”Because AI is 100% right and safe”: User Attitudes and Sources of AI Authority in India

Understanding the role and measurement of trust continues to be an active topic of research for several decades, across various disciplines such as interpersonal relationships [104], organizational studies [89], consumer relations [87], and technology [61]. Several definitions and frameworks have emerged from the widespread and varied interest in trust [58, 83, 104]. Lee and See [73] define trust in automation as “an attitude that an agent will help achieve an individual's goals in a situation characterized by uncertainty and vulnerability.’’, and used the framework by Fishbein and Ajzen [42] to characterize trust as an attitude of the individual, rather than an intention or a behavior. Beyond defining what it means to trust, researchers have devised frameworks for the factors that contribute to trust. Mayer, Davis and Schoorman [83] present an integrative model of trustworthiness with three components: ability, integrity, and benevolence. Ability refers to the competencies that are required for a trustee to have influence in a specific domain. Integrity relates to following a set of principles that are acceptable to the trustor. Benevolence refers to the extent to which the intentions of the trustee align with the trustor. We utilize these frameworks in the following sections to reflect on our results and present implications.

Extensive prior research has explored constructs to understand the emergent dynamics between human and algorithmic interactions (e.g., AI utopia [111], algorithmic appreciation [79], algorithm aversion [32], described below). Santow [111] described AI utopia as an imagined future with improvements in every aspect of life. In this research, we draw inspiration from Lustig and Nardi's work [81] that introduced the concept of algorithmic authority grounded in their mixed-methods research of the Bitcoin community. Users perceived Bitcoin as an ‘apolitical currency’, as more reliable and trustworthy than banks or governments, and the algorithmic was seen as a form of explicit resistance against traditional financial institutions. The concept of trust is of particular importance to algorithmic authority. Bitcoin users place trust in the algorithmic authority: trust is an attitude that mediates relationships between humans and the algorithmic authority. Lustig and Nardi discuss the ways in which Bitcoin was perceived to possess a predictability which led to greater trust in the algorithmic authority. However, the algorithmic authority of Bitcoin was mediated by human judgment, the human oversight even if the algorithmic was perceived as ‘self-contained’. In this research, we extend the work of Lustig and Nardi to examine the perceptions and acceptance of AI-based applications among Indian users, and the factors which contribute to these perceptions.

The communities of HCI and CSCW have investigated the perceptions of algorithmic decision-making in comparison with human decisions (e.g., [13, 19]) with conflicting results. Most studies find that individuals are more likely to trust human decisions over algorithmic decisions for tasks that require human-like skills (e.g., subjective judgment, emotion) (e.g., [80]). For instance, Lee et al. [74] explore perceptions of fairness and trust for four managerial decisions and find that human decisions are considered fairer and more trustworthy than algorithmic decisions for tasks requiring human skills (hiring and work evaluation). Dietvorst et al. [32] found that human forecasters are resistant to using algorithms, especially after seeing them perform (algorithm aversion). Logg, Minson, and Moore [79], however, present contradicting results: they find that people rely on algorithmic advice over human advice for forecasting (algorithmic appreciation), but this effect waned when they had to choose between their own judgment vs. an algorithm. Others in the community have studied human perceptions of AI through a wide-ranging set of methods [39, 63]. Cave, Couglan and Dihal [20] conducted a survey to examine public responses to AI in the UK, and find that 25% participants associated AI with robots. Prior research has also investigated worker perceptions of technology in organizations [76, 93]. Höddinghaus et al. [48] compared workers’ trustworthiness perceptions of automated vs. human leadership. Their results indicate workers perceived automated leadership agents as being higher on integrity, but lower on benevolence than human leadership agents. Nagtegaal [88] demonstrate the role of task quantifiability/complexity in perceptions of procedural justice: decision automation was perceived as less fair in tasks with higher complexity (and lower quantifiability [71]).

A related area of research examines the factors influencing perceptions of trust, algorithmic fairness and justice (e.g., [75, 136]). People's perception of fairness can be complicated and nuanced, and go beyond formal algorithmic constraints. Araujo et al. [7] explored the role of individual characteristics in perceptions about automated decision-making, and report that people with more knowledge about AI are more optimistic about algorithmic decision-making. Ashoori and Weisz [10] report that model interpretability and transparency about training/testing datasets played a critical role in establishing trust. However, recent research by Wang et al. [130] argues that for non-expert stakeholders, the effect of development procedures on perceived fairness is much smaller as compared to algorithmic outcomes and biases.

