Embodying algorithms,enactive artificial intelligence and the extended cognition: You can see as much as you know about algorithm

1.1. Embodied AI and enactive algorithm

Embodied AI grows from the idea of embodied cognition [12] that the actions of users can play a role in the development of algorithms and AI. The idea has been applied to personalised algorithms to improve their performance and functionality. Process automation, chatbots, advanced robotics, autonomous drive technology and personal companions could all benefit from embodied AI. How people interact with AI becomes important when justifying algorithm design and development [10,13]. Personalisation algorithms, for example, influence what people have chosen in the past, what they Machine learnings are based on programmed algorithms that learn and optimize their automated operations by analyzing input data to make predictions [1]. With the continuous feeding of user data, algorithms can make more accurate and personalized predictions. How users perceive personalization algorithms, to what extent users approve of the use of their data, and how much users are gratified with their personalized results all determine the operation and performance of personalization algorithms. User heuristics regarding algorithmic qualities consider questions such as: how do users determine the qualities or features of AI? How do users perceive them and with what effects? Because algorithm-based content such as personalised algorithms brings numerous competitive edges, it is essential to examine what users’ prior expectations are and how these expectations are constructed. Also important is how users’ trust affects emotion and satisfaction, subsequently influencing user intentions [14]. Heuristic–systematic processing (HSP) model is an appropriate frame for this task insofar as the model posits that pre-behaviours and post-experiences influence user evaluation, which in turn leads to satisfaction and continuance intentions [15]. We apply HSP for examining how normative value and performance can influence the way users understand algorithmic processes and evaluate algorithmic qualities. The theory is suitable for this inquiry because it is designed to examine user behaviours depending on expectations, performance and user confirmation of beliefs based on cognitive processes. As algorithms and AI afford people unique experiences, HSP can be extended by integrating algorithm-specific factors such as antecedents of trust and by incorporating usefulness and ease of use as part of performance expectancy.

HSP has been extensively applied to analyze how people shape their motivations and beliefs. The framework proposes a dual-process, heuristic and systematic. Heuristic processing occurs based on the behavior to maximize evaluations with the least amount of cognitive effort [15]. Systematic processing involves when users have strong motivation and capability to process information, they scrutinize the information content, utilizing prior knowledge and beliefs to make a reasonable judgment [16]. HSP can be nicely used in artificial intelligence systems to explain information-seeking behaviors during algorithmic exposure [9]. When users receive algorithmic information, they either actively evaluate the information by elaborating cognitive resource or simply see through it based on their limited knowledge and understanding.

Recently, HSP has been used in algorithms and AI. Alarcon et al. [17] applied HSP to the perceptions of trust in algorithms and found that transparency, readability and source reputation significantly influenced trust perceptions. We use HSP as a lens of which to search, evaluate, and make sense of personalization in reference to algorithmic attributes.

1.2. Embodied cognition of FATE

The embodied and enactive approach to AI has become a feasible alternative to algorithm design with respect to the practical goal of developing personalised AI, which can behave in a robust and reflexive manner [12,18]. Nevertheless, some concerns have recently been raised with regard to FATE [1,7]. Algorithm journalisms, for example, present challenges that are both operational – how might we present what we think people like – and normative – how do we know what people like and what are we assuming about who people are? These challenges are related to FATE, which has been the most common topic of debate in AI [10]. Generally, the ‘E’ in the FATE is referred to as ethics in other studies. This study replaces ethics with explainability because the notion of ethics is exceedingly overarching and broad, making it difficult to operationalise and measure effectively. In addition, the issue of explainability and related concepts such as interpretability and understandability become important [6], and have become more frequent topics of discussion in AI, providing opportunity to conceptualise and operationalise. These issues can be turn into emergent questions.

First, how can we ensure fair algorithms and reliable AI such as personalised recommendation? Second, how can we develop AI that is more accountable and transparent (Diakopoulos) [19]? In much of the current discussions on trust in AI, issues of FATE are frequently touted without specific knowledge of how these issues are related to trust [1,19]. The journalism industry, for example, should be prepared for questions regarding FATE because structurally biased and discriminatory algorithms may turn into severe risks in reported news, thereby lending urgency to the debate on how to develop algorithm journalism that is fair, transparent and accurate [20].

