AI acceptance and attitudes : People’s perception of healthcare and commercial AI applications

(1)

Linköpings univeristy | Department of Computer Science Bachelor thesis, 18 ECTS | Cognitive Science SP 2021 | LIU-IDA/KOGVET-G--21/003--SE

Linköpings universitet SE-581 83 Linköping 013-28 10 00, www.liu.se

AI acceptance and attitudes

People’s perception of healthcare and commercial AI

applications

Josef Jönsson

Supervisor: Erkin Asutay Examiner: Mattias Forsblad

(2)

ii

Upphovsrätt

Detta dokument hålls tillgängligt på Internet – eller dess framtida ersättare – under 25 år från publiceringsdatum under förutsättning att inga extraordinära omständigheter uppstår. Tillgång till dokumentet innebär tillstånd för var och en att läsa, ladda ner, skriva ut enstaka kopior för enskilt bruk och att använda det oförändrat för ickekommersiell forskning och för undervisning. Överföring av upphovsrätten vid en senare tidpunkt kan inte upphäva detta tillstånd. All annan användning av dokumentet kräver upphovsmannens medgivande. För att garantera äktheten, säkerheten och tillgängligheten finns lösningar av teknisk och administrativ art. Upphovsmannens ideella rätt

innefattar rätt att bli nämnd som upphovsman i den omfattning som god sed kräver vid användning av dokumentet på ovan beskrivna sätt samt skydd mot att dokumentet ändras eller presenteras i sådan form eller i sådant sammanhang som är kränkande för upphovsmannenslitterära eller konstnärliga anseende eller egenart. För ytterligare information om Linköping University Electronic Press se förlagets hemsida http://www.ep.liu.se/.

Copyright

The publishers will keep this document online on the Internet – or its possible replacement – for a period of 25 years starting from the date of publication barring exceptional circumstances. The online availability of the document implies permanent permission for anyone to read, to download, or to print out single copies for his/hers own use and to use it unchanged for non-commercial research and educational purpose. Subsequent transfers of copyright cannot revoke this permission. All other uses of the document are conditional upon the consent of the copyright owner. The publisher has taken technical and administrative measures to assure authenticity, security and accessibility. According to intellectual property law the author has the right to be mentioned when his/her work is accessed as described above and to be protected against infringement. For additional information about the Linköping University Electronic Press and its procedures for publication and for assurance of document integrity, please refer to its www home page: http://www.ep.liu.se/.

(3)

iii

Abstract

The relevance of AI is ever increasing. The goal of the wide implementation is usually either to boost task efficiency or for public comfort. To fuel this progression, more personal data is being used and Artificial intelligence inhabits the role of the human expert, in many different applications. This study investigated the attitudes and rates of acceptance to said AI applications and if they differed in relation to each other. Additionally, this study explored if general positive and negative attitude towards AI influence AI acceptance. Applications studied came from two different domains,

E-commerce/Marketing and Healthcare. It was found that acceptance levels did in fact not significantly differ between the two domains. However, a significant positive correlation was found between positive attitude and acceptance rates, while an inverse significant correlation was found between negative attitude and acceptance rates.

Keywords

(4)

iv

Acknowledgments

I would like to give special thanks to my family and close friends for all the love and encouragement, through thick and thin. Furthermore, I would express thanks to my project partner Theodor. We have many times both agreed on that writing this project on our own would have been impossible. While everything else during Covid-19 has been an utter struggle, working together proved to be so enjoyable. Finally, I would like to thank my supervisor Erkin Asutay. Me and Theodor has greatly valued all the insight and support you have given us along this journey.

(5)

v

1 Introduction

Rose spends many hours a day listening to heavy metal music on Spotify. Should Spotify ever

recommend her a genre of music that is not heavy metal? Rob plays a lot of games on his PC. Should google search engine target Rob with personal advertisement, stating that video games are a possible cause for violent actions?

An algorithm is a set of instructions that helps to solve a problem or complete a task. The internet runs on such algorithms. Every type of online searching is done with the help of algorithms. Emails being sent always find their way to the right inbox because of algorithms. Every single social media app is nothing but a clever set up algorithmic environments. To play through a videogame is to play through algorithmic storytelling. A movie recommendation or a dating app is nothing else than algorithms. When google maps recommend the fastest path between Stockholm and Gothenburg, it is all done through smart algorithms. To construct an algorithm, the practitioner must adopt math theory and write computer code. It is these different recipes of code that make all online activities possible. Similarly, Artificial Intelligence (AI) operates on algorithms.

If the algorithm is the recipe, then the AI is the chef cooking with that recipe. With enough recipes, the AI can now cook a meal – perform a task. To become a proficient chef, the AI needs to be able to modify its recipes, to fit the taste pallet of the restaurant guest, and to improve its prospects of continuously cooking well. Thus, an AI needs to be able to modify its algorithms when needed or create new ones, as a response to learned inputs and collected data. This ability to change and act with “intelligence” is the reason why AI has such a broad use case today.

According to Gartner (2019), in the last four years, enterprises implementing AI have grown by 270 percent. As such, 37 percent of all organizations, in all industries, now utilize some kind of AI (Gartner, 2019 ). In the health sector, AI has proved to play a vital role in diagnostics, monitoring hospital patients, and in the analyzing of various diseases. Through AI’s ability

to analyze patterns in big data sets, autonomous vehicles now drive on the streets, automatically understanding their surroundings and how to drive in it safely. AI has brought drastic change to many fields, automating systems, improving efficiency, and refining performance. However, AI has a more sinister set of practices as well. AI tools are often used in a somewhat more manipulative way, such as personally targeted pollical ads or the exploitation of data, infringing on the privacy of vulnerable people.

