The Threats of Artificial Intelligence Scale (TAI)

In recent years, applications making use of Artificial Intelligence (AI) have gained re-newed popular interest. Expectations that AI might change the face of various life domains for the better are abundant [1‐3]. Be it medicine, mobility, scientific progress, the economy, or politics; hopes are that AI will increase the veracity of input, effectiveness and efficiency of procedures as well as the overall quality of outcomes. Irrespective whether changes apply to the workplace, public management, industries producing goods and services as well as private life: As usual with the diffusion of new technologies there is tremendous uncertainty as to how exactly developments will play out [4], what social consequences will manifest and to what extent respective expectations of stakeholders and societal groups will materialize. Oftentimes, there will be some people that immensely profit from socio-technological innovations, while others are left behind and cannot cope with the unfolding of events [5]. Thus, whenever new technologies bring about social change, the success of their implementation or failure depends upon the reaction of the affected people. People might happily accept new technology, they might not care nor use it at all, or they may even show severe reactance towards it [6]. There is first empirical evidence suggesting that the general public itself shows some considerable restraint when it comes to the broad societal diffusion of AI applications or robots that might even border on actual fear of such technology [7‐9]. However, as fear and respective threat perceptions are presuppositional theoretical constructs, they necessitate a more fine-grained approach that goes beyond broad claims of concerns or even fear regarding autonomous systems.

Accordingly, in this paper, we argue for an improved assessment of the perceived threats of AI and propose a survey scale to measure these threat perceptions. First, a broadly usable measurement would need to address perceived threats of AI as a precondition to any actual fear experienced. This conceptual difference is subsequently based on the literature on fear and fear appeals. Second, the perceived threat of AI would need to take into account the context-dependency of respective fears as most real-world applications of AI are highly domain-specific. AI that assists in the medical treatment of a person’s disease might be perceived vastly different from an AI that takes over their job. Third, not only do perceptions hinge on the domain in which people encounter AI applications, it would also be necessary to differentiate between the extent of an AI’s actual autonomy and reach in inflicting consequences upon a person. Thus, it needs to be asked to what extent the AI is merely used for analysis of a given situation, or going even further, whether the AI is used to actively give suggestions or even making autonomous decisions.

As the field of application is crucial for the mechanism and effects of threat perceptions concerning AI, any standardized survey measure needs to be somewhat flexible and individually adaptable to accommodate the necessities of a broad application that considers AI’s functions and the context of implementation. That is why our scale construction opts for a design that can easily be adapted to varying research interests of AI scholars.

Consequently, we developed a scale addressing threats of AI that takes into account such necessary distinctions and subsequently tested the proposed measure for three domains (i.e. loan origination, job recruitment and medical treatment that are subject to an AI application) in an online survey with German citizens. In our proposed measure of perceived threats of AI, we aim to cover all aspects of AI functionality and make it applicable to various societal fields, where AI applications are used. Thereby, we highlight three contributions of our scale, that are addressed in the following: The collected data supports the factorial structure of the proposed TAI scale. Furthermore, results show that people differentiate between distinct AI functionalities, in that, the extent of the functional reach and autonomy of an AI application evoke different degrees of threat perceptions irrespective of domain. Still, such distinct perceptions do also differ between the domains tested. For instance, recognition and prediction with regard to a physical ailment as well as the recommendation for a specific therapy made by an AI do not evoke substantial threat perceptions. Contrarily, autonomous decision-making in which an AI unilaterally decides on the proscribed treatment was met with relatively bigger apprehension. At the same time, the application of AI in medical treatment was generally perceived as less fearsome than situations where AI applications are used to screen applicants on a job or a financial loan.