While fairness has been touted along with rising AI, there is a lack of generally acknowledged definition of fairness in algorithmic terms [1,11]. It is critical to consider what fairness means in the specific context of users. In the case of designing AI systems, an important consequence to avoid is the creation or reinforcement of unfair bias [21]. Contrary to general perception, the reality of AI operation is that AI does not always function fairly. Without a careful approach to fairness in every process of AI design, that system can produce discriminatory results for certain groups or individuals. The fairness of an algorithm can be measured based on its accuracy (e.g. the rate of right prediction), recall (e.g. the capability to locate related results) and precision (e.g. the capability to produce exact results).

The concept of transparency in the context of personalised algorithms requires that recommendations made by algorithmic processes are obvious – that is, transparent – to the users (Ananny and Crawford) [22]. Social media platforms, for example, apply personalisation algorithms to ensure the content shown to the user is engaging and relevant [1,22]. Yet, there is normally no transparency in the criteria used for ranking the information, and even if there is some information about the process, such information is limited and hard to understand for the layperson. There is an increasing call for transparency and explainability to make algorithms comprehensible to those affected by them [23]. There is debate regarding what constitutes a transparent explanation and what level of transparency is needed, as well as transparent to whom and for what purpose [24]. In this light, Shin and Park [11] propose a term of algorithmic transparency and define it as the requirement that the users understand how a decision/prediction is made by the AI system. Diakopoulos and Koliska [21] define transparency as the algorithmic inputs together with the algorithm itself should be visible, and thus understood. Algorithmic transparency is related to terms such as algorithmic visibility, explainability and interpretability [11]. In this way, the decisions involved in algorithm outputs are interpretable, and in addition, algorithmic processes and intentions can be properly accounted for in their entirety [25]. For example, transparency is a mechanism to determine whether legitimate data are used, whether the implemented algorithm is statistically and mathematically appropriate for the task at hand and whether the goal of the algorithm is valid and justifiable [26]. When people know how algorithm works and how operations are executed in machine learning, they are more probably to use the content properly and trust the algorithm and resulting outputs [1,9].

While similar to the concept of transparency, explainability or interpretability in an algorithm refers to how methods and techniques in the application of AI should be understandable by humans [27]. Explainability means algorithms that is transparent in its processes so that users are able to understand and trust decisions [6]. Explainable AI stipulates that the internal operations of a machine learning should be understood in human languages. As algorithm technologies become black boxes in decision-making processes, there has been debate about the extent to which people are able to understand how the resulting decision-making algorithm works [18]. There has been increasing calls for more transparent AI explainability, as part of attempts to establish ethical codes into the design and development of AI systems.

The debate of accountability in AI falls on who is liable for the outcomes of AI services [11]. Diakopoulos [19] argues that the providers of an algorithm should be held accountable for the consequences of their AI. While there is a divergent view on who is responsible for AI, the consequences of unintended and unpredictable circumstances are severe. Examples include self-driving cars creating traffic accidents, Google Home devices accidentally locking in, and facial recognition AI discriminating against certain races and gender. Because an algorithm cannot be held legally accountable, human accountability needs to be embedded in all stages of news creation. Platform providers such as Google, Amazon and Netflix hold great power in informing and shaping public opinion because AI is used to decide the relative value of content that appears in front of users. While better measures are needed to ensure that intentionally misleading content is rooted out, the question stands as to how to combat fake news and disinformation in the algorithm ecology.

Predicting the possibility of unintended consequences is important in addressing algorithmic accountability in AI. Numerous scholars have begun to scrutinise the process of assigning responsibility for harm if algorithmic news reporting results in inequitable and discriminatory output [11]. While algorithmic accountability has been neither conceptualised nor operationalised in a journalism context, the applicability of policies of accountability for algorithms has been discussed alongside practical strategies. Journalism industries, as well as AI industries overall, should be able to answer pointed questions about accountability because biased and black-boxed algorithms can result in harmful risks. Such great risks lend urgency to a discussion of how to make algorithms fair, accountable, transparent and interpretable, and therefore, trustworthy and widely accepted.