Relatively less work has been done to uncover the perceptions and acceptance of algorithmic systems in India (e.g., [52, 84]). Okolo et al. [92] studied the perceptions of community health workers (CHWs) in rural India towards AI, and find that CHWs considered AI applications trustworthy with expertise greater than their own. Thakkar et al. [122] examined the perceptions and practices of vocational workers towards automated systems. We extend this body of work by focusing on the ways in which perceptions and intentions are shaped through factors beyond system design. We deliberately focus on ‘AI systems’ as opposed to ‘algorithmic systems’, given the specific associations that people make through terminologies and conceptual metaphors [41].

Algorithmic folk theories are conceptions to explain algorithmic behavior and outcomes that are informally shared among similar members of a culture, but are not necessarily accurate [28]. Prior CSCW and HCI literature has explored people's understanding of social media algorithms [31, 45], and find that users develop these folk models through their experience with technologies and social interactions. Their results indicate that a majority of people have high-level folk theories about how these algorithms work. Rader and Gray [102] studied user beliefs about news feed curation on Facebook, the algorithmic behaviors that users notice and the behaviors to which they respond. Taking it further, Eslami et al. [36] uncovered and codified the folk theories of the Facebook News Feed curation. Most people become aware of algorithmic involvement through unexpected behavior, which causes them to adjust their future actions.

Prior research has also made headway in examining user awareness of algorithms, and the role of that awareness in their interactions. Bucher [17] developed the concept of the algorithmic imaginary as the “ways of thinking about what algorithms are, what they should be and how they function [...] plays a generative role in molding the Facebook algorithm itself.’’ Hargittai et al. [45] study people's algorithmic skills, and report on the methodological challenges. Algorithmic skills involve the ability to identify that a dynamic system is in place (awareness), and reflecting on how those systems work (understanding) [45]. Another recent thread of work focuses on the ways in which users adapt behaviors around algorithms [53, 55]. Eslami et al. investigate the ways in which users perceive and manage bias in online rating platforms [38], and propose the concept of ‘bias aware design’ to harness the power of users to reveal algorithmic biases.

Most of this research has been conducted with US or EU-based respondents, largely with social media, and/or MTurk users who are known to be tech-savvier than the average Internet user [46]. US and EU tend to have critical media discourse around the use of algorithmic systems [64, 119], scrutiny about algorithmic biases from activists and civil society [3], and existing/emerging laws and regulations around the use of AI [25, 54, 105]. This is in contrast with India, which currently lacks substantial research on responsible AI development [107]. Algorithmic folk theories are shaped through the above elements, which differ greatly between the contexts previously studied and our study site. In addition, our goal is not to provide explicit folk theories, instead, we connect folk understandings with users’ intentions to engage with, and act upon AI decisions or recommendations.

Prior work in HCI and ICTD has studied the discourses around technology in India, and how it has been frequently tied with notions of development [95]. The last two decades, in particular, have been crucial in shaping technology as the means for prosperity in India. Digital technologies are viewed as a vehicle for progress, as a solution to societal problems in developing countries [16]. The following is an excerpt from Prime Minister's speech in 2018 [86]: “Can Artificial Intelligence help us detect serious health conditions before they manifest physically? Can Artificial Intelligence help our farmers make the right decisions regarding weather, crop and sowing cycle? Friends, our Government is of the firm belief, that we can use this power of twenty-first century technology [AI] to eradicate poverty and disease. In doing so, we can bring prosperity to our poor and under-privileged sections. We are committed to achieving this vision.”

The Indian government's vision for technological development manifested through two major initiatives over the recent years. First, the introduction of Aadhaar— a biometric identification system for 1.3 billion citizens, which was legitimized through a promise for poverty reduction and financial inclusion [120]. Secondly BHIM, an application for digital payments introduced soon after demonetization as the future for cashless payments [97]. Technology played a symbolic and functional role in enabling the ‘leapfrogging’ into the modern era. Bozarth and Pal [14] examine the social media (Twitter) discourse, and find that politicians have an inclination to discuss technology in connection with development as part of their political messaging. Public discourse around technology and AI has been hyper-optimistic from the general public, the tech industry, and the government [96], with several deployments underway [90]. Sambasivan et al. [107] describe how due to the aspirational role played by AI, high risk AI solutions are often adopted too quickly in India, including in predictive policing and facial recognition, without sufficient checks and balances. Our research aligns with this perspective on effects of techno-optimism and extends this body of literature by presenting findings of the ways in which this aspirational value is shaping AI authority among Indian technology users.

