Content area
While Artificial Intelligence (AI) and computer algorithms have become increasingly embedded in everyday life, concerns over biases in these systems have also been rising. Although much attention has been devoted to data-centric approaches that see the source of bias in the training data fed to these systems, this paper focuses on the second source of bias: biased programmers. This stance defends that programmers might unintentionally and unconsciously embed their worldviews into their codes. Drawing on an ontology of “bias in automated decision making” that distinguishes between first- and second-level discrimination and arbitrariness, we propose a novel twofold transparency concept to address second-level arbitrariness. To this goal, we transpose and adapt methodological tools from the social sciences: reflexivity and positionality statements. First, we advocate for the adoption of Algorithm Designers’ Reflexivity Statements (ADRSs), namely confidential internal written reflections that encourage programmers to critically examine and articulate their assumptions and potential biases. Second, we propose synthesising these reflections into an internal ADRSs Report and then into a public AI Positionality Statement (AIPS), which communicates to end users the residual and inherited biases that may skew algorithmic outputs. This dual approach not only enhances internal bias awareness but also equips AI users with a contextual framework to interpret algorithmic decisions, thereby promoting fairness and increasing trust in AI systems.
Introduction
Alongside the rapid integration of artificial intelligence (AI) in various aspects of our lives, and due to high-profile cases that attracted media attention, concerns over biases in algorithms and AIs have been rising (Amini et al., 2019; Floridi, 2024b). This issue has also recently gained momentum in the academic discourse, when calling for the urgent “need to develop strategies and tools for detecting incorrect, hallucinated and intentionally misleading content” (Nature Machine Intelligence, 2024).
In response to these concerns, two main strategies have emerged. The first (and the most researched) approach focuses on the data used to train the algorithms or AIs. The core idea here is straightforward: garbage in, garbage out – also known as the “BIBO problem” (Floridi, 2024b). If the training data is biased, the outputs generated by the algorithm will likely reflect and perpetuate those biases. This data-centric approach has led to various mitigation efforts, such as creating fairer datasets (Mittal et al., 2024), documenting their provenance (N. Vincent, 2024), or even conducting algorithmic audits (Morewedge et al., 2023).
The second strategy, which has received less attention (Johansen et al., 2023), shifts the focus from data to the individuals who develop these systems, namely the programmers, and how these perceive the world according to “lived experience, specific understandings, and historical background” (Finlay, 2002). Hence, this approach builds on the idea that “being-in-the-world means that researchers [and programmers] cannot help but bring their own involvement and fore-understandings into the research [or codes]” (Finlay, 2002). Referred to as “biased programmers” (Cowgill et al., 2020), this stance contends that algorithms (and consequently AIs) are written by people who inevitably encode their worldview into their algorithms, thus biasing what and how data is used and, consequently, the output of AIs. As aptly summarised by Baeza and Yates (2018), “each programme probably encodes the cultural and cognitive biases of their creators” and these “implicit values inherent in automated decision-making systems are contestable on both epistemic and normative grounds” (Binns, 2018).
In this paper, we focus on this second, less researched source of bias: programmer bias – or sometimes also referred to as ‘coder bias’. Adapting from the social sciences, an instrument used to tackle the issue of researchers’ biases in qualitative social science research, we propose a twofold transparency strategy to address the transposition of bias from biased programmers to algorithms and AIs. First, we advocate the adoption of algorithm designers’ reflexivity statements (ADRSs), i.e., written reflections (for programmers or organisations’ internal use only) in which algorithm designers explore their assumptions, perspectives, and worldview, focusing on the biases they might hold and that they may unintentionally and unconsciously introduce into their code and design choices. Writing an ADRS is an exercise or training that creates awareness by making programmers “embrace their own humanness” (Walsh, 1995) so as to reduce inherent biases.
Second, granted that algorithms and AI systems are coded by multiple contributors, we propose, within an organisation, to have the multiple ADRSs internally collected and, through a careful and rigorous synthesis process, to be condensed into, first, an internal ADRSs Report, and, second, into a public AI positionality statement (AIPS): a declaration that clarifies inherited, residual, potential biases inherent in the AI and that could materialise in the AI’s output. While the suggested ADRSs Report enables organisations to gain insight into their internal worldview’s diversity (allowing for affirmative or corrective actions when necessary), AIPSs serve as a crucial interpretive tool for AI users to better understand and, more importantly, contextualise AI outputs (Mahajan, 2025) in light of unintended residual biases.
To defend these proposals, we have structured the paper into two sections. Section I, which is divided into three chapters, provides the paper’s theoretical framework. First, the two primary sources of bias in algorithms and AIs, namely biased training data and biased programmers, are presented; the paper focuses on and addresses this second and less-studied source. Second, a typology of bias is presented. While various philosophical analyses exist (Bueter, 2022; Kelly, 2022), we base our argument on Ulrik Franke’s (2022) theoretical framework. This choice is motivated by Franke’s development of an ontology of bias within the discourse on AI, making it particularly well-suited to the objectives of this paper. In this work, Franke (2022) presents and defines first-level and second-level discrimination and arbitrariness. Third, always drawing from Franke (2022), a gap in the literature on how to mitigate second-level arbitrariness is identified; the paper aims to fill it by developing two instruments that cover the three possible approaches suggested by Franke (2022) on how to decrease such type of bias: (1) “articulating and scrutinising […] choices of standards”; (2) “creating design situations with more diversity”; (3) “ensuring proper competition and consumer choice”.