1.3. Hypotheses

Based on the above discussion, the following are hypothesised in this study:

2.1. User sensemaking of algorithmic characteristics

Machine learning algorithms have a growing influence over major decisions. The decisions that machine learning supports have a major impact on the way people consume information and the way they communicate, and thus how they see the world [22]. Without FATE, algorithmic innovation would be difficult to achieve the desired goals of user centricity.

Recommending content/items requires a more detailed engagement of issues of FATE. Taken together, FATE brings up key considerations in the design and development of algorithm services [9,21]. Personalised algorithms such as Netflix and Hulu are essentially designed to produce accurate recommendation systems [18]. How these personalisation processes are conducted, whether the recommendation results actually reflect user preferences and whether the results are reasonably accountable remain unanswered questions. Thus, FATE emerges as the most fundamental issue of AI services [6,28]. In particular, there have been rising concerns about the trustworthiness of AI news services [29,30]. There are increasing concerns about whether and to what extent we should believe what we hear about algorithms and what the algorithms recommend [11]. Whether or not users trust algorithms depends on the objectiveness and consequence of the task, and also on the manner the algorithm itself is presented. Perceptions of trustworthiness impact on algorithms and consequently, influence a user’s decision and behavior related to the service or product. Since beliefs about atechnology’s effectiveness are key determinants for its adoption, user trust in algorithms is key to its development. Ethical considerations such as FATE behind the design or deployment of an AI-based service or product can impact perceptions of trust.

People are inclined to use trustworthy systems because they are familiar with how data are collected, how they are processed for personalisation and thus, how recommendations are produced [6]. When there is a transparent mechanism, users can revise their input data to improve recommendation outputs. Algorithm users are able to understand the logic of recommendation system [18]. Providers of algorithms strive to ensure accuracy and legitimacy of results to increase user trust. Together, transparency, fairness and accuracy play critical roles in algorithm services by improving user trust in algorithms [9]. When transparent, fair and accurate services are guaranteed, users are more probably to perceive higher credibility with regard to the news. High levels of transparent algorithm can afford users a sense of personalisation. Fair and accountable news affords users a sense of trust, which in turn promotes a sense of satisfaction and continued use. User awareness and understanding of why and how a particular recommendation is generated are known to be significant. Great visibility and clear transparency for relevant feedback improve search performance and satisfaction with recommendation systems. The previous work shows that explanation can improve users’ overall satisfaction with a recommendation system [31]. Shin and Biocca [32] argue that a user’s confirmation level directly influences satisfaction in the context of technology adoption. The work of Zhang, Wang and Jin [33] confirms a causal relationship between explainability and assurance in the context of algorithm services.

2.2. Performance expectancy

Perceived value is the worth or benefits a user ascribes to a product or service [32]. In the algorithm context, extensive research has consistently shown that perceived usefulness and ease of use are related to user acceptance and the adoption of news recommender systems [11,34]. The work of Zheng et al. [35] shows that perceived usefulness and ease of use significantly influence user attitude and intentions to use. In regard to algorithmic media, perceived usefulness and convenient use can be quantified in ways that are more concrete and specific than abstract conceptualisations of value. Users may want to clearly understand what exactly is useful, how easy a system is to use and what facilitates them to act. In algorithm-driven recommendation services, users already know about the usefulness and ease of use of pertinent services. Users may pursue specific valued features through algorithm services, such as utility and ease of use [36]. As algorithms for content curation show people what is relevant to users, users consider AI acceptance in terms of how useful and how convenient AI services are when it comes to actual use.