Despite the widespread use of AI, their impact on society is not yet clear. For instance, how widely do people know about the various algorithms that affect its digital environment? Most of the algorithmic

(7)

2

decisions affecting people's lives are made by software companies. There are still many aspects of AI, and what effects it has on people, that are unclear. How accepting are people of the overall change to a more algorithmic-driven society?

1.1 Aims

Kozyreva et al. (2020) aimed to probe the German population in questions of AI, in their study Representative Survey of Public Attitudes in Germany (Kozyreva, Herzog, Lorenz-Spreen, Hertwig, & Lewandowsky, 2020). They claimed that very little is known about public awareness and attitudes regarding modern AI applications. Taking inspiration from Kozyreva et al., this study aims to further investigate people’s attitudes and acceptance levels towards AI, specifically constrained to the E-commerce/Marketing and Healthcare domain. The domains were chosen to represent two different AI use cases. Commercial-related applications are used in a less serious manner, every day, while healthcare applications most likely are used less often, with potentially much greater consequences. This study analyses the rates of acceptance in the two different domains while investigating the interaction of positive and negative attitudes to the acceptance rates. These are the two studied research questions:

How accepting are people regarding AI applications in the E-commerce/Marketing and Healthcare domains?

And

How does general positive and negative attitude towards AI influence AI acceptance in healthcare and E-commerce domains?

2 Theoretical Background

2.1 Algorithm Aversion

There currently exists an abundance of evidence supporting the claim that algorithms outperform humans in many different tasks. One such task is prognostications. Paul Meehl (1954) was able to show this already in 1954 by reviewing 20 forecast studies from various fields of research. Whether the algorithms created forecasts regarding academic performance or parole violations, the results showed that algorithms were superior to the human counterpart (Meehl, 1955). Similarly, following the work of Grove, Zald, Lebow, Snitz, and Nelson (2000), evidence was found in 136 studies

(8)

3

the predictions of humans. On average, the algorithms displayed 10% better performance in

comparison. Thusly, the large span of different tasks was overwhelmingly in favor of the accuracy of algorithm forecasts (Grove, Zald, Lebow, Snitz, & Nelson, 2000).

The fact that algorithms outperform human forecasting should realistically mean that they are the common choice for such a task. The reality is, they are not. Expert and laypeople tend to be skeptical to use algorithms in forecast purposes, choosing inferior human forecasts (Diab, Yueh Pui, &

Yankelevich, 2011). This phenomenon is referred to as “algorithm aversion”. But why does this occur? Burton et al. (2019) came up with five characteristic themes to further explain this (Burton, Stein, & Jensen, 2019 ).

The first theme deals with false expectations and algorithmic literacy. A human decision-maker tends to exist misguided in what they expect out of the algorithm when it is utilized to help in the decision-making. The person builds up certain expectations for what they think that algorithm can and should do, and how it actually functions. People tend to pick up such expectations through their own personal use of algorithmic aids or in experiencing decision-making recommendations. Similarly, friends and media also affect the user’s expectations. This creates an opinion paradigm shift over time, resulting in humans over-scrutinizing the credence of algorithm advice. So, what are the underlying mechanism of these expectations? According to Burton et al. (2019), this can be explained by a few different factors. One being the human tendency to associate human error with randomness, something that is

repairable, while algorithmic error is something systemic. Furthermore, when seeking advice humans tend to seek a bond with the source of the advice - a social or parasocial relationship – which is a distinctly different experience with algorithmic agents. Finally, the problem of expectations can also be linked to individual differences (e.g., professional experience or age), showing the significance of individual attachments and sentiment regarding algorithm aided decision-making (Burton, Stein, & Jensen, 2019 ).

The second theme is the lack of decision control. When a human is in the process of making a decision, the impression of being in control is an important factor to confidently make that decision. To achieve such control, the human agent could either have a very high level of understanding of the algorithmic process, or the algorithm designer could present the interface in a differently, indirectly creating user control. Furthermore, control in decision-making is connected to trust. To calculate trust in decision aid Muir (1987) points to predictability, dependability, technical competence, reciprocity, and morality (Muir, 1987). Such predictability and dependability are critical in how humans feel in control when interacting with algorithmic decision-making aids. Individuals tend to find it more burdensome to satisfy the psychological self-interests needed to feel in control, while being assisted by algorithmic aids (Colarelli & Thompson, 2008). Additionally, the human decision-makers trust in such

(9)

4

algorithms has been shown to rapidly degrade, should it make a flawed judgment. Burton et al. (2019) note that this fact is probably linked to human decision makers feeling the lack of agency and control. Theme number three discusses the Lack of incentivization problem. Many studies have shown that people favor being incentivized by a human expert above an algorithm (e.g., Alexander, Blinder, & Zak, 2018; Brown, 2015). Burton et al. (2019) argue that this can be explained through that a decision is made through many different judgments. Every judgment needs a certain level of incentive, which adds up being a significant amount of motivation. The creation of such an algorithmic aid is deemed difficult, for the reason that the success of algorithmic decision-making involves motivating and incentivizing human decision-makers (Burton, Stein, & Jensen, 2019 ).