We conducted interviews with 32 adult Internet users based in different regions within India to understand perceptions of AI and their acceptance of AI-based decisions. Participants were located in various tier 1 and tier 2 cities of India, belonging to diverse age groups and occupations. Our sample also consisted of a mix of internet experience, with 16 nascent internet users that first accessed the internet only in the last 2 years, and the remaining 16 more ‘experienced’ that have been online for more than 2 years. We interviewed 17 male and 15 female participants. Refer to table 1 for details on participant demographics.

Participant recruitment and moderation. We recruited participants through a combination of a UX Research database internal to our organization, and a market research company Dowell. We conducted ’screener’ conversations with potential participants to ask if “[they] had heard of the words AI or Artificial Intelligence?”, and if yes, we asked them to describe what it means. We selected people who were aware of these terminologies and did not filter any participant based on their descriptions of AI. The sessions were conducted using video conferencing, due to COVID-19 travel limitations, and each interview lasted 75-90 minutes. We conducted interviews in three languages (Hindi, Tamil and English) by allowing participants to select the language with which they are most comfortable. The interview notes were recorded through field notes and video recording, and transcribed within 24 hours of each interview by the corresponding moderator. Each participant received a thank-you gift in the form of a gift card worth 27 USD or 2000 INR.

Interview structure. In line with our goal, each interview had structured sub-sections beginning with the participants’ general perceptions and understanding of AI. We asked participants to share “in [their] opinion, what do [they] think Artificial Intelligence means?”, and, “overall, do [they] think AI is doing more good or more harm for us? Why? In what ways?” To explore whether participants are able to recognize AI, they were then shown a series of images each containing commonly used AI applications in India across domains [131], and requested to identify if and how those products use AI or not. In the main part of the interview, we used a scenario-based approach (details in the following section) to probe experiences, attitudes and intentions to act. For each scenario, we began by presenting a scenario description accompanied with visuals, and then invited participants to share their initial reactions. We explicitly clarified that the applications in the discussed scenarios are built using AI. Following questions aligned with our goal of understanding participants’ acceptance of AI technologies, how it relates to their conceptions of AI, and the factors that influence these perceptions. We asked questions around their beliefs and intentions (e.g., if AI made a decision on your loan application, would you believe it to be correct?), their preferences for human vs. AI decision-maker (e.g., What are the differences between a human making a decision on your loan application and an AI making that decision?), the kinds of information they would like to know about the system, and the level of control they believed the AI application should have in the given scenario. Within each scenario, we also elicited reactions for a negative outcome for the individual (e.g., loan rejection, wrong medical diagnosis). The final part of the interview focused on reflecting on qualities of an ideal, responsible AI, the kinds of tasks that an AI system should or should not do.

Scenario selection. We draw inspiration from prior research on examining various perceptions of trust [10, 74], fairness [7, 136], justice [13] and explanations [33] in using scenarios³ to investigate perceptions. We used four scenarios in the study: (1) loan assessment, (2) financial advisor, (3) medical diagnosis, and (4) hospital finder (details in appendix A). The scenarios were a part of a vignette experiment with 2 (decision-support vs. decision-making) x 2 (health vs. finance) within-subjects design. We randomized the four scenarios across these participants using Latin Square Design [15] to control for ordering effects, if any.

Healthcare and finance are a visible, explicit policy focus (see NITI Aayog Strategy of 2018 [1], and the latest responsible AI strategy of India [2]), and key areas for private sector deployments [44, 100]. Participants were the subject of the decision-making in these scenarios, and thus, we selected relatively common interaction use cases [13] within which most participants could situate themselves (at present or in the future), even if they had not previously encountered those specific algorithmic systems. We developed the scenarios to represent decision-making with different kinds of agency to affect decision-making (e.g., decision support where the AI makes a recommendation and the user is the deciding agent, and decision-making where the AI makes a decision about or on the behalf of the user). Each scenario was based on real-world examples of AI systems. We used hypothetical scenarios (instead of commonly-used consumer applications) to control for differences in actual prior experiences across participants, so that these effects do not manipulate the results of our study.

