Broadly speaking, artificial intelligence (AI) is any type of technology that automates processes to solve problems that are usually associated with human intellectual capabilities (Ertel, 2017). More specifically, AI aims to solve problems and achieve goals with limited or no human supervision. A closely related term is machine learning (ML). Originally coined by Samuel (1959), ML can be defined as a collection of algorithms (mainly statistical and mathematical) to build computers capable of learning through experience (see Jordan and Mitchell, 2015). While the terms AI and ML are often used interchangeably, ML may be considered a more appropriate term than AI. Stereotypically, AI tends to be associated with rather unrealistic narratives depicting agents capable of human behavior (see Cave et al., 2018), and such examples are not yet feasible (also known as general AI). ML refers to algorithms designed to perform specific tasks in an automated way (also known as narrow AI) (Dawson et al., 2019).

ML relies on data and algorithms (see Whittlestone et al., 2019), which together permeate many sectors of society (e.g., Schwab Intelligent Portfolios, Shanmuganathan, 2020). While algorithms can be defined as step-by-step procedures for solving a problem, data can be defined as numerical and categorical information about objects, events, processes and people that is digitally encoded (see Whittlestone et al., 2019). For example, the following steps represent a solution algorithm for estimating the central tendency in a vector of numbers: (i) sum all the numbers, and (ii) divide the result of the sum by the number of elements in the vector. This algorithm is known as the arithmetic mean (or average). The caveat of this algorithm is that it will be biased if the data does not follow a Gaussian shape. In other words, the output of this algorithm is only reliable if the data can be confidently shown to have a normal shape (e.g., via normality tests). In the context of AI-related technologies, algorithms are procedures designed to perform automated tasks using data sets to support human reasoning and decision making. In other words, data is used to feed algorithms, and algorithms in turn are used to drive AI agents (Siemens et al., 2022). Thus, algorithms are the “ghost in the shell” behind any AI agent. The Figure 1 illustrates this relationship between algorithms, data and AI (here ADA) (Whittlestone et al., 2019).

Algorithms influence our daily lives. Whether it is defining our interests through our browser history (DeVito, 2017), determining what music we should listen to (Werner, 2020), or where we should go for dinner (Tenemaza et al., 2020). On a massive scale, algorithms are being used to extract information from so-called “big data” and support decision making in areas as diverse as surveillance (Tsakanikas and Dagiuklas, 2018), traffic management (Modi et al., 2021), and financial markets (Shanmuganathan, 2020). More recently, a new field of human-algorithm interaction mediated by natural language generation (NLG) systems has emerged, such as the Generative Pre-trained Transformer 3 model (better known as GPT-3) (Brown et al., 2020). GPT-3 produces human-like texts that are difficult to distinguish from texts written by humans (Clark et al., 2021), and this has begun to raise concerns about its use in various contexts, such as academic plagiarism (Dehouche, 2021) or computer programming (Ugli, 2020). While algorithms are increasingly embedded in our digital experiences, it is important to distinguish between their varying levels of impact on human lives.

As such, the majority of algorithms are used in a context that does not significantly affect our lives. We refer to these instances of algorithmic use as low-stakes scenarios. More recently, however, AI and ML algorithms have been used in scenarios that could have a significant impact. For example, algorithms are being used in hiring and promotion decisions (Drakopoulos et al., 2020), the criminal justice system (Rawat et al., 2021), and self-driving cars (Badue et al., 2021), to name a few. We call the latter a high-stakes scenario. That is, the above situations represent two types of scenarios in which algorithms could affect our daily lives: one with little involvement and almost no consequences (low-stakes scenario), and the other with great involvement and consequences (high-stakes scenario).

However, our interactions with algorithms are not limited to low-stakes and high-stakes scenarios and often involve preconceptions related to fear and distrust (Dwork and Minow, 2022). The literature suggests several explanations for why people do not trust algorithms, including a cost-benefit oriented logic where people tend to distrust algorithms even when presented with evidence of their superior performance, as they weigh potential risks more heavily than potential benefits (Debad, 2018). Many see algorithms as an “enigmatic technology” because they are difficult to understand (Beer, 2017) or in some cases, because people believe that algorithms are not capable of learning from their mistakes (Prahl and Van Swol, 2017), but at the same time they also believe that they could be replaced by computers (Granulo et al., 2019; Frank et al., 2019). Algorithmic bias can also affect trust [see examples in medicine (Vayena et al., 2018; Rajkomar et al., 2018). For a recent comprehensive report on trust in AI, see Gillespie et al., 2023].