The fourth theme is called combatting intuition. The complexity of algorithmic decision-making requires the simultaneous integration of two decision-making processes: those of the human and the algorithmic aid. These two processes need to be aligned and mapped so that they can be easily accessed and augmented. Burton et al. write that it is crucial that the two processes must work in parallel with each other, or the person might find himself needing to decide on which decision process to use. As such, the cognitive compatibility between the two processes becomes crucial for effective and low-effort decision-making. Thus, being able to form this link is of high importance for the algorithm aid, or it will rather hinder the human intuition (Burton, Stein, & Jensen, 2019 ).

The fifth and last theme discusses conflicting concepts of rationality. According to Burton et al., the “heuristics‐and‐biases program” is the basis for the present-day research of algorithm aversion. The program touches on the subject of human incapability of making rational calculations. Taking this perspective, much has been learned about algorithmic decision-making, through identifying human cognitive limits and inadequacies. But whilst this has complemented the creation and research of algorithmic aids, it has locked algorithmic decision-making research in a single line of research theory. Burton et al. points to “fast‐and‐frugal heuristics” as an additional line of research that would benefit in algorithmic decision-making, which deals with human decision uncertainties. The authors worry that numerous insights, about how humans actually make decisions, would be lost without a broadened theory scope. Especially in how humans make decisions while being uncertain, which is highly

connected to algorithm aversion. Here are the authors own words: “Put simply, human decision-makers often operate in a world of uncertainty (where alternatives, consequences, and probabilities are unknown and optimization is unfeasible) whereas algorithms operate in a world of risk (where

probabilities are known or calculable and optimization should be the objective; Hafenbrädl et al., 2016). The best decision strategy under risk is often not the best decision under uncertainty. So, when a human decision-maker or an algorithmic aid is unable to reconcile its view of what constitutes a good decision under the specific circumstances of a given task (i.e., the environment) with the other,

(10)

5

algorithm aversion is observed.” (Burton, Stein, & Jensen, 2019 , s. 226).

2.2 Acceptance

Acceptance is a person’s psychological willingness to approve or accept the currently experienced situation. The situation could be of a positive or negative variety, but to accept is to not act in an attempt to change it. The same description applies for the acceptance of an application, or of being aided in a decision. But what processes lead to accepting or rejecting new technology? To better understand the acceptance of AI technology, the next sections will explore three perspectives: Decision-aids, how acceptance can be comprehended in Kozyreva et al. (2020) and Technology Acceptance Model.

2.2.1 Decision-aids

A large portion of corporate and common people’s experience with AI technology is through the form of decision-aids (Doumpos & Grigorousdis, 2013). In AI decision-aids, a vast underlying set of data can assist the user through recommendations. For the benefit of the common people, AI decision-aid reduces the requirements of knowledge and mental processing of a task. Likewise, recommendation and projection systems greatly benefit the corporate realm. Through a high-quality set of information, companies can build robust frameworks for strategic operational decisions (AIMultiple, What is data-driven decision making? Step-by-step guide, 2019). To further explore these types of AI, Hilb, (2020) categorizes decision-aiding AI into five different scenarios:

Assisted intelligence

This type of assisting AI is generally accepted by society and the common man, being well regulated and accepted. The AI plays only a selective supporting role to the user, which is in full control of the decision-making.

One of the main objectives of an assisted intelligence AI is to improve on a process that is already in use, communicating with the decision-maker who then takes the final decision. For this reason, many different assisted intelligence applications are being used in health care, monitoring patients, and alerting doctors. Similarly, assisted intelligence is vital in the transportation industry, where much of the navigation is done by GPS.

Augmented intelligence

Similar to the assisted intelligence, the user has full control over the decision, being supported by AI. In this case, the AI technology outperforms the cognitive abilities of humans. Such tasks could be to

(11)

6

store and analyze a large amount of data, to later report automatically to the user. Augmented intelligence is undergoing a phase of regulation as of today, having more and more implications, which hinders it from being fully accepted. This type of AI is very politically debated and therefore has many implications for social acceptance.

Augmented intelligence is utilized in several industries, finance, healthcare, and cybersecurity to name a few. Storing and interpreting large data sets means that a finance company can make robust

projections, that health care companies can diagnose illnesses and cybersecurity companies can analyze new potential areas of cyber risk.

Amplified intelligence

Amplified intelligence means that decisions are done collaboratively, between the user and the AI. By utilizing the full potential of both man and machine, the recommendations and approvals of a decision can be sent back and forth in co-operation until a final decision is established.

Amplified intelligence is being used in highly technical situations where the human and the AI work as a system. The nuclear power plant operative safely operates facility through such a collaboration. Similarly, a jet plane cockpit utilizes the decision-making and action-taking from both the machine and the human operator.

Autonomous intelligence

Autonomous intelligence is capable of making decisions autonomously and without constant inputs. This allows machines to perform within a predefined range. Some examples of this are self-regulate control mechanisms and robots. As for the amplified intelligence, similar regulatory debates are going on with autonomous intelligence, making it lack accountability and liability in terms of acceptance. Autonomous intelligence often handles large amounts of data, providing online stores with automated customer analytics and projections. Additional use can be seen in political think tanks identifying undecided voters, factory automation tasks, and identification of stock market patterns (Boschert, Coughlin, & Ferraris, 2019).