Analysis and coding. After transcribing, we translated all the Hindi and Tamil interviews into English, our primary language of analysis. We followed the qualitative data analysis approach by Thomas [123] to conduct multiple rounds of coding at the sentence and paragraph level. We began by open coding instantiations of perceptions, assumptions and expectations of/around AI. Two members of the research team independently read all units multiple times, and identified categories (unit of analysis) together with a description and examples of each category, until a saturation point was reached. Our upper level categories were guided by the evaluation aims, comprising (1) general perceptions about AI, (2) perceived harms and benefits, (3) sources of AI perceptions, (4) acceptability of AI interventions, (5) willingness to give authority, and (6) human vs. AI vs. human and AI. These codes were discussed and iteratively refined through meeting, diverging and synthesizing. We consolidated our codes into two top-level categories on defining AI authority, and the sources of AI authority.

Research ethics and anonymization. During recruitment, participants were informed of the purpose of the study, the question categories, and researcher affiliations. Participants signed informed consent documents acknowledging their awareness of the study purpose and researcher affiliation before the interview. At the beginning of each interview, the moderator additionally obtained verbal informed consent. We stored all data in a private Google Drive folder, with access limited to the research team. To protect participant identities, we deleted all personally identifiable information in research files. We redact all identifiable details when quoting participants. Members with research backgrounds in HCI, AI, psychology, and interaction design constitute our research team. Authors located in India moderated the interviews. All researchers were involved in the research framing, data analysis, and synthesis.

The interviews revealed that AI authority manifested, in part, as acceptance of AI-based decisions. We further confirm this observation by examining the acceptability of AI outcomes through a survey with a larger audience. Our qualitative research informed the operationalization process, and we use acceptability as the construct for measurement. AI authority is a multi-dimensional concept of which a likelihood to accept AI decisions is only one element. The survey was completed online and received a total of 459 respondents. We used a scenario-based approach for the survey, similar to prior research in this space [7, 10]. The scenarios were a mix of high- and low-stakes situations, which we describe below.

Sample. Members of an online market research panel, Cint were invited to participate in our survey. All participants were based in India, and above 18 years of age (minimum age of consent). We began the survey by eliciting informed consent from respondents. The inclusion criteria for our survey was similar to the interviews. The screener question asked, “how much do you know about Artificial Intelligence?’’ We included respondents that had at least heard of the words Artificial Intelligence. Those who answered, “Never heard of AI’’ were not asked any further questions. After the screening and attention check questions, we were left with 459 respondents that completed the entire survey. Each respondent received $1.70 (INR 150) as compensation for their participation. We describe their demographic details in table 2.

Questionnaire. The survey questionnaire was implemented in Qualtrics. Our questionnaire consisted of 14 questions in total, with a mix of multiple-choice (11) and open-text questions (3). It was divided in four sections: understanding of AI (2), AI acceptability in general (1), scenario-specific AI acceptability (7) and demographics (4). On average, it took respondents 6.5 minutes to complete the online survey. After completing the screener, the first question focused on their understanding of AI. The open-ended question asked, “how would you explain Artificial Intelligence to a friend?’’ We then requested respondents to provide up to three examples of how AI is used today. The seven questions about measuring AI acceptability ranged on a Likert scale of 1 being “Not at all accepting’’ to 5 being “Extremely accepting’’. The questions were worded with the respondents as the subject of decision-making. Each question on AI acceptability was phrased with a variation of, “How accepting would you be of a decision made by an AI for you?”, with the scenarios embedded at the end of the question and highlighted for readability.

After understanding their general willingness to rely on AI for decision-making, we used six scenarios to understand the level of AI acceptability in varying situations/domains: three high-stakes, and three low-stakes. The six scenarios are (1) medical diagnosis, (2) loan approval, (3) hiring decision, (4) song recommendation, (5) route recommendation, and (6) auto rickshaw pricing. We use a within-subjects design, where all respondents were shown all 6 scenarios, presented independent of each other in a randomized fashion to account for any ordering effects. In addition to the Likert scale items, participants were asked to describe why they were willing to accept decisions made by an AI, if at all. Finally, we also included demographic questions (optional to answer) about respondent's age, gender, education, and internet exposure.