“Technophobia”, a term coined by Rosen and Mcguire in the 1990s, describes the anxiety caused by a potential interaction with computers or computer-related technology, usually accompanied by negative attitudes toward computers (Rosen and Maguire, 1990; Kim, 2019). Recent demographic analyses have revealed nuanced patterns in technology anxiety. Research indicates no significant gender differences in technophobia scores between males and females, challenging earlier assumptions about gender-based technological comfort levels. The age distribution suggests that technophobia manifests across multiple generations, from young adults through middle age, rather than being concentrated among older populations as often assumed. Professional background data shows particular prevalence among educators and students, with experience levels primarily ranging from novice to moderate. While the studied sample was predominantly White, it also included smaller representations from other ethnic groups, such as Caucasian, Indian, and African American participants (Khasawneh, 2018; Kim, 2019). These findings suggest that technophobia's relationship with demographic factors is more complex than previously assumed, transcending traditional socio-demographic boundaries and affecting individuals across various social, professional, and cultural groups.

Similar existential fears dominate the public debate around concerns such as autonomous weapons (Human-AI Teaming, 2022; Warren and Hillas, 2020). One of these sociological fears is the fear of autonomous robots. This is a widespread fear in different countries (Liang and Lee, 2017; Gnambs and Appel, 2019), even though most people have not had contact with this type of robot. These fears could be the result of exposure to the way robots are portrayed in science fiction or social constructs related to robots, such as the possibility of being replaced by a robot at work (Liang and Lee, 2017; Gnambs and Appel, 2019). This polarization against robots and AI is fuelled by attention-grabbing events such as the recent confirmation by Blake Lemoine, a Google engineer, that the chatbox LaMDA has the ability to express thoughts and feelings like a human child (Luscombe, 2022) or the concerns about text generated by GPT-3 (Dale, 2021). These examples further distract the public from the most legitimate and worrying problems of these systems, such as “data colonialism” or the disturbing parallels between AI development and European colonialism (Adams, 2021). These parallels manifest in several ways: the extraction and exploitation of data from marginalized populations, mirroring colonial resource extraction; the use of Global South populations as testing grounds for AI systems developed in the Global North, reminiscent of colonial medical experimentation; and the imposition of Western conceptual frameworks of intelligence and ethics onto diverse cultural contexts. The field's emphasis on “ethics” often serves, paradoxically, as a form of technocratic rationalization similar to how ethical arguments were used to justify colonial expansion (Adams, 2021). Additionally concerning is that algorithms may reinforce preconceived stereotypes (Bender et al., 2021) and mishandle our personal data or who our data is shared with (Olhede and Wolfe, 2018), perpetuating historical patterns of discrimination and surveillance that characterized colonial governance. In addition, how the data given to algorithms is annotated has a direct impact on algorithmic performance (Tubaro et al., 2020), raising questions about whose worldview and categories are being encoded into these systems.

The media plays a significant role in shaping public perception of AI by covering two main sources of concern: autonomous technology and computer technology (Crépel et al., 2021). Autonomous technology refers to intelligent machines capable of making decisions independently, while computer technology encompasses software that supports communication and computation. The media tends to distinguish between these two categories and also differentiates between fear and criticism when discussing AI. This dichotomous approach to presenting the issues surrounding AI introduces a bias in how we perceive the risks associated with the technology. Consequently, this bias influences the level of trust we place in AI systems. The way the media frames the discussion about AI has a substantial impact on public opinion and can lead to a distorted understanding of the actual risks and benefits of the technology.

Developing a better understanding of how algorithms work and how to modify them can help reduce distrust in these systems, as suggested by several authors (Beer, 2017; Debad, 2018; Marmolejo-Ramos et al., 2022). When people have knowledge about how algorithms work, they can use this information to empower themselves as users. For example, music fans have acted collectively to boost the rankings of certain bands by engaging in massive streaming or downloading (Kang et al., 2022). Another example is Linkedln Brazil, which changed its algorithms to allow job ads targeted at Afro-Brazilians following social pressure (Milton Beck, 2022). These cases show that understanding how an algorithm works can both minimize suspicion and empower users. It is not necessary to understand all the technical details of how an algorithm works, but rather to understand that algorithms use statistical methods to classify, sort, rank and order information. This understanding of statistical concepts is called statistical literacy (François et al., 2020).