Autopoietic intelligence

According to Hilb (2020), this type of AI does not exist today, outside that of science-fiction. Autopoietic intelligence means that the system can restructure and recreate its own elements. That means it would never need any human decision cooperation since the AI would be able to dynamically alter itself for whatever decision that needs to be analyzed. As a result, no relevant argument or discussion regarding acceptance can be made at this time.

As Burton, Stein, & Jensen, (2019) explains about algorithm aversion, for people to accept AI, many criteria has to be satisfied. All but the first of Hilb’s types of decision-aid AI fails to do so, which explains why the acceptance can vary so much from person to person. Governments giving technology

(12)

7

the green light, through regulation, heavily affecting people’s sense of control, expectations, and beliefs. But how else can AI acceptance be understood?

2.2.2 Acceptance in Kozyreva et al. (2020)

In Kozyreva et al. (2020), most respondents agree that it is acceptable to collect and use personal data to provide personalized entertainment and leisure services. Participants give out high acceptance ratings to the recommendations of music (77%), events (78%), and shops and restaurants (78%). On the other hand, they oppose the use of personal data in politics and news, giving worse ratings to political advertising (61%), customized social media feeds (57%), and customized search (e.g., Google) (63%).

Respondents considered information regarding age and gender to be somewhat acceptable (59% and 64% respectively). But when asked about personal tragedies, household income, religious views, political views, sexual orientation, or personality traits (e.g., “outgoing”), the respondents answered in opposing “not acceptable “ or “not very acceptable” rates (83%, 77%, 73%, 71%, 71%, and 62% respectively).

2.2.3 Technology Acceptance Model (TAM)

Technology acceptance model (TAM) was invented in 1989 by Fred. Davis. Davis argues that the model can be used to predict human acceptance levels for new technology. The model has been validated by many scholars (e.g., Chin & Todd, 1992; Agarwal & Karahanna, 2000).

(13)

8

Figure 1: Technology Acceptance Model (adapted from Davis, 1989)

The model states that a vast number of individual factors, like age, gender, and past experiences etc., become the external variables in the model, to affect people’s level of technology acceptance. Next in line, Davis explains that for an individual to experience acceptance, first two essential conditions must be fulfilled. As seen in Figure 1, Perceived usefulness and Perceived Ease of Use is crucial in assuring the user feeling a positive attitude towards the use of the technology. The user’s intentions of using the technology derive from his attitude. The last step is to actualize the use of the technology(Davis, 1989).

Using the reasoning of TAM, acceptance of technology becomes highly variable based on a wide range of individual factors. The main goal of the AI architect should be to make the AI feel useful while being easy to use. Different AI applications, in different domains, might need different

approaches. But acceptance is nevertheless a crucial factor to understand in how people interact with varied AI applications.

3 Methodology

3.1 Recruitment and Participants

On behalf of Jedi-lab research group, this survey was conducted from Linköping (Sweden). The study used a sample from The United States, containing N = 508participants (233 women, 265 men, and 10

(14)

9

non-specified), in the span of 18 to 79 years of age (mean age = 34.79, SD = 11.72). The survey had a within subjects design, measuring the result of the same subjects’ answers in the same questionnaire. Ultimately, 29 participants were excluded from this study, resulting in a final sample of N = 479. The reasons for being excluded was either to have finished the survey in under three minutes or to have responded with the same answer to all questions. The survey also contained a control question testing the participants’ attention and seriousness, which required them to correctly answer a simple question, or they would not be able to continue with the study.

The recruitment was done through prolific.co, an online platform for recruiting human participants. Prolific can guarantee their participants to be fluent English speakers, to be in a diverse age-span, and that close to half of the target participants are male, and the other half is female. Participants from prolific.co were compensated for taking part in this survey (Prolific.co, 2017).

3.2 Survey Design and Materials

To create, design, and distribute the web-based survey, the program Qualtrics was used. The software can be used anywhere in the world, granted the participants have access to a computer, tablet, or smartphone with internet-connection (LibGuides: Statistical & Qualitative Data Analysis Software: About Qualtrics, 2018). The survey was administered and sent out by the Jedi-lab research group that has robust previous experience of using the platform, for similar studies.

3.2.1 AI applications

The survey consisted of 20 questions and was divided into two different themes, called domains. 10 questions were on the subject of “Marketing/E-commerce” and the other 10 contained questions regarding “Healthcare”. The questions aimed to inquire about the participants’ level of perceived acceptance and risk/benefits in these two domains. All the applications can be viewed in Sect 4.2,

table 1. Perceived levels of acceptance were presented on a point scale (1-not acceptable at all,

5-Very Acceptable).

Inside the survey, general personal information is gathered about the participant, including age, gender, and tech-familiarity. Additionally, the survey incorporated a few additional questions from two previous proven study-questionnaires, “Initial validation of the general attitudes towards Artificial Intelligence Scale” (GAAIS) (Schepman & Rodway, Initial validation of the general attitudes towards Artificial Intelligence Scale, 2020) and “Technology Readiness Index” (TRI 2.0) (Parasuraman & Colby, 2015). TRI 2.0 was not used for any further analysis and was ultimately not part of this study. TRI proved to be superfluous, for planned analysis. A correlation analysis between the results of TRI

(15)

10

and acceptance rate was deemed too similar to the correlation analysis between GAAIS and acceptance rate.

In this study, healthcare is defined by how it interacts with artificial intelligence and its applications. It is used as an overarching term to describe every type of healthcare application that uses artificial intelligence or machine-learning algorithms to execute any healthcare tasks, typically done by a human. E-commerce/Marketing is similarly defined as an overarching term for AI applications within the areas of marketing and advertisements, in an online environment.