Analysis. We performed several kinds of analysis to determine the prevalence of AI acceptability across various scenarios. We present two key measures from our analysis in table 2(a), (1) the mean acceptability in our sample, (2) ‘significant acceptability’ which represents the percentage of respondents that selected midpoint (>=3 i.e., “Moderately accepting”) or above on the Likert scale for AI acceptability. The questions on levels of AI acceptability were analyzed by calculating percentages and cross-tabulated to view demographic-specific percentages for certain questions. We performed an independent samples t-test to compare AI acceptability between males and females. We conducted a one-way ANOVA to study the influence of age and internet exposure on acceptability. Finally, we performed a qualitative analysis to the open-ended questions using an approach similar to the interviews (see section 3.1). We conducted multiple rounds of coding at the response level in conjunction with participants’ survey ratings to surface high-level themes. We include direct quotes from our survey respondents in the Findings with the prefix ‘S#’, to differentiate from our interview participants prefixed ‘P#’.

We draw upon Weber's definition of authority, as a form of power whose use is legitimized by those on whom the power is exercised [132], and extend the work of Lustig and Nardi [81] to define AI authority as the legitimized power of AI to influence human actions, without requiring adequate evidence about the capabilities of a given system. AI was seen as reliable and infallible, and thus considered worthy of authority. Participants were willing to trust and accept AI-based outcomes, and perceived AI as a better alternative to the human institutions with which they currently interacted. Thus, AI authority emerged as a form of ‘voluntary compliance’ [127] that was seen as justified by users.

The conceptual metaphors and terminologies can play an important role in our interactions with technology [41]. We broaden the conceptualization of authority from ’algorithmic’ [81] to ’Artificial Intelligence’ for the unique disposition of ‘AI’ to gain authority due to the hype, optimism and narratives surrounding its use. The techno-optimistic narratives, among other sources, played a key role in generating legitimacy for AI authority among the participants in our study. The acceptance of AI decisions was often due to reasons beyond system design characteristics and performance. Participants perceived AI as a panacea— a tool to prevent discrimination and injustice that they had previously encountered in interactions with traditional institutions, as opposed to the ways in which Bitcoin users in Lustig and Nardi's research [81] viewed the algorithm as a technology of resistance.

We find that over 79% survey respondents reported significantly high acceptance (i.e., selected midpoint (>=3 “Moderately accepting”) or above) of decisions/recommendations made by an AI system, averaged across the six scenarios. A willingness to accept AI-based outcomes for respondents in both studies was often guided by the consequences associated with a decision. On average, 84.1% survey respondents were willing to accept decisions in ’low-stakes’ AI scenarios, whereas only 73.8% respondents were willing to accept decisions in high-stakes AI scenarios (see table 3). We combined the three high-stakes scenarios (Cronbach's α = 0.83), and three low-stakes scenarios (Cronbach's α = 0.77) into two distinct scales to examine the effects of personal characteristics on acceptability. Gender was relevant for AI acceptability across the three measurements, with females seeing AI systems as significantly more acceptable than males. Finally, we found no significant association between internet exposure and AI acceptability. We present the responses across the acceptability scale for each scenario in Appendix A.

Interview participants also indicated a willingness to take high-risk actions based on outcomes. Several participants noted an acceptance of the medical diagnosis by AI, with an openness to undergo medical treatment, for instance, “the AI has given me a diagnosis. I will go to the hospital and get the necessary treatment. [...] I have more confidence on an AI instead of a doctor.” (P21) Participants were willing to accept loan assessments, and in fact, preferred that AI evaluated their application over a human loan officer. For the decision-support scenarios, most participants reported an inclination to follow AI recommendations (e.g., visit a hospital, budget income, invest in financial schemes), often mediated by their own judgment. As P11 expressed, “I would definitely invest in businesses suggested by an AI, because I am 100% sure that it would be the right recommendation.’’ During each interview, we explored situations in which users might receive opposite recommendations from a human expert vs. an AI. Several participants reported not only an intention to act on AI outcomes, but also an inclination to follow AI-based recommendations over those by a doctor or financial advisor. As P7 described, “I would definitely go with AI if there is a contradiction. There is a clear logic behind an AI.” Overall, participants demonstrated a tendency to accept AI decisions as correct, and believed AI was worthy of authority unless they had evidence to the contrary.

Participants’ trust in the AI applications of the scenarios was underpinned by their beliefs about AI's high competence and benevolence, where different levels of trust could result in diverging intentions or behaviors. AI was considered reliable, that it is high-performing and capable of making correct decisions or recommendations. Participants also perceived AI as infallible and emotionless— that systems are based on facts, logic, rules and conditions or parameters, for example, a loan assessment AI looks at whether an individual's salary is above a certain threshold to determine whether they are eligible for a loan.