In recent years there has been a somewhat re-newed interest in applications of AI based on recent developments in computer technology that allows for use of extensive processing power and the analysis of vast amounts of so-called Big Data applications of Machine Learning, Deep Learning and Neural Networks. Such applications gather under the label of AI, which is ascribed a huge impact on society as a whole [10]. Thereby, AI has especially seen widespread use in business and public management [11]. As a consequence, the public discourse regarding AI is mainly driven by companies that provide AI technology looking for customers and markets for their products [1, 12, 13]. Meanwhile, empirical evidence from survey research supports the assumption that AI is not per se perceived as entirely positive by the public. A cross-national survey by Kelley et al. [10] shows that AI is connected with positive expectations in the field of medicine, but reservations are prevalent concerning data privacy and job loss. Another concern is raised by Araujo et al. [14], who state that citizens perceive high risks regarding decision-making AI. Moreover, a representative opinion poll by Zhang and Dafoe [15] illustrates that Americans as well as citizens from the European Union (EU) believe that robots and AI could have harmful consequences for societies and should be carefully managed. Additionally, Gnambs and Appel [16] show that attitudes towards robots have recently changed for the worse in the EU. Especially, when it comes to the influence of robots in the economy and the substitution of workforce, people express fear [7, 9]. On a broader level, a recent study by Liang and Lee [8] inquiring about the fear of AI even found that a considerable amount of all Americans reported fears when it comes to autonomous robots and AI.

Using data from the Chapman Survey of American Fears, Liang and Lee [8] set out to investigate the prevalence of fear of autonomous robots and artificial intelligence (FARAI). They come to the conclusion that roughly a quarter of the US population experienced a heightened level of FARAI. In the respective study, participants were confronted with the question “How afraid are you of the following?”. The FARAI-scale was afterwards built out of these four items: (1) “Robots that can make their own decisions and take their own actions,”, (2) “Robots replacing people in the work-force”, (3) “Artificial intelligence” and (4) “People trusting artificial intelligence to do work”. All items were rated on a four-point Likert scale with answers ranging from “not afraid (1), slightly afraid (2), afraid (3), to very afraid (4)” [8].

While the authors shed some first light in addressing threat perceptions of artificial intelligence and generate valuable insights into various associations of FARAI with demographic and personal characteristics, there is also need for a potential enhancement of the existing measurement of FARAI. As the FARAI scale was developed out of a broad questionnaire concerning many possible fears people in the US might have, the measurement was not specifically developed for measuring the distinctive fear of robots and AI, respectively. The FARAI scale also varies in its scope. While item 3 broadly queries the fear of AI in general, item 2 specifically inquiries about its specific impacts on the economic sector. Items 1 and 4 query a specific functionality of AI, with item 4 focusing on the human-machine connection. Thus, the items are mixed in their expressiveness and aim at different aspects of AI’s impact. Accordingly, the scale does not allow for distinct assessments of AI and necessary specifications concerning its domain of application and the employed functions.

Besides, the public understanding of robots and AI might be influenced by popular imaginations from pop-culture, science-fiction, and the media, as is also already implied by Liang and Lee [8] and Laakasuo et al. [17]. Due to the popularity of vastly different types of autonomous robots and AI in literature, film and comics, it is hard to pin down what exactly comes to a person’s mind inquired about both terms. Delineating boundaries may not be possible when it comes to the public imagination. As a survey research in the UK by Cave et al. [18] suggests, a quarter of the British population conflates the term AI with robots. Accordingly, a conceptual clarification concerning the distinction between the terms robot and artificial intelligence is required to begin with. In the FARAI measure by Liang and Lee [8], there is a mixture between both terms as two question items focus on each term, respectively. This terminological distinction is often conflated in empirical research [19]. We believe that the mixture of the terms might lead to avoidable ambiguity and maybe even confusion, since people may think of two distinct and even completely different phenomena or might not be able to distinguish between the two constructs at all. According to the Oxford English Dictionary a robot is “an intelligent artificial being typically made of metal and resembling in some way a human or other animal” or “a machine capable of automatically carrying out a complex series of movements, esp. one which is programmable” while AI is “the capacity of computers or other machines to exhibit or simulate intelligent behaviour”. There is certainly some conceptual overlap by definition, especially with regard to the capacity of intelligent behavior demonstrated by an artificial construct, hence, something that does not exist naturally, but is human-made. It also cannot be ruled out that appraisal of robots may be strongly associated with AI, especially when such robots are depicted as autonomous.