Healthcare and E-commerce/Marketing were chosen as domains because they highlight very different aspects of AI applications. In addition to investigating the general levels of AI acceptance, this study furthermore aims to investigate if one domain gets a different response than the other.

The domains were chosen to represent two different areas of AI applications. Commercial-related applications are used casually, every day, while healthcare applications most likely are used more rarely, with potentially much greater consequences. Additionally, applications were chosen with the goal of having the participants feeling roughly equally familiar and knowledgeable about the subject in the two different domains.

This study will be used as part of a larger future study. Here, the participants assessed acceptance, risk and benefits, attitudes and answered several other questions regarding new technology. It should be said that for this study, the focus was on investigating the level of “acceptance” and “attitudes” towards AI, only.

3.2.2 Scales

GAAIS was used to measure the participants’ general attitude to AI (Schepman & Rodway, 2020). The attitude scale has two subscales – positive and negative. Attitude rates were presented on a 5-point scale (1-Strongly disagree, 5-Strongly agree). We chose 4 items from each subscale to include in this study (item 1-4 come from positive subscales, item 5-8 came from negative subscales). The items chosen had the highest factor loading to each subscale and can be seen here (Schepman & Rodway, 2020):

Positive attitude scale:

• I am interested in using artificially intelligent systems in my daily life • There are many beneficial applications of Artificial intelligence • Artificial intelligence can provide new economic opportunities

(16)

11 • Artificial intelligence is exciting

Negative attitude scale:

• I think that Artificial intelligence is dangerous • Organizations use Artificial intelligence unethically

• I shiver with discomfort when I think about future uses of Artificial intelligence • Artificial intelligence is used to spy on people

From the eight original GAAIS questions, two categories were set up. The categories consist out of four positive attitudes and four negative ones. By calculating the mean scores from every application within one category, two different composite scores were created. These scored could be used as participant attitude composite scores. To support this conversion, a Cronbach test was done on both categories, resulting in a Cronbach α = 0.854 for the positive composite score and a Cronbach α = 0.765 for the negative category. According to George and Mallery (2003), a Cronbach score of >.7 equals to “acceptable” and a score of >.8 qualifies as “Good” (George & Mallery, 2003).

3.3 Procedure

The participants were initially presented information stating that this study contain a general attitude towards artificial intelligence (AI) technologies survey, regarding self-learning computer programs that analyze large amounts of data. They could then read that their participation was anonymous, and all their answers were treated confidentially. The results would be used for scientific purposes and will only be presented at a group level. Participation in the study is voluntary and the participants read that they could cancel their participation at any time (see appendix 1). Continuing, the participants were asked to complete an attention check, to assure that they partook in a seriously manner. Following, the participants were informed of how to correctly answer and rate the following survey questions, being split up into three sections: “Acceptance”, “Risk” and “Benefit”. The block orders were

counterbalanced. When the participant has successfully answered the three following sections, they continued to give answer to how familiar they estimated themselves being regarding AI technology, in four different questions. Next, the participants were presented with four different statements from the “Technology Readiness Index” (TRI 2.0) followed by eight statements taken from the “initial

validation of the general attitudes towards artificial intelligence scale”. By the end of the survey the participants were asked to fill out their gender and age, which concluded the survey.

(17)

12

3.4 Ethics

Ethical considerations were approved through the regional ethic committee

(Etikprövningsmyndigheten, 2021). Ethical steps taken were also in line with the four research ethics principles i.e., requirements concerning information, consent, confidentiality, and usage

(Vetenskapsrådet, 2002). In accordance with these requirements, the participants were informed about the study’s contents, its purpose and they all had to give their written consent beforehand.

Additionally, stored user data was processed in confidential means, though prolific has knowledge about participants, through website profiles. Access to these profiles were only available to prolific. All stored data was only to be used in this study, exclusively. Participants could leave the study at any point in time. If participants chose to exit the survey, before completion, their data was discarded.

3.5 Data analysis

The results were analyzed using a few statistical tests found on the JAMOVI platform. To analyze correlations between different variables, regression tests were done in the form of Pearson’s correlation coefficients test. As for correlations done between domains and between positive or negative attitudes and acceptance rates, two mixed model effects tests were done.

4 Results

The following paragraphs present results from the online survey, including descriptive statistics of term familiarity, acceptance to Artificial intelligence, and general attitude towards Artificial intelligence. These results are reported along with analysis done on the correlation of acceptance between domains in addition to a correlation between acceptance and general attitudes towards Artificial intelligence. Both of the latter correlations were found to be statistically significant for the set conditions, which will be further discussed in section 4.3.

(18)

13

4.1 Participant familiarity

Figure 1: Procentage of participants responding to how well they know which term

As shown in figure 1, the level of term familiarity, amongst the American participants, is shown to be relatively high across the board. 95% of the participants answered being familiar with the term

Artificial Intelligence, 60% were familiar with the term Machine Learning, 80% were familiar with the

term Computer algorithms and finally, 81% of the participants responded being familiar with the term

Personalized/Targeted Advertising.

4.2 Acceptance overview

Table 1: Overview of average participant acceptance per application, mean and SD.