Participants in our study ascribed a positive or neutral intent to AI in their responses (more in 4.2.1), and displayed a motive-based trust in the systems presented in the scenarios, which was also observed by Kramer and Tyler [67]. The social conception of trust suggests that people are influenced by their attributions of the motives of authorities [67]— “attributions of positive intent lead members to trust the authority, and accept the decision onto themselves.” Next, we present the four user attitudes of faith, forgiveness, self-blame, and gratitude, that accompanied the authority of AI, and the ways in which these attitudes affect people's interactions with AI applications. The following attitudes towards AI made participants more likely to accept AI-based outcomes, even in situations where they receive a wrong or unfavorable outcome.

In this section, we present four sources through which AI acquired authority. A propensity to confer authority to AI was influenced by: (1) interactions with ineffectual institutions, (2) techno-optimistic narratives and misleading terminologies around AI, (3) users’ prior interactions with applications, and (4) unavailability of alternative systems. AI authority was often derived through one or all of the above sources, often a combination (e.g., a good experience with a different healthcare mobile application, and the unavailability of a medical doctor in their village). We emphasize that AI authority could be misplaced: if AI derives legitimacy and influence over people's actions through sources which lie outside the AI system and are unindicative of the efficacy of the given system, then an intention to act upon AI decisions introduces a potential for individual harm. These sources of AI authority are extraneous to the system— they provide no evidence that the system will be effective, and moreover, unbiased to the user. In reality, an AI system could be dysfunctional (poor performance and/or unfair), and still gain authority among users.

Participants in our study indicated a propensity to actively rely on AI for decision-making or recommendations. This might seem desirable or align with business goals, indeed, many organizations utilize consumer trust as a metric for success [125, 129]. Greater trust might seem to indicate a well-performing product for users. However, our findings indicate that users’ experiences with alternative, human systems could easily confound measurements of trust. Is it possible that a high trust is simply an indication that users place authority in AI instead of existing human institutions? For instance, an application might be dysfunctional, unsafe, or unfair to certain users, but they would still rely on it because AI is perceived as a better alternative. Overall, high perceived trust might not be an indicator about the system performance. Efforts to reduce acceptance of AI outcomes might seem antithetical to business goals: regardless, if authority is well-calibrated over time, then it might mitigate harm, retain users, and increase overall satisfaction with the application, leading to success. Overall, unwarranted authority might initially seem desirable, but it can negatively harm product experience in the long-term, as people continually receive outcomes that do not align with their expectations.

We emphasize the need to calibrate authority towards an appropriate level aligned with the actual trustworthiness of AI [51, 124], instead of a trust built through the proxy, confounding sources that we document in our findings. The gratitude shared by our participants was not isolated to the applications that they were currently using (e.g., social media, navigation, voice assistants), but often extended to all of AI systems, including those in our scenarios. Designers can consider introducing features to communicate the actual capabilities of the system while optimizing for understanding [5]. One can set the right expectations from early-use about the system's capabilities. In particular, designers can consider making the limitations of the system explicit before and during early use, by describing the ways in which a system operates and arrives at a decision, offering examples of situations in which the system is likely to provide unreliable results (see the PAIR Guidebook [94], and the Microsoft Human-AI Toolkit [4]). More research is needed to discover the nuances of the ways in which users might adjust these components of trustworthiness with continued interactions with a system. Future work can consider designing alternative metrics for measuring success, beyond user trust in a product or feature, that explicitly takes into account the contextual factors and their prior experiences which contribute to acceptance and AI authority.

Individuals’ beliefs and understanding about AI was a major factor that contributed to AI authority. Participants conceptualized AI systems as rule-based, as embedded with clear, specifiable conditions. Users’ understanding of the ways in which AI works heavily deviates from the non-deterministic, deep learning models which are primarily used in current product experiences. Indeed, the communities of HCI and CSCW have long been interested in understanding the ways in which people perceive algorithmic systems (particularly social media applications) [31, 37, 62]. We find that AI is not understood in a standardized manner across cultures, contexts, internet exposures and age groups. In fact, respondents do not come with a ‘common denominator’, a shared understanding about AI systems when researchers are measuring various aspects of AI perceptions. AI carries a qualitatively different meaning, especially owing to the contextual narratives surrounding its use.