2.3. Attitudes

Over the last several years, emotions have received growing attention in several AI-related fields, most prominently in human–robot interactions, where emotional receptiveness and expressivity are essential [37]. In recent algorithm, efforts have been made to understand the role of emotion in AI such as emotional AI [18]. For example, chatbots (i.e. chatting robots) have been used as interactive platforms to improve communication channels. With chatbots, marketers use emojis and emoticons to express a range of emotions in messaging. Relevant research has suggested that emotions play an important role in user interactions with AI [20]. The goal of AI is to ensure that users are not just satisfied, but are pleased to use with AI positive emotional valence. A strong positive emotion derived from the assessment processes assures users, providing them with satisfaction and intention to use the algorithm.

Recently, emotion has received great attention in AI (as exemplified emotion AI) as it plays a critical role in interactions with an algorithm, particularly in tandem with the rise of humanizing AI [1]. Incorporating emotion is key to this effort to create human-like AI because it is hoped that instantiation of emotion will ultimately lead to AIs that have ethical and moral codes, which then be able to develop mutual rapports, facilitating the use of AIs as human friends. Prior literature in technology acceptance model has confirmed shown that user-perceived usefulness and ease of use influence the attitude towards the system. When interacting with AI, many emotions make up attitudes, as attitude has a significant impact on human emotions. When users confirm the value of a system, their emotion towards the AI algorithm becomes positive. The relationship between usefulness/ease of use and emotion has been widely validated in various contexts, including algorithms (Figure 1) [11].

OPEN IN VIEWER

Figure 1.

Embodied cognitive process in human–AI interaction.

3.1. Data collection and sample

We collected a total of 395 individuals residing in the United States via Amazon Mechanical Turk in exchange for a cash reward (US$2). The sample was targeted to respondents who had prior experience with algorithm services (personalised recommendation, content suggestions, online news aggregation, etc.). To ensure the quality of survey results, multiple validation check questions were included to warrant the accuracy and trustworthiness of the responses. Of the collected responses, 19 incomplete responses with missing information were excluded, resulting in a total of 395 responses being used for data analysis. Among the participants, 51% were female. With regard to age, 36% were in their 30s, 40% were between the ages of 20 and 29 years, 13% were between the ages of 40 and 49 years, 8% were between the ages of 50 and 59 years and 2% were over 60 years.

Respondents were asked to surf, view and read auto-generated recommendations on algorithm-based sites for about 2 h (Figure 2). Respondents were instructed to surf personalised recommendation algorithm such as Netflix (hybrid recommender systems), Amazon (Amazon personalise) and Hulu (streaming algorithm). They were instructed to submit their data on user characteristic, behaviours, demographics and preferences on certain items (personalising user experience), and evaluate the recommendations on how they are customised/relevant to their needs. They were instructed to search and shop what they normally would for any preferred items, such as news, music, books, and movies. They were given information about the contexts of machine learning algorithms, FATE, and personalizations. The main experiment took about 30 minutes to complete and this was followed by the experimenter asking each participant questions from a self-reported survey questionnaire. For most participants, the experiment took about 30 minutes, but that time was variable based on how much time the participant spent interacting with the algorithms before starting the searching. Participants were allowed to repeat the experiment if they would like to.

OPEN IN VIEWER

Figure 2.

Experiment design.

3.2. Scales and measurements

The measurements used in this study were derived from published literature, which have been validated acceptable measurement reliability and validity in previous studies. We modified the measurements to suit the specific context of algorithms and AI. The final scales were seven-point Likert type, ranging from 1 (strongly disagree) to 7 (strongly agree). The FATE measurements were derived from [20] and [23]. The measurements of utility [32], [34], [35], ease of use [34], [35], [36], and satisfaction [15], [31], [33] were based on technology acceptance literature. The trust measurements were derived from [11],[38]. Finally, the emotion measurements were derived and modified from [20], [23].