Application Mean sd

Personalise commercial advertisement _3.00 _1.22

Personalise political advertisements _2.30 _1.25

Send personalized notifications regarding sales at local supermarkets _3.51 _1.19

Recommend travel destinations _3.44 _1.18

Recommend services based on personal events and/or tragedies _2.33 _1.15 Recommendation regarding further education _3.20 _1.16 Give reminders for car service schedules _3.91 _1.10 Utilize chatbots to provide customer support in online stores _3.67 _1.19 Produce online marketing and commercial texts _3.26 _1.21 Run social media accounts of businesses _2.99 _1.26

(19)

14

Recommend treatments based on patient journals _2.73 _1.22

Assist in surgical procedures _3.06 _1.22

Diagnose medical conditions _2.89 _1.21

Make financial decisions for patients suffering from dementia based on their previous such decisions

2.45 1.21

Analyse the results of CT-scans _3.34 _1.17

Analyse and propose prevention tactics concerning large-scale spread of contagions _3.53 _1.15 Automatically contact emergency services in cases of emergency _3.94 _1.03 Assist in the development in new pharmaceuticals _3.54 _1.12 Select which patients to prioritize in intensive care _2.61 _1.22 Predict risk group inhesion in patients _3.10 _1.11

As seen in table 1, the mean acceptance level varies between 2.30 – 3.94. Summed up, the average acceptance level, throughout all applications, measured M = 3.14 (sd = 1.24). The overall acceptance level was slightly above the mid-point. “Personalized political advertisements” had the lowest acceptance rate (M = 2.3, sd = 1,15), which shows a general concern with AI in relation to political aspirations. “Automatically contact emergency services in cases of emergency” was the highest-rated application (M = 3.94, sd = 1.03). In this specific life-or-death situation, the participants showed a greater level of acceptance, regarding the implications and technology that could be involved. Furthermore, basic recommender systems scored above average if they did not cross any personal boundaries, such as “Recommend services based on personal events and/or tragedies” (M = 2.33, sd

= 1.15) or “Recommend treatments based on patient journals” (M = 2.73, sd = 1.22). Similar low

results were found in financial decisions for patients when a question became personal.

The assistance of health care AI proved to be somewhat accepted on average, in questions as “Assist in

the development in new pharmaceuticals” (M = 3.54, sd = 1.12) or “Analyse and propose prevention tactics concerning large-scale spread of contagions” (M = 3.53, sd = 1.15). But when the decision

was normally done by a medical professional, the participants scored lower acceptance rates. “Diagnose medical conditions” (M = 2.89, sd = 1.21) and “Select which patients to prioritize in

intensive care” (M = 2.61, sd = 1.22) scored below average, which shows a general concern for

autonomous AI applications in healthcare.

4.3 Mixed Effects Model

For this study, a mixed effect model was used. It gives the relevant information measures needed for this within-subject study. A mixed effect model gives the advantage of systematically accounting for

(20)

15

both item- and participant-level variability. The model contained fixed effects of domain, age, and positive and negative attitudes, as well as random intercepts and slopes at both the participant and item levels.

Table 2: Fixed effects parameters estimates for AI acceptance model

Fixed effects Estimate SE df t p

(Intercept) 3.24 0.14 45.7 22.89 <.001

Domain -0.04 0.23 19.2 -0.19 0.85

Positive attitudes 0.37 0.04 475.0 9.90 <.001

Negative attitudes -0.18 0.04 475.0 -4.99 <.001

age -0.003 0.002 475.0 -1.31 0.19

In this mixed effect model analysis, the acceptance rating from each participant in each application was used as the dependent variable. As shown in table 2, no significant difference was found I the acceptance ratings between the two different domains (B = -0.04, p = .85, t(19 )= -0.19). However, a significant positive correlation was found between the composite score of positive attitudes (Sec 4.3) and acceptance (B = 0.374, t(475) = 9.90, p = <.001). An inverse significant correlation was found between the negative composite score of attitudes and acceptance (B = -0.17, t(475) = -4.99, p

=<.001). As such, positive attitudes had a stronger effect than negative attitudes. These findings show

that a positive attitude correlates with having a positive attitude. A negative attitude instead meant that the participant showed a low level of acceptance.

4.4 Correlation Matrix

For further analysis, a correlation matrix was done on the variables positive and negative attitudes, acceptance ratings and age.

(21)

16

Table 3: Correlation matrix done on positive and negative attitudes, age, and mean acceptance rating on all participants.

Age Acceptance Positive attitude Negative attitude Age - Acceptance r=-0.04*** - Positive attitude r=-0.07*** r=0.26*** - Negative attitude r=-0.05*** r=-0.19*** r=-0.32*** - Note. * p < .05, ** p < .01, *** p < .001

As presented in table 3, all the correlation tests came back as significant. Starting off, age showed a significant negative correlation with participant acceptance ratings r = -0.044, p = <.001. Similarly, age showed a significant inverse correlation with positive and negative attitudes, r = -0.073, p =

<.001 and r = -0.058, p = <.001, respectively. This shows that the older participants are on average

less accepting of the different AI applications and showed lower positive and negative attitudes, compared to the younger participants.

Furthermore, acceptance ratings showed a significant correlation with positive attitude r = 0.264, p =

<.001. Acceptance rating also showed a significant inverse correlation with negative attitude r = -0.058, p = <.001. These findings indicate a higher rate of positive attitude, while being more

accepting and similarly a lower positive attitude when the person is less accepting.

Finally, positive attitude showed a significant inverse correlation with negative attitude r = -0.329 p =

<.001.