We argue for embracing the variability in research participants’ perceptions about AI which heavily influence their intentions and behaviors around algorithmic systems. Perhaps, instead of trying to create a shared meaning of AI or algorithmic systems, could we elicit user conceptions as a basis to guide our analysis of trust, fairness or usefulness? How might we foreground these conceptions or use them to contextualize our research findings? Comparisons of human vs. algorithmic decision-making must be foregrounded in participants’ beliefs about AI's functioning, which in turn influences their intentions and behaviors. Creating a standardized understanding across users (e.g., by defining AI for the participant) might not reflect real-world situations, or effectively erode/replace users’ own beliefs. In our study, several participants suspected the presence of AI in certain applications, but their reason for believing that those applications are created using AI were frequently distorted. For instance, many people alluded that an application which requires Internet or GPS is AI-infused. Overall, people might be able to accurately indicate which systems contain AI by conflating non-AI technologies with AI. AI literacy tests may be unable to reveal users’ actual competencies or understandings about AI. Combining quantitative tests for AI literacy or awareness with a qualitative approach could probe these beliefs and expectations in a holistic way.

Lateral comparisons of the various constructs of AI perceptions (e.g., acceptability, trust, fairness) across countries must be conducted with extreme caution. Prior research documents the various kinds of survey biases (e.g., acquiescence bias [133], social desirability [121]). People in different cultures have different response styles. Subjective Likert scales are often compromised due to response artifacts [91] or the reference group effect; “the more cultures differ in their norms (i.e., the more cultures are really different on the dimension), the more the cultural comparisons are confounded.’’ [47]. A straightforward comparison of beliefs and attitudes across cultures with widely differing norms might lead to measurement errors, and thus requires a careful wording of questionnaire design that proactively accounts for variations in response styles and regional differences. For instance, in the EU, perceptions might be very different considering the attention to GDPR [126], and the resulting awareness about algorithmic systems and its biases [60]. This applies to interpretations as well: there is a need to cautiously interpret pre-reported data from different countries, especially if there is a chance that the contextual factors might be lending authority to an AI.

Though the research on trust and fairness perceptions offers mixed results (often domain- and task-dependent), several prior studies find that respondents in the EU, US indicate a tendency to trust humans more than algorithmic systems [19, 32, 70], especially for tasks requiring human skills [74] (with some exceptions such as Logg et al. [79] and Lee and Rich [77]). In our technological context, we find a case of acceptance of AI outcomes with a confidence greater than a system might deserve. Misplaced authority is consequential: AI systems have potential to cause harm through incorrect or biased outcomes, especially because users might not actively seek out information about the capabilities of the given system. The effects of misplaced authority are exacerbated and far-reaching in high-stakes scenarios (e.g., hiring, finance, social benefits). Several participants considered a lack of human involvement in making the decision as desirable. Even the use of general-purpose products in critical situations (e.g., virtual assistants for job interview reminders) could lead to adverse outcomes for users.

Mayer, Davis, and Schoorman [83] present an integrative model of trustworthiness with three components: ability, integrity, and benevolence. We use this framework to reflect on our findings and present a path forward. Prior work (based in US or EU) reports that users rate technology companies (and products) with low trust (benevolence) [11, 85], but high ability. Therefore, when a technology system goes wrong, it is seen as a benevolence issue. Our results suggest that both—ability and benevolence—components of AI trustworthiness are viewed as high in India. Users are willing to give authority to AI because they have faith in the competence and benevolence of AI systems. When a system makes a mistake, neither its capabilities (ability) nor its intent (benevolence) were readily questioned. This resulted in self-blame or placing blame on other actors, without a recognition that AI could have made that error. Especially for the loan assessment scenario, many participants in our study believed that either they deserved an unfavorable outcome or could be their own fault. The self-blame was exacerbated when participants indicated a faith in the capabilities of AI. Low or non-contestation of AI outcomes has potential to cause individual harm if people accept decisions which might be wrong. In addition, users reporting decisions or behaviors to the system represents important opportunities for model feedback. This scope for mitigating harm and improving models is lost if users believe the outcomes to be correct, with a possibility of causing harm to other users.