Recently, the term AI, particularly, has renewedly received widespread attention and describes techniques from computer science that gather many different concepts like machine learning, deep learning or neural networks, which are the basis of autonomous functionality and pervasive implementation. As a consequence, we decided to focus our measurement solely on AI as it depicts the core issue of the nascent technology, i.e. autonomous intelligent behavior, which applies to many use cases that do not necessarily include a physical machine in motion.

There is plenty of literature on the subject of fear, especially from the field of human psychology. Altogether, fear is defined as a negative emotion that is aroused in response to a perceived threat [20]. When it comes to the origins of emotion, many studies rely on the appraisal theory of emotion: “The appraisal task for the person is to evaluate perceived circumstances in terms of a relatively small number of categories of adaptional significance, corresponding to different types of benefit or harm, each with different implications for coping” [21]. Accordingly, the authors define relational themes for different emotions. According to Smith and Lazarus [21] anxiety, respectively fear, is evoked, when people perceive an ambiguous danger or threat, which is motivationally relevant as well as incongruent to their goals. Thereby, a threat is seen as an “environmental characteristic that represents something that portends negative consequences for the individuum” [22]. Furthermore, people perceive low or uncertain coping potential. In other words: Fear is the result of a persons’ appraisal process, where a situation or an object is perceived as threatening and relevant as well as no avoiding potential can be seen. If theses appraisals are processed, people react with fear and try to avoid the threat [21], i.e. in turning away from the object.

Many scholars build on appraisal theory to develop more specified theories on the mechanisms of fear. Especially in health communication much work on so called fear appeal literature has been done [22, 23]. In a nutshell, most fear appeal theories state that a specific object, event or situation (e.g., a disease) threatens the well-being of a person [24, 25]. With the development of the Extended Parallel Process Model (EPPM), Witte [25] theorizes that this threat is at first processed cognitively. Thereby, severity and susceptibility of the threat as well as coping potential (self and general), i.e. the amount of efficacy [25] respectively control [26], is rated. Depending on the weights of these cognitive apprehensions, people react differently. Fear emerges, when the threat perception is high, while the coping perception is low. As a result, message denial arises that is mostly characterized by avoiding the threat. On the other hand, when threat as well as coping potential are perceived as high, message acceptance results. If this happens, people actively engage with the threat, for example in gathering information about the threat or actively combating potential harms. In this case, fear does not emerge. Whereas the empirical examination of the EPPM found no clear proof [27] and many suggestions for extending the model have been made [28, 29], scholars agree upon the central persuasion effects of threat and coping perceptions [30]. Moreover, the EPPM commonly serves as a framework for further research [31].

Transferred to the subject of perceived threats of AI, we belief that AI is best described as an environmental factor that might cause fear. However, AI should not be treated as a specific fear itself. In our view, fear may be a result of a cognitive appraisal process, where AI depicts a potential threat origin. Thus, we explicitly focus on threats of AI, not fear of AI. This idea becomes more prevalent in thinking about an actual situation. For example, a person is confronted with an AI system that decides over an approval of a credit. This person most likely will not be afraid of the computer system, but will rather evaluate cognitively the threat that such a system might pose to its well-being. The person then rates the probability of the system causing harm (e.g., if it denies the credit). If the outcome of this process ends in a negative evaluation for the person, fear will be evoked. However, this fear is based on the threat the AI systems poses and not on the AI system itself. This is crucial for our understanding of threats of AI.