5 Discussion

The results of the study showed that the participants had a relatively high level of familiarity with the different AI terms, where “machine learning” was the most unfamiliar term (Sect 4.1). The average acceptance, throughout all applications mentioned in the survey measured to be M = 3.14, a very minor higher-than-average acceptance rate (Sect 4.2). Comparing the two domains of applications, no significant correlations were found between the two different acceptance rating. A significant positive correlation was however found between the composite score of positive attitudes and acceptance rates. Similarly, a positive significant correlation was found between the negative composite score of

(22)

17

attitudes and acceptance rates (Sect 4.3). As presented in Sect 4.4, all correlation tests done on between the factors age, acceptance, positive and negative proved to have significant results The following sections will further explore and reflect on the results. Additionally, considerations regarding methodology as well as validity and reliability will be elaborated upon. Finally, future studies and improvement are discussed.

5.1 Reflection of the results

The lowest scoring application in acceptance rate was “Personalized political advertisements”, showing a mean acceptance rating of 2.3. The explanation might be rooted in the nationality of the American participants. Since Donald Trump won the election in 2017 there have been countless news stories about the potential sinister voter fraud, ministered by Russian AI-bots. It is possible that AI in conjunction with politics is somewhat of an infected subject, affecting the acceptance. Additionally, a report done on American attitudes and trends showed an “overwhelmingly majority of Americans (82%)” regarding robots/AI as something that should be carefully managed (Dafoe & Zhang, 2019). In a similar study presented in the introduction, 71% of the German participants answered to be “Not accepting at all” or “not very accepting” for their political views to be stored and used for

personalization (Kozyreva, Herzog, Lorenz-Spreen, Hertwig, & Lewandowsky, 2020). That could indicate that type of AI usage is frowned upon in more places than in the US.

From the mixed effect model, a significant positive correlation was found between the composite score of positive attitudes and high-rated acceptance. Having a positive attitude towards AI

applications meant that the participants also felt more accepting. The opposite was true as well, in the significant positive correlation between the negative composite score of attitudes and low rated acceptance. These correlations felt close to being expected but contextualizing the underlying connecting mechanism might be difficult. As stated in algorithm aversion (sect 2.1) and TAM (sect 2.2.3), the sentiment and expectations for AI differ depending on individual factors. Personal attachment or previous experience might play a part. Age is certainly a very large possible factor for distribution. In the correlation matrix made in sect 4.4, age showed a significant negative correlation with participant acceptance ratings, which further proves that theory. This would support the

moderating effect of age on attitude towards technology, shown in many other studies (e.g., Elias & Smith, 2012; Czaja & Sharit, 1998).

Contrary to the theory of how acceptance change depending on individual factors in combination with different applications in algorithm aversion and TAM, no significant difference was found in the acceptance ratings between the two different AI domains. This fact puts into question if ten questions in each domain are enough to get a robust result. To increase the sample validity of a similar future

(23)

18

study, the number of AI applications could be increased in all domains. Then the problem becomes a balance of sample validity and respondent bias, not wanting to tire out the participants with too many survey questions.

When asked, participants rated “Recommend services based on personal events and/or tragedies” the second-lowest acceptance level of 2.33. These results are in line with Kozyreva et al. (2020), where 83% of the respondents answered “not acceptable at all” or “not very acceptable” for the use of such personal data. This shows that people are not keen on the idea of AI applications to collect and use such sensitive data.

5.2 Limitations to the study

All participants in this study were based in the United States. Because of cultural nuances this means that the results found here cannot be generalized to any other part of the world. The participants used the Qualtrics web-based service, which meant they got paid to complete this survey. There is always the risk of abuse in such a system. If the participants complete a survey hastily, they might be able to complete more surveys in a day. Would this be the case, this could endanger the robustness of the results; participants not answering seriously. Qualtrics does have systems in placed to check for such behaviour, whilst our survey used an attention-check questions for comparable reasons. Hopefully, this combated some of that effect, but one can only speculate about the risks involved.

The questions used in the survey were designed to be compactly formatted and diverse in topic. But one could argue that the format puts the robustness of the questions into question. What would a participant answer really say, if the questions are only at a shallow surface-level? A phrase like “assist you in” makes room for own interpretations while answering something that a study would want to control for. Could one get a more accurate assessment of participant answers if the questions would be reformulated as extended and more specific? This type of change would also get away from the problem of potential habituation bias; providing similar worded questions. In the end of the day, considerations were done, inspired by the survey design in the previously named German study (Burton, Stein, & Jensen, 2019 ). The final assessment was that a compact question-format was needed to fit the web-based and mobile phone survey environment.

(24)

19

The relevance of AI is increasing in all parts of society. Although most people utilize AI every day, it is clear that everyone does not agree on acceptance throughout. To make a smooth transition to an even heavier AI-implemented society, a better understanding of the different cognitive and social factors involved are needed. This study made some early attempts at exploring how people experience acceptance for different types of AI applications. The two studied domains in this study showed no meaningful difference in rates of acceptance, contrary to the results of Kozyreva et al. (2020). This answers the first research question of this study. What was found however was that having a more positive attitude to AI applications also meant that the person felt more accepting of them. The inverse correlation was also true, finding lower acceptance rates in the negative attitude participants. Thus, answering the second research question.

