann, Grossklags, & Berendt, 2001).

First of all, we did not find an interaction effect between the privacy context and the relationship between privacy concerns and information disclosure as

displayed by the factorial analysis of variance. We did find that the privacy paradox was present for the chatbot condition as presented in the one-way analysis of variance split for both the chatbot condition and the online survey condition, in line with the research by Spiekermann, Grossklags, and Berendt (2001). This finding indicates that privacy concerns in the chatbot condition did not lead to lower information disclosure by the participants, as discussed in the section above, and that the privacy paradox was apparent in the chatbot condition. Furthermore, instead of finding a moderating effect of the privacy context on the relationship between privacy concerns and

information disclosure, we found a direct effect of the privacy context on information disclosure that will be discussed below.

Spiekermann, Grossklags, and Berendt (2001) found in their research on information disclosure to an “anthropomorphic 3-D bot” that privacy concerns did not

Referenties

GERELATEERDE DOCUMENTEN

In this paper we examine how simulated social touch by a virtual agent in a cooperative or competitive augmented reality game influences the perceived trustworthiness, warmth

The results of the multiple regression analyses showed relationships between burnout, strain and job characteristics, such as work overload, role ambiguity, task characteristics,

Younger participants’ negative beliefs and dislike of ITF determined their lower product acceptance and intended consumption of the less modernized dishes (samples

Development of novel anticancer agents for protein targets Estrada Ortiz, Natalia.. IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish

The political State diversity on the other hand shows significant results at some points, as firms located in Democrat States obtain more financial performance than firms located

Based on an analysis of related work, we dis- tinguish explicit and implicit concept drift, which in automatic subject indexing translates to settings with new descriptor terms and

Objective: This study aimed to conceptualize wearable technology embodiment, create an instrument to measure it, and test the predictive validity of the scale using