The knowledge required to understand and critically evaluate statistical results in order to make decisions based on them is defined as statistical literacy (SL) (François et al., 2020). Since its inception, the concept of SL has evolved (Haack, 1979) to include elements related to the context in which statistical reasoning can be applied (Wallman, 1993). SL plays a crucial role in society (Watson, 1997) and the communication of statistical information is now more important than ever (Gal, 2002). More recently, SL is leading individuals to recognize the importance of mathematics in the world (OECD, 2013).

Due to the statistical nature of algorithms, some level of SL is crucial to understanding what algorithms are capable of, but this understanding will also depend on the level of transparency or explainability of the algorithms (Friedrich et al., 2022). Explainability refers to the interpretability, comprehensibility or readability of the algorithm. Most of the latest algorithms are based on complex multi-layer networks, the basis of deep learning, which use an internal logic that experts cannot fully understand (Carvalho et al., 2019). These systems are called “black box” algorithms and various efforts have been made to promote their transparency (Du et al., 2019). Black box algorithms are less trusted than transparent models because they cannot be explained (Pasquale, 2016).

Several approaches have been proposed to increase the transparency of AI models and reduce systematic errors that affect their performance. One such approach is based on the concept of “model cards for model reporting” (see Figure 1 from Mitchell et al., 2019). This approach suggests that a comprehensive list of information should accompany the description of how the model was trained. This information should include details of the technician who developed the model, the intended use of the model, and the demographic or phenotypic groups on which the model has been tested. In addition, the model card should list the decisions made to optimize the model's performance and the various analyses carried out during the training process. Similar efforts to provide a framework for identifying biases associated with the data used to build or train AI models include the REVISE (REvealing VIsual biaSEs) (Wang et al., 2022) and The Spotlight (d'Eon et al., 2022) projects. These initiatives aim to increase transparency by systematically documenting and disclosing potential biases, enabling more informed use and interpretation of AI models.

Another more complex concern, also related to explainability, is the principle of explicability, a concept that combines intelligibility and accountability as the basis of an interpretable AI model (Herzog, 2022). The latter concept points to the importance of transparency, in the sense that all procedures and details used to build, train and test the AI model should be available during its development and use. This principle is part of the four principles endorsed by the OECD (2019) and the European Commission's High Level Expert Group on Artificial Intelligence (HLEG) to guide the development of “trustworthy” AI: respect for human autonomy, prevention of harm, fairness and accountability (Hagendorff, 2020). Despite consensus on these four principles, we are still far from creating a legal framework that guarantees accountability mechanisms in AI development (Mittelstadt, 2019).

In this context, our work presents an experimental study that looks at factors that might explain why people trust algorithms, such as: SL, explainability, stake levels, demographics, among others.

Data from 3,260 participants were available from 20 countries (Armenia, Australia, Bulgaria, Brazil, Cameroon, Colombia, Czech Republic, Spain, Indonesia, India, Italy, Japan, Nigeria, Philippines, Thailand, Turkey, Taiwan, UK, USA, and Vietnam). However, only participants who provided complete data were included in the analyses (n = 1,921) (see Figure 2, M_age = 26.03 ± 9.88 SD; 59.5% women, 38.2% men, 1.8% other). Each participating laboratory obtained ethical approval from its local ethics committee, and data collection began only after ethical approval (the ethics approval for the leading research group in Australia was granted by the University of South Australia, with the approval number 203238. This approval was then used by the other participating laboratories to obtain their own respective ethics approvals). All participants voluntarily accessed the internet link for this study and agreed to participate after reading the information page and agreeing to take part. They were recruited via social media using convenience sampling.