4.1. Data validation

Although the scales used in this study have been used in previous research, some minor modifications to the wording of specific scale items were made. Hence, reliability for each of the scales used in this study was analyzed by calculating Cronbach’s alpha. The Cronbach’s alpha for the scales ranged from 0.760 to 0.905, suggesting that the items have high internal consistency (Table 2). A confirmatory factor analysis is conducted to confirm the structure of the scale using subsample. The result show that the items had acceptable factor loadings providing support for the 9-factor model. In order to evaluate validity (concurrent, convergent, and divergent), correlation analysis was conducted using Pearson’s correlation. Our correlation analysis shows no sign of collinearity. The latent variable discriminant validity was confirmed by calculating whether the square root of the average variance extracted (AVE) from each construct was above the correlations of the other latent variables. All AVE values were above 0.6, thus acceptable. The measurement model results show that the validity and reliability criteria were satisfied and that constructs for this measurement model are valid in testing the structural model and proposed hypotheses.

OPEN IN VIEWER

Table 1.

Reliability and validity.

Variables	Mean	Standard deviation	Cronbach’s alpha	AVE	Composite reliability
Transparency	4.61	1.041	0.854	0.763	0.906
	4.64	1.141
	4.42	1.066
Accountability	4.34	1.192	0.842	0.762	0.905
	4.60	1.213
	4.06	0.983
Fairness	4.04	1.257
	4.46	1.208	0.797	0.755	0.787
	4.07	1.065
Explainability	4.05	1.212	0.760	0.719	0.771
	4.04	1.266
	4.16	1.271
Satisfaction	4.38	1.119	0.885	0.817	0.930
	4.55	1.191
	4.41	1.299
Usefulness	4.19	1.063	0.805	0.714	0.882
	3.84	1.100
	4.10	1.048
Emotion	4.40	1.116	0.768	0.663	0.723
	4.09	1.285
	4.19	1.231
Trust	4.36	1.121	0.905	0.840	0.940
	4.23	1.057
	4.13	1.046
Ease of use	4.40	1.200	0.797	0.814	0.929
	4.32	1.142
	4.64	1.217

AVE: average variance extracted.

Measured with seven-point scales.

OPEN IN VIEWER

Table 2.

Model fit indices.

Fit statistics	Structural model	Suggested value
χ 2/(df)	1453/308 = 4.71	<5
p-value	0.000	<0.05
RMSEA	0.095	0.05 < x < 0.10
CFI	0.841	>0.90
NFI	0.908	>0.90
IFI	0.843	>0.80
TLI	0.905	>0.85
RFI	0.865	>0.80
AIC	810
Hoelter	105

RMSEA: root-mean-square error of approximation; CFI: comparative fit index; RFI: relative fit index NFI: normed fit index; IFI: incremental fit index; TLI: Tucker–Lewis index; AIC: Akaike information criterion.

In evaluating the model goodness of fit, we used widely recognized indices. From the results presented in Table 3, it was determined that the model goodness of fit was acceptable since all the indexes are within the recommended ranges. Overall, the data fit the model reasonably well. Hence, results show that there were insignificant errors in measuring the endogenous constructs in the model.

OPEN IN VIEWER

Table 3.

Path results.

Paths	Beta weights	SE	CR	p
H1. Transparency → trust	0.316	0.091	3.501***	0.00
H2. Fairness → trust	0.119	0.030	2.765	0.06
H3. Accountability → trust	0.284	0.067	4.924***	0.00
H4. Explainability → trust	0.381	0.111	2.493**	0.03
H4. Trust → utility	0.707	0.053	12.909***	0.00
H5. Trust → ease of use	0.765	0.062	10.817***	0.00
H6. Utility → emotion	0.150	0.053	3.329***	0.00
H7. Ease of use → emotion	0.114	0.069	2.202*	0.028
H8. Emotion → satisfaction	1.003	0.039	24.94***	0.00

SE: standard error; CR: critical ratio.

*

1.96: 95% (0.05); **2.58: 99% (0.01); ***3.29: 99.9% (0.001).