In the future, more research must be done in the area of study, in many more countries. Through Algorithm aversion and TAM, we learn that many factors affect the acceptance of technology. More theory and research must be done on the underlying mechanism of technology acceptance to better understand why different applications differ. More AI domains should therefore be compared and studied. Conclusions from these findings might give some insights on what factors can be of importance in understanding AI acceptance.

(25)

20

References:

(2017). Retrieved from Prolific.co: https://www.prolific.co/

Agarwal, R., & Karahanna, E. (2000). Time Flies When You're Having Fun: Cognitive Absorption and Beliefs about Information Technology Usage. MIS Quarterly, 665-694.

AIMultiple. (2018). 18 AI Applications / Usecases / Examples in Healthcare in 2021. Retrieved from https://research.aimultiple.com/: https://research.aimultiple.com/healthcare-ai/

AIMultiple. (2019). What is data-driven decision making? Step-by-step guide. Retrieved from https://research.aimultiple.com/data-driven-decision-making/

Alexander, V., Blinder, C., & Zak, P. J. (2018). Why trust an algorithm? Performance, cognition, and neurophysiology. Computers in Human Behavior, 279– 288.

Boschert, S., Coughlin, T., & Ferraris, M. (2019). Symbiotic Autonomous Systems. digitalreality.ieee.org.

Brown, R. V. (2015). Decision science as a by‐product of decision‐aiding: A practitioner's perspective.

Journal of Applied Research in Memory and Cognition 4, 212– 220.

Burton, W. J., Stein, M.-K., & Jensen, T. B. (2019 ). A systematic review of algorithm aversion in augmented decision making. Journal of Behavioral Decision Making, 220-239.

Chin, W. W., & Todd, P. (1992). On the Use. Usefulness, and ease of use of structural equation

modeling in MIS research: A note of caution. Calgary, Alberta : University of Calgary .

Colarelli, S. M., & Thompson, M. (2008). Stubborn reliance on human nature in employee selection: Statistical decision aids are evolutionarily novel. Industrial and Organizational Psychology, 374–351.

Czaja, S. J., & Sharit, J. (1998). Age differences in attitudes toward computers. J Gerontol B Psychol Sci

Soc Sci, 29-40.

Dafoe, A., & Zhang, B. (2019). Artificial Intelligence: American Attitudes and Trends. University of Oxford: Center for the Governance of AI, Future of Humanity Institute.

Davis, F. D. (1989). Perceived Usefulness, Perceived Ease of Use, and User Acceptance of Information Technology. MIS Quarterly.

Diab, D. L., Yueh Pui, S., & Yankelevich, M. (2011). Lay Perceptions of Selection Decision Aids in US and Non‐US Samples. International Journal of Selection and AssessmentVolume 19, Issue 2, 209-216.

Doumpos, M., & Grigorousdis, E. (2013). Multicriteria Decision Aid and Artificial Intelligence: Links,

Theory and Applications. Wiley-Blackwel.

Elias, S., & Smith, W. L. (2012). Behaviour and Information Technology. Behaviour and Information

Technology, 453-467.

Etikprövningsmyndigheten. (2021). Retrieved from

https://etikprovningsmyndigheten.se/for-forskare/sa-gar-det-till/

Gartner. (2019 ). Gartner Survey Shows 37 Percent of Organizations Have Implemented AI in Some

(26)

21

https://www.gartner.com/en/newsroom/press-releases/2019-01-21-gartner-survey-shows-37-percent-of-organizations-have

George, D., & Mallery, P. (2003). SPSS for Windows step by step: A simple guide and reference. Boston: Allyn & Bacon.

Grove, W. M., Zald, D. H., Lebow, B. S., Snitz, B. E., & Nelson, C. (2000). Clinical versus mechanical prediction: A meta-analysis. Psychological Assessment, 12(1), 19–30.

Kozyreva, A., Herzog, S., Lorenz-Spreen, P., Hertwig, R., & Lewandowsky, S. (2020). Artificial

intelligence in online environments: Representative survey of public attitudes in Germany.

Berlin: Max Planck Institute for Human Development.

LibGuides: Statistical & Qualitative Data Analysis Software: About Qualtrics. (2018). Retrieved from

Kent.edu: https://libguides.library.kent.edu/statconsulting/qualtrics

Meehl, P. (1955). Clinical Versus Statistical Prediction: A Theoretical Analysis and a Review of the Evidence. Psychology.

Muir, B. M. (1987). Trust between humans and machines, and the design of decision aids.

International Journal of Man‐Machine Studies, 527–539.

Parasuraman, A., & Colby, L. C. (2015). An Updated and Streamlined Technology Readiness Index: TRI 2.0. Volume: 18 issue: 1, 59-74.

Schepman, A., & Rodway, P. (2020). Initial validation of the general attitudes towards Artificial Intelligence Scale. Computers in Human Behavior Reports, Volume 1.

Vetenskapsrådet. (2002). Forskningsetiska principer inom humanistisk-samhällsvetenskaplig

(27)

22

Appendix 1

Survey – Participant information

This is a study about general attitudes towards artificial intelligence (AI) technologies,

self-learning computer programs that analyze large amounts of data.

Participation in this study is anonymous and your answers are treated confidentially.

The results will be used for scientific purposes and will only be presented at a group

level. Participation in the study is voluntary and you can cancel your participation at

any time (by closing down the page).

Do you agree to participate in the study?

o Yes, I have read the information above and agree to participate. I understand that I have right to withdraw from the study whenever I want.

AI acceptance and attitudes : People’s perception of healthcare and commercial AI applications