This online survey consisted of four sets of questions: (1) a demographic questionnaire in which participants were asked about their first language, country of residence, age, gender, level of education, level of familiarity with ADA (their level of familiarity with ADA was assessed using a visual analog rating scale (VAS) ranging from 0 [not very familiar] to 5 [very familiar] and using up to two decimal places); (2) a VAS rating scale version of the six-item ‘propensity to trust scale items' from Merritt et al. (2013), with a range of responses from 0 (strongly disagree) to 5 (strongly agree), using up to two decimal places; (3) a selection of 14 items (questions 2, 4, 9, 10, 12, 14, 18, 19, 27, 31, 34-37) from the 37-item Basic Literacy In Statistics (BLIS) scale (Ziegler and Garfield, 2018). The 14 items from the BLIS were chosen to cover different statistical concepts equally, i.e., items 2 and 4 relate to data production, items 9 and 10 to graphs, items 12 and 14 to descriptive statistics, items 18 and 19 to sampling distributions, items 27 and 31 to hypothesis testing, items 34 and 35 to the scope of conclusions, and items 36 and 37 to regression and correlation (these items are available in the supplementary material via the Qualtrics files). Finally, (4) 12 scenarios related to situations in which algorithms are used (half related to low-stake situations and the other half to high-stake situations). Each scenario was followed by two questions (see below), which were answered on a VAS rating scale from 0 (not at all likely) to 5 (very likely), using up to two decimal places. The results of expert judgement of these items are provided in the supplementary material. All phases of the study were programmed and distributed using Qualtrics™.

Two scenarios were created to illustrate different situations in which people interact with algorithms. Half of them represented low-stake situations, i.e., (1) algorithms to make restaurant recommendations, (2) to select stories for online news, (3) to organize and sort emails, or (4) to suggest new restaurants, (5) new clothes, and (6) new music. The other half represented high-stakes situations, i.e., (7) algorithms to support court decisions based on psychological profiles, (8) to select CVs, (9) to make hiring recommendations for a job, (10) to select the best candidate for a position at a university, (11) to control the brakes of autonomous vehicles, and (12) to decide the priority of care in a medical context.

Each scenario contained a sentence related to its explainability. These sentences contained information about a specific machine learning method used by the algorithm (e.g., clustering learning methods, classification learning statistical methods, logistic regression methods, dimensionality reduction techniques, supervised statistical methods and clustering statistical methods). The sentence also briefly mentioned the quality of the method.

The following are examples of two different scenarios used to evaluate trusting algorithms:

The experiment is a 2 × 2 factorial design: the importance of the situation in which an algorithm is used (low and high stake situation) and the explainability of the algorithm (with and without). These factors were implemented in the 12 scenarios via two lists; list 1 = six low-stake scenarios with explainability and six high-stake scenarios without explainability, and scenario list 2 = six low-stake scenarios without explainability and six high-stake scenarios with explainability. The four sets of questions were counterbalanced across participants, resulting in four experimental conditions (see Figure 3). Qualtrics ensured that participants were randomly assigned to each condition and that a balanced number of responses were collected for each condition. While the median time to complete the task was 24 minutes, there was some variation, with an interquartile range of 27 min (i.e., half of the participants completed the task within a 27-minute time span).

Data analysis was conducted using multilevel linear models implemented in the R packages lmerTest and lme4 (Kuznetsova et al., 2015, 2017). The significance level for all statistical tests was set at α = 0.05. The model tested was: p~e*S*BLIS+g+a+ADA+c+(1|id) + (1|i) where “p” is the probability of trusting/recommending/using algorithms, “e” is the presence of explainability, “S” is the stake level (i.e., high and low stake), “BLIS” represents statistical literacy (frequency of correct answers), “g” represents participant gender, “a” represents participant age, “ADA” represents participant familiarity with ADA, “id” represents subject identification, “i” represents each of the 12 scenarios, and “c” represents participant country (“*” represents main effects and interactions. Only numeric variables are shown in teletype font; other variables are categorical. The variable “propensity to trust scale” was not added as a covariate as it showed a high correlation with the dependent variable, r₍₁₇₆₈₎ = 0.69, p < 0.001).

A stepwise backward model/variable selection algorithm was applied to this model to produce a significant and parsimonious model. The initial and final models were evaluated using metrics such as AIC and AICc weights (Wagenmakers and Farrell, 2004), R² (coefficient of determination) for conditional (both fixed and random effects) and marginal (fixed effects) models, and performance score. These metrics were estimated using the performance R package (Lüdecke et al., 2021).

Once a parsimonious model was found, the marginal and conditional R² values were estimated using the r2 nakagawa command from the performance R package (Lüdecke et al., 2021), then, the variance components of the random factors were estimated using the gstudy command from the gtheory R package (Huebner and Lucht, 2019).

For access to all materials and analysis codes, including a machine learning approach, visit the following link: https://figshare.com/projects/Trust_in_algorithms_An_experimental_approach_-_Data_repository/156212.