4.2. Structural model testing

The structural path analyses show the causal relationships between variables in the model are largely significant. (Figure 3 and Table 3). Except one (H2), all the path coefficients between the variables in the model were significant (p < 0.001). Even the rejected H2 can be considered marginally significant as p is 0.06. Trust is significantly influenced by FATE. These factors altogether account for 58.5% of trust variance. Performance expectancy values are greatly influenced trust. Satisfaction was significantly influenced by emotion. The model explained a significant portion of variances in each construct. The model explained 58.5% (p < 0.001) of the variance in ease of use, 50.0% (p < 0.001) of the variance in usefulness and 39% (p < 0.001) of variance in satisfaction. The strong paths imply a fundamental connection between trust and its antecedents. Based on prior studies on the relation of trust and performance, it is reasonable to estimate mediating effects of trust on performance expectancy. When users trust algorithms, they tend to allow more data to be collected, more data produce more accurate predictive analytics.

OPEN IN VIEWER

Figure 3.

Heuristic and systematic embodied cognitive process.

To test mediating effect, we employed a bootstrapping procedure, which is a non-parametric method depends on resampling multiple times with replacement. We selected 5,000 random samples of our data to generate our estimates. It is possible for mediators and suppressors to co-exist in models involving multiple intervening variables. Thus, bootstrapping tests of specific indirect effects would help describe the separate roles played by single intervening variables. We used an SPSS macro on testing multiple mediation models to test the mediation analyses. The standardised indirect effect table shows that exogenous latent constructs have partial mediation effects towards emotion through trust. There are partial mediations, which mean that trust has indirect effects on the relationships, which can be significantly reduced without trust, but the relationships still hold.

6.1. Contributions to research: Integrating user cognition into the algorithmic process framework

Our research contributes to ongoing debates in the academic literature on HAII, FATE, and algorithms in the context of AI. This study identifies the antecedents of user trust in AI and tests the heuristic role of those antecedents and emotion. These findings contribute to theoretical advances by proposing how algorithmic trust is created and what effects of trust there are in AI use, and how trust can be theorised, measured and analysed with reference to embodied AI. The significant relationships of performance expectancy with trust are one piece of evidences linking to embodied cognition. How users perceive algorithmic features and how they embody them are related to how much they trust and how they evaluate the algorithmic performance. Perceived notions of FATE are positively associated with usefulness and ease of use through trust mediation. In other words, users assess performance and quality of AI through a dual process: first, through FATE heuristics and second, via systematic processes through trust. Users process algorithm services both heuristically and systematically. Heuristic processing involves the use of simplifying decision assessment of FATE to quickly assess service quality. Systematic processing entails deliberative processing of usefulness and ease of use. Our dual process is distinct from the existing dual process, which underplayed heuristic processing by putting it as a peripheral route and systematic processing as a central route [39]. By highlighting users’ heuristic processing, our model is able to integrate diverse, including embodiment, emotion and context into a cognitive model as well as acknowledging important concepts such as FATE. Trust connects the two processes linking heuristic and systematic mechanism [3,40]. It is shown that the effect of algorithm quality on users’ attitudes is partially mediated by perceived trust. This trust link can be key to understanding algorithmic qualities, algorithm experiences and users’ interactions with AI. Certain algorithmic features provide users with cues for trust, and trust allows users to use algorithms with feelings of usefulness and efficacy. It can be inferred that trust formed through heuristic processing is more probably to have cognitive attributes that reflect FATE assessment, whereas that formed through systematic processing is more probably to have effects on performance evaluation due to reliance on established FATE cues.

Second, the HSP in the study advances the current user heuristic literature, specifically the users’ cognitive process debate, by identifying the role of trust and the underlying relationships [37,38,41]. Our models not only support the HSP’s classic argument that decision-making is largely influenced by heuristic cues [19], but also provide additional insights for the trust influence as a mediator between heuristics and systematic processes [3]. Previous research on HSP focuses on dichotomic and separate process of heuristic and systematics, neglecting how the two are related and intersect. Advancing from previous works, our results show that User cognition arises through a dynamic interaction between human and algorithms, and users have to make sense of algorithmic environment [17]. Many features of algorithmic cognition are shaped by processes of the heuristic-systematic evaluations of FATE. Algorithmic performance is brought about and enacted by the active interaction of users and algorithms. User’s cognition is embodied through heuristic-systematic processes when it is dependent upon features of the algorithmic attributes. Because the user cognition is embodied and arises out of an active processing and interacting with algorithms, algorithms are reflections of their creators, human. Algorithms and human are co-evolving, and they shape each other through heuristic-systematic mechanism.