It is also important to address the social situation, in which a threat is perceived. Smith and Lazarus [21] already stated that an “appraisal can, of course, change (1) as the person-environment relationship changes; (2) in consequence of self-protective coping activity (e.g. emotion-focused coping); (3) in consequence of changing social structures and culturally based values and meanings; or (4) when personality changes, as when goals or beliefs are abandoned as unservicable”. Furthermore, Tudor [32] proposed a sociological approach for the understanding of fear. He developed a concept, in which he distinguishes parameters of fear including environments, cultures as well as social structures.

Thereby, contexts can vary in manifold ways. A rather simple example for what Tudor [32] refers to as an environmental context is the case of the wild animal: for instance, a tiger could face a human being; however, arguably there is a huge difference in fear reaction if one is confronted with the tiger in a zoo or in its natural habitat. Thus, the environmental factor “cage” does have a huge impact on the incitement of fear. Additionally, cultural backgrounds can affect the way threats are perceived: “If our cultures repeatedly warn us that this kind of activity is dangerous, or that sort of situation is likely to lead to trouble, then this provides the soil in which fearfulness may grow” [32]. Lastly, social structures described that the societal role of an individual might influence threat perceptions. For instance, that could be the job position of an employee or just the belonging to a specific societal group.

Furthermore, different social actors are able to influence the social construction of public fears, i.e. if and how environmental stimuli are treated as threats [26]. According to Dehne [26], the creation of fear, among other factors, is dependent upon transmission of information in a society. In that, especially scientific, economic, political and media actors affect the social construction of threats. However, depending on the actors that take the highest share in the public discourse, different threat perceptions might emerge. For instance, given an AI application in medicine, we assume that science as well as media lead the debate. On the other hand, an AI application in the field of recruiting will probably be led by economic actors. It is plausible that there are specific context dependencies (who informs the public about a specific AI application) that have an influence on (threat) perceptions.

In summary, there are many (social) factors that shape the way emotions are elicited leading to the conclusion that threat perceptions heavily rely on the context in which an individual encounters AI. Of course, we are not able to cover all possible contexts of AI related threats. However, we distinguish two context groups, which are important for the understanding of TAI: AI functionality and distinct domains of AI applications.

As outlined in Sect. 4, perceived threats of AI are a precondition of emotional (fear) reactions. Thus, we assume that threat perceptions concerning AI actually trigger fear reactions. Accordingly, hypothesis 2 reads as followed.

As threat perceptions of AI functionalities and domains may differ vastly from each other, we are interested in whether the amount of perceived fear (if any) that is explained by our proposed measure also differs by context. Arguably, not all threat perceptions necessarily need to cause the same fear reactions. For instance, if subjects perceive high levels of efficacy in dealing with the potential threat, a far less strong emotional reaction is likely to occur [25]. This becomes particularly obvious, when comparing the recommendation and the decision-making functionality. Decision-making AI takes control away from the individual, whereas in recommendation at least an (other) human still has control over the process.

Therefore, we test whether the measure is able to capture induced fear reactions across the different contexts and whether these differ. Accordingly, we pose and address the following research question:

Thus, the aim is to reliably and validly assess the extent to which respondents perceive autonomous systems as a threat to themselves. Moreover, the scale needs to be standardized allowing for comparisons between samples from various populations, but flexible enough allowing for application in distinct domains of AI research. It thus needs to be as concise as possible affording to be included in brief questionnaires.

We tested our scale using a non-representative German online access panel. We used an online questionnaire with a split survey design comparing the threat perceptions towards AI in the three sectors loan origination, job recruitment, and medical treatment. Participants were randomly assigned to one of the three groups. We chose the method of matching urns after completion of the questionnaire to ensure an even distribution among the different questionnaire groups.

Participants were recruited from the SoSci Open Access Panel between the 30th September and 14th October 2019 [51]. All in all, 917 subjects completed the questionnaire. In the data cleaning process, we had to drop 26 cases from our data set. Data elimination was based on two criteria: minus scores deployed by the access panel as well as time for completing the questionnaire. The minus scores are calculated on basis of the sum of the deviations from the average time for answering individual questionnaire pages in order to identify respondents’ inattentiveness. For our data cleaning process, we chose the access panel’s recommended value for a conservative cut-off criterion to ensure a high-quality data set. Moreover, we checked the overall time score participants needed to fill out the questionnaire. We excluded all data from participants with an answering time below five minutes, since we defined this minimum value after a pre-test of our questionnaire. Thus, our final sample consists of n=891 participants.