There is a well-established body of work on explainable AI [9, 49], and how explanations impact user trust and reliance [98, 118]. The focus of this research is to explore when to explain (high vs. low stakes, e.g., [18]), what to explain (global vs. local [40, 69]), and how to explain [138]. To make the system more transparent, there's increasing interest within the HCI, XAI communities to find human-centered approaches for explainability [78]. There are broadly three levels of building competencies on AI use: (1) outcome/in-the-moment explanation of a certain outcome, (2) an understanding of how a given AI application works, (3) general, accessible education about what AI is, how it works, and what are its strengths and limitations. Most explainability approaches have been context-agnostic, however, there needs to be particular attention to emerging internet users that often have less familiarity with technical jargon, often coupled with lower literacy [92]. Additionally, in-the-moment explanations are not always feasible due to legal or usability constraints. Consider a credit card approval application. Designers might be unable to ‘explain’ the reason behind a particular decision due to anticipated legal issues. Even without legal constraints, explanations are most useful when actionable [56]. A credit card rejection for an applicant with AI authority might mean that they believe that the system made the correct decision. As a result, they might not seek alternative platforms when it could easily be the case that the AI made an error or gave a biased outcome. The onus is on the designers/developers to add safeguards, ensure that users do not share a Utopian view of the systems with which they interact [111], and acknowledge the range of system errors that are possible. Another approach to consider is to leverage existing capacities [134] by educating the user about how the system works in general (e.g., [43]) that could be a valuable starting point for users to calibrate their responses based off their general understanding of how an application provided an outcome.

The narratives shared by our respondents were often too fantastical and polarized. They described science fiction narratives (e.g., movies like Terminator), and several respondents exclusively associated AI with robot applications. Participants dismissed the more pessimistic narratives (e.g., of killer robots, see also [20]) which were hard to imagine or relate with. On the other hand, the optimistic narratives were entertained and adopted. AI was perceived as an innocuous, neutral tool to ‘make [their] life easier’. Our work reinforces the sentiment shared by Cave et al. [21] that “narratives might influence the development and adoption of AI technologies.” Then how might we develop alternative narratives about the ways in which AI technologies are currently used? An important emerging research area is to create and disseminate AI narratives that align with contemporary issues of algorithmic bias. Realistic accounts could gradually help people calibrate their authority by painting a less skewed, lopsided picture of the capabilities of AI.

People generally have a tendency to engage with mainstream media (articles, movies), and our results indicate that the polarized narratives to which people were exposed greatly shaped their acceptance and the authority of AI technologies. Overall, an optimistic portrayal of AI, especially through state-sponsored initiatives, could result in less-than-rigorous technologies making their way into usage by the general public, causing irreversible individual and societal harm. Several participants actively acknowledged the human involvement in AI development, however, we note a faith in the intentions of AI applications— that these systems are built to support users and improve current workflows, with neutral/good intentions, partly owing to the narratives about the benefits of emerging technologies.Several state-sponsored bodies are pushing for AI curriculum in schools and dedicated programs in universities [30, 106], strategy documents are reporting credible information [2, 114], however, it is equally important to disseminate these to a broader audience. For instance, “The Social Dilemma” [68] is a film that gained notoriety in the US, which documents the dangers of AI, specifically, social networking. Mainstream media modalities that educate people about the complexities and contextual concerns about AI, but more importantly initiate dialogue about these topics, might play a crucial role in calibrating AI authority.

Traditional auditing methods (e.g., involving researchers, activists) [103, 110] can be inadequate in surfacing harmful behaviors when the auditors lack the cultural backgrounds or lived experiences to recognize if something is problematic [113]. Shen, Devos et al. [115] propose the concept of everyday auditing through which users detect and interrogate problematic biases/behaviors through their everyday interactions with algorithmic systems. However, auditing can be premised on informed users with the capacities, and even more, the intentions to recognize and report bias. What if users are not likely to seek out instances of biased behaviors? What if, beyond that, users are defensive about AI systems, or unquestioning of its abilities to provide good results? Eslami et al. [38] report that users were able to notice bias in the rating platforms. User auditing approaches predicated on skeptical users might not generalize well to contexts that demonstrate techno-optimism and faith towards AI and its capabilities. People with high acceptability of AI might not see the shortcomings of using these systems. Sambasivan et al. [107] extensively report the ways in which a straightforward porting of responsible AI tenets can be inadequate and often harmful. How might we involve users with optimistic views about AI into algorithmic audits? Certainly, their perspectives would be valuable contributions in surfacing biased outcomes. Empowering users to interrogate these systems might be an approach to calibrate AI authority towards an appropriate level. Future work might investigate how we could leverage existing capacities in users that have high confidence in AI to acknowledge bias. This could benefit platforms in two ways: visibilising bias and mitigating harm, but also through building alternative, realistic narratives about AI that are better aligned with the capabilities of the system.