Having defined the paper’s theoretical framework and the literature gap, Section II introduces ADRSs and AIPSs. To this goal, first, the origins and main features of reflexivity and positionality statements are explained, and their effectiveness in the social sciences (especially anthropology and ethnography) in acknowledging the presence of researchers’ bias and providing more rigorous, transparent, and trustworthy research is emphasised. Second, granted their benefits, an adaptation of these instruments to address the issue of second-level arbitrariness in the AI field is carried out and followed by a proposal for its operationalisation. Finally, the benefits that such instruments could provide to the field and to AI users are underscored.
Section I
Bias in AI: two approaches
Algorithms and artificial intelligence (AI) systems are becoming increasingly ubiquitous, revolutionising entire sectors (Binns, 2018; Crawford & Calo, 2016; de Laat, 2018; Peters, 2022), from healthcare (Bjerring & Busch, 2021; Gichoya et al., 2023) to education (Bellas et al., 2024; X. Wang et al., 2024; Zhao et al., 2024), from HR management (Prasad & De, 2024) and job hiring (Bogen, 2019) to banking (Vasilogambros & Journal, 2015), from green finance (Q. Wang et al., 2025) to sustainable development (Kirikkaleli et al., 2025). This rapid, ubiquitous proliferation has been driven by an “explosion in the availability of online data and low-cost computation” (Jordan & Mitchell, 2015), which fostered a “virtuous circle where new algorithms, more online data and cheaper computing power have reinforced each other” (Franke, 2022). However, as AI becomes more pervasive, concerns about “machine biases” (Johansen et al., 2023) and algorithmic bias have emerged. Although there is no consensus on the definition (Gichoya et al., 2023), one of the agreed features of bias, which differentiates it from mere errors, is its “systematic tendency to commit the same type of error over time or in different situations” (Johansen et al., 2023).
The presence of biases in algorithms and AIs often creates dissonance in the population as these are believed to be objective and accurate, namely, immune to such processes (Johansen et al., 2023). However, the truth is far from the broader perception of objectivity: in fact, algorithms have been shown to express racial, gender, colour, and other personality traits biases (Amini et al., 2019; Chen, 2023). Examples that show such biases are numerous: from machine learning methods for detecting COVID-19, which, due to “underlying biases stemming from small training data sets, data set variability, and limited integration of non-imaging data” (Gichoya et al., 2023), were deemed unfit for clinical use (Roberts et al., 2021), to racial biases in an “algorithm used to manage the health of populations” (Obermeyer et al., 2019), or even algorithms used for sentencing decisions to assess probabilities of reoffending (Rudin et al., 2020) that have been accused of being biased against black people (Nature, 2016). While among the cases that caught the public attention, there were Amazon’s AI tool for job recruiting (J. Vincent, 2018), which was accused of “systematically downgrade women’s CVs and so displayed a gender bias” (Peters, 2022) or Facebook’s discrimination of Black men, where an AI put the labelled as “primates” videos of Black men (Mac, 2021).
Given the increased presence of such systems in all aspects of life and considering the pernicious effects of bias, a “need for curation and safeguarding of high-quality data and information” has been deemed as “more crucial than ever” (Nature Machine Intelligence, 2024) – also to ensure a long-term acceptance of these systems (Courtland, 2018). To this goal, attention has been addressed in two directions: (mostly) towards training data and (less) towards programmers.
The first and prevailing explanation for bias and incorrect output of algorithms identifies training data as the source of the problem. The “biased training data” (Cowgill et al., 2020) framework defends that “[t]he decisions made by AIs are shaped by the initial data it receives. If the underlying data is unfair, the resulting algorithms can perpetuate bias, incompleteness, or discrimination” (Chen, 2023). Fundamentally, this ‘biased data in, biased data out’ perspective, “reflect[s] and amplify[ies] societal biases inherent in their training data”(Floridi, 2024a). Lum and Isaac (2016) underline how, for example, “if police focus attention on certain ethnic groups and certain neighbourhoods, it is likely that police records will systematically over-represent those groups and neighbourhoods” and when such biased data is employed to train an AI system, the “model will reproduce […] those same biases”. To tackle this issue, current quality-securing strategies have been focusing on creating fairer datasets (Mittal et al., 2024), documenting their provenance (N. Vincent, 2024), conducting algorithmic audits (Morewedge et al., 2023), creating datasheets (Gebru et al., 2021), or establishing ethical governance frameworks for AI development.