Turning to the distribution by questionnaire groups, 296 participants were assigned to the ‘loan origination’ group, 294 to the ‘job recruitment’ group and 301 to the ‘medical treatment’ group. Of all participants 445 identify as female (50.0%) and 438 as male (49.2%), while seven (0.8%) respondents identify as non-binary. The average age of the participants is approximately 46 years (SD=15.66). Because of the demographic structure of the access panel, 82 percent of our participants have the highest German school leaving certificate. Arguably, our data is not representative for the German population, which should be acknowledged when interpreting the descriptive results. However, our primary interest of introducing a valid scale remain unchallenged by this limitation.

To test our measurement, we performed confirmatory factor analyses (CFA) with the lavaan package [53] in R (version 4.0) as well as several test statistics with the semTools package [54]. For visualization we used the semPlot package [55]. Firstly, we calculated a CFA with configural invariance. To check, if the factor loadings differ between the applications, we secondly calculated a model with measurement invariance. Thirdly, we constrained the intercept of the measurement to check for scalar invariance, i.e. to analyze whether threat perceptions are different between application areas. Lastly, we set one intercept free to gain our final model; that is, our results suggest the TAI scale as a measurement with partial scalar invariance. To check the influence of the TAI scale on emotional fear, we built a structural equation model (SEM) with emotional fear as dependent variable. We will further elaborate on our findings.

In this paper we introduced a scale to measure threat perceptions of artificial intelligence. The scale can be used to assess citizens’ concerns regarding the use of AI systems. As AI technologies are increasingly introduced into everyday life, it is crucial to understand under what circumstances citizens might feel themselves to be threatened. Threats can be understood as a pre-condition of fear. Subsequently, according to fear appeal literature, being frightened can lead to denial and avoidance of the threatening object. Thus, if people perceive AI as a serious threat, it could cause a non-adoption of the technology.

However, AI is an umbrella term for a huge variety of different applications. AI applications can fulfill various functions and applications are used in almost every societal field. Arguably, there are huge variances in threat perceptions of different functions and domains of application. With the TAI-scale we propose a measurement to account for this context-specificity.

First, the results suggest that threat perceptions of distinct AI functions can be reliably differentiated by respondents. Recognition, prediction, recommendation and decision-making are indeed perceived as different functions of AI systems. However, depending on the context evaluated the measure showed diverging factorial validity. In one case the indicator items had significant shared variance with more than one dimension. This impairment of discriminatory power indicates that thorough pre-testing of the adapted measures and data quality control are of utmost importance when devising the survey instrument in subsequent study designs. In doing so, researchers need to make sure that respondents fully comprehend the item wording and that the object of potential threat is clearly recognizable. Especially, this becomes important when respondents are confronted with new and technically sophisticated AI systems, for which there not yet exists enough direct personal experience.

Second, threat perceptions are shown to vary between different domains, in which AI systems are deployed. This suggests that the notion of a general fear of AI needs to enhanced in favor of a broader conception not only of what actions AI is able to perform, but also what exactly is at stake in a given situation. In cases where AI systems seem useful and the consequences of its application appear insubstantial, the introduction of AI in another domain might evoke entirely opposite reactions. Thus, while general perceptions such as general predispositions concerning digital technology certainly do play a role when it comes to the evaluation of innovative AI systems, a more fine-grained approach is necessary and appears to be fruitful with the developed measurement. Respondents’ threat perceptions in this study varied considerably between domains. Especially, the use of AI in medical treatment was only perceived as lightly threatening, whereas threat perceptions were quite higher in the domains of job recruitment and loan origination. Regarding the levels of threat perceptions concerning the functionalities, it is evident that the decision-making function is perceived as most threatening within all three domains. Arguably, this might be based on the loss of humans’ autonomy. As this is only a hypothesis at this point, further studies should elaborate on these findings. Future applications of the TAI scale will yield further insights concerning the items’ and scales’ sensitivity with regards to different domains of AI applications.