The third contribution of this study is that our model shows more integrated perspectives of how users perceive AI qualities, how their trust is created and sustained, what cognitive affordances are realised and what behavioural results are derived from the processes. The findings of this study, particularly the heuristic–systematic process, will enable future studies to increase both the rigour of existing literature and the questions addressed in the area of HAII. As the findings imply, the functional features of algorithms are processed through users’ understanding and perceptions regarding perceived affordances, which are mediated by trust. Affordances thus influence the cognitive processes of quality, performance, emotion and satisfaction on the part of users [32]. We have detailed ways to conceptualize embodiment in an age of AIs, how enactivism relates to algorithms, and how extended cognition might be extended into the AI realm by underscoring the heuristic-systematic processes through which reciprocal and social interactions of users and algorithms are maintained. Future studies may examine theoretical justifications of the mediating effect of trust in the association between AI features and human emotion. Understanding the human emotion in AI would provide a clue on the development of user-centred interfaces for AI.

6.2. Contributions to practice: How to better design AI

Our research contributes to practical human-centered algorithm systems as well as the research of embodying algorithms and enactive AIs. To guide AI development, this study provides a heuristic–systematic approach to relate social concepts of trust and emotion with the technologies used in AI-based services and products. We conceive trust as discussed in the FATE framework and use a recently proposed mapping of FATE onto the qualities of an algorithm. These FATE frameworks provide practical guidelines for the development of user-centred or trust-based algorithm design guidelines. For the providers of AI or other similar algorithmic services, the implications of this study can be useful in designing AI interface and usability. As AI continues to transform the way we interact with technologies, how to warrant transparent interactions and fair algorithms while including explainability in the interface are important issues to address.

Our findings have practical implications for FATE in algorithms. Issues of transparency and explainability have been recent topics in AI, and users seek guarantees on such issues when using AI. Based on the FATE model, we can infer that trust is related to these issues as it plays a key role in developing user confidence and credibility. When users are assured of issues of FATE, users’ trust increases, and they are willing to provide more of their data to be collected and analysed. The more trust between users and AI, the more transparent processes can be put into practice. In turn, greater amounts of data enable AIs to produce more precise and accurate results tailored and individualised to user preferences and personal histories. Trust serves as a critical liaison to bridge between users and AI systems, enabling a positive feedback loops. The results of this study provide guidelines on how to actualise and integrate transparency and explainability issues with other factors – for example, how to collect user data and/or implicit feedback effectively while upholding users’ trust and emotion.

Our findings raise the need for addressing FATE when designing and developing AI systems and applications. An important implication of this is that building consensus and achieving collaboration across key stakeholders (such as clients, users and society) are a prerequisite for successful adoption of AI in practice. Another key implication is that explainable AI should provide answers on how to address questions like why did the AI make a specific recommendation and why did not the AI do something else? One practical implication to gain credibility in AI systems is to use algorithms that are inherently explainable and interpretable. For instance, basic elements of algorithms such as logic classifiers, path trees and other algorithm information that have certain levels of transparency and traceability in their decision-making can provide the visibility needed for AI.

Another key message to emerge from this study is that industries should address algorithm experience in AI. Designing algorithm systems and features poses new challenges for user experience practitioners. To understand user attitudes and algorithm behaviours, research must consider algorithm quality, user heuristics and recognised value (Bolin and Schwarz) [42]. In particular, the insight from the user heuristics can be used for designing heuristic algorithm. Developing effective user-centred algorithm services requires an understanding of users’ cognitive processes together with the ability to reflect these processes in an algorithm design [11]. User perceptions and psychological state of mind are essential in rationalising how and why people perceive and feel what they do regarding the issues surrounding AI, as well as how they accept and experience AI services [42].