Third, threat perceptions are reliable predictors of self-reported fear. As the measurement of actual fear is rather complicated via the means of a survey instrument, the inquiry of threat perceptions appears not only to be preferable. It also suggests that it is reasonably well connected to individual self-reports of experienced fear. Further studies should elaborate on these findings and focus on the behavioral impact of AI-related threat perceptions. As the fear appeal literature suggests, one might expect that high levels of AI-induced fear lead to rejection of the technology or even protest behavior.

Highlighting the good fit of our scale, we encourage researchers to implement the TAI scale in research focusing on public perceptions of AI systems. As outlined earlier, the TAI scale can be seen as a toolbox. Hence, it is possible to integrate only those functional dimensions in a survey that actually fit the AI system under research. Anyhow, we also advise researchers to be mindful when using the scale. In practice, there is a lot of confusion about the terminology of AI - even within the scientific community. Researchers using the scale have to make sure that the AI system under research actually performs AI tasks. Given the fact that usually non-experts serve as respondents, scholars have the responsibility to inform them rightfully about what the specific AI system under consideration is and what it is able to do. Otherwise, researchers would make claims about a threat of AI perceptions without actually examining AI.

Future studies should test the TAI scale in surveys employing representative sampling to make statements over the actual level of threat perceptions regarding the different functionalities of AI systems in various domains. With that, it would be possible to grasp the public threat perception of AI systems and draw conclusions for further implementation of AI in society. However, such research needs to be thoroughly theorized, as the mere information concerning levels of threat perceptions of AI with the public is of little academic value.

Another promising future direction of research could focus on the role of knowledge in attitude building. Knowledge about AI technology could influence the way individuals feel threatened by AI. Additionally, with this extension of research one may test, whether the perceptions of what an AI system is capable of do in fact match the technical level. It may be possible that the imagination of individuals and the real performance of AI systems might not correspond.

As this study attempts to develop a more fine-grained approach to measuring threat perceptions regarding AI, its focus lay on scale construction and testing the application of the scale in an online survey. The results are thus limited to German online users from a non-representative online access panel. Further research should extend the scope of the domain of AI applications as well as addressing further groups of stakeholders and, especially, behavioral consequences of perceived threats of AI.

Furthermore, a translation of the scale to other languages appears as another promising avenue. As Gnambs and Appel [16] showed, based on longitudinal data from the Eurobarometer attitudes towards robots and autonomous systems vary between countries and might be subject to cultural influences that warrant research illuminating divergent perceptions and their antecedents.

Finally, we point out that, although we refer to the periodic system of AI [34] and the study of Hofmann and colleagues [33], the functional classes may be considered somewhat arbitrary. As AI is a very broad term, there might be other possibilities for dimensional structures of a scale focusing on public threat perceptions. However, our results give support for the dimensional structure we proposed.

The public perception of Artificial Intelligence will become increasingly important as applications that make use of AI technologies will further proliferate in various societal domains. A populace that perceives AI as threatening and that in consequence fears its proliferation may prove as detrimental as a blind trust in the benevolence of actors that implement AI systems as well as a general overestimation of the veracity of assertions and decisions made by AI. Consequently, the survey of threat perceptions of various AI systems is of great research interest. In this paper, we proposed and constructed a measurement of threat perceptions regarding AI that is able to capture various functions performed by AI systems and that is adaptable to any context of application that is of interest. The developed TAI scale showed satisfactory results in that it reliably captured threat perceptions regarding the distinct functions of recognition, prediction, recommendation, and decision-making by AI. The results also suggest that the developed scale is able to elucidate differences in these threat perceptions between distinct domains of AI applications.