However, focusing solely on the problem of biased training data overlooks a critical dimension of the problem: bias in algorithms also stems from “biased programmers” (Cowgill et al., 2020). This perspective, which has unfortunately received comparatively less attention in the academic discourse (Johansen et al., 2023), rests on the theoretical background according to which
“each person […] perceive[s] the same phenomenon in a different way; each person brings to bear his or her lived experience, specific understandings, and historical background. This way of being-in-the-world means that researchers [in this case, programmers] cannot help but bring their own involvement and fore-understandings into the research [and codes]” (Finlay, 2002).
As such, the biased programmers approach posits that algorithms are written by people who – intentionally or unintentionally, consciously or unconsciously (Stinson, 2022) – inevitably embed their axiological, epistemological, and ontological assumptions (namely, their worldview) into codes. As noted by Baeza and Yates (2018), “each programme probably encodes the cultural and cognitive biases of their creators”. Such a form of bias is more subtle. It can manifest in the “garden of forking paths” (Gelman & Loken, 2013) taken in a coding process or in other various forms: from the selection of target variables and class labels, to the labelling, handling and processing of the data, to even the choice of proxies or even the data set itself (Barocas & Selbst, 2016). This is more evident in predictive AI when the developer must create a simplified model of the real world: the developer has to “make choices about what attributes they observe and subsequently fold into their analyses” (Barocas & Selbst, 2016); however, such a simplification of reality introduces arbitrary decisions based on personal values and beliefs, which are “are contestable on both epistemic and normative grounds” (Binns, 2018). For example, one can imagine a scenario where an AI must predict whether a job applicant will be a good employee or not. In this case, the characteristics of what a ‘good’ employee is are arbitrary and subjective (Zuiderveen Borgesius, 2018): one programmer might hold that laziness is a vice, while someone else might hold the opposite, defending that laziness is actually a virtue, as a lazy employee will find how to do the job in less time and with less effort.
In this paper, we are going to focus on this second source of bias and propose a strategy to address it.
The ontological matrix of bias
In an article from 2022, Ulrik Franke expands on Robert Nozick’s ontology of bias presented in his The Nature of Rationality (Nozick, 1994). Franke first develops Nozick’s only-sketched differentiation between discrimination and arbitrariness, defining the first as the “intentional” bias “done in order to exclude certain cases” (Franke, 2022) – here, intentionality, exclusion and the production of detriment are necessary conditions – and the latter as the “unintentional [bias] […] not done in order to accomplish anything in particular, or if it is intentional, [it is] not done in order to exclude certain cases” (Franke, 2022), but to include them. By combining such differentiation with the two-level categorisation of bias already developed by Nozick, an ontological matrix (see Table 1) composed of four types of bias (first-level discrimination, second-level discrimination, first-level arbitrariness, and second-level arbitrariness) is constructed. These distinctions are of fundamental importance as they represent the theoretical scaffolding on which the paper develops, and they help identify the target of the instruments this paper advocates for. For clarity, we present these biases in more detail using a common scenario, namely, a job hiring process. When an AI-related case is particularly well-suited to illustrate a specific bias, we include it alongside the hiring scenario to provide a more comprehensive analysis.
Table 1. Bias typologies.
Bias | Discrimination | Arbitrariness | |
|---|---|---|---|
I level | Definition | “Intentional uneven application of existing standards in order to exclude certain cases” (Franke, 2022) | “Other uneven applications of existing standards” (Franke, 2022) |
Formal | if A → S1 but if B → S2 | A → ¬S1 | |
Example | When, in hiring processes, black candidates are requested to provide their criminal records while white ones are not | When the standard requires candidates to provide their criminal record, but the job offeror forgoes this requirement for a specific candidate | |
II level | Definition | “Intentional uneven choice of standards in order to exclude certain cases” (Franke, 2022) | “Other uneven choices of standards” (Franke, 2022) |
Formal | S1 → ¬B | S {S1, S2, S3, …, Sn} → Sx | |
Example | The requirement, in a job offer, of having specific titles, degrees, or capacities that the job offeror knows that members of a given group do not have but those of another do, so as to exclude the former from applying | When in a job offer, the employer might decide to require candidates to have a university degree from a specific institution (S1) without considering equally relevant alternatives, such as degrees from other accredited institutions (S2) or equivalent professional experience (S3). | |
First-level discrimination
It consists of an intentional – meant to exclude a specific group – biased application of existing standards (Franke, 2022). A first-level discrimination example is when, in hiring processes, black candidates are requested to provide their criminal records while white ones are not. In this case, for members of group A (in the example, black candidates), standard S1 (provide the criminal record) is applied, while for members of group B (white candidates), standard S2 (no criminal record is requested) is applied. More formally, in this case, the “intentional uneven application of existing standards”, which gives rise to the discrimination, is based on a race criterion: if a member of A, then S1 applies, but if a member of B, S2 applies (if A → S1 but if B → S2). A hypothetical case, based on a scenario proposed by Zuiderveen Borgesius (2018), sees a company developing a “pregnancy prediction score” (Zuiderveen Borgesius, 2018) through which, based on some patterns and behaviours, it can predict whether the employee is pregnant (or willing to) and, based on this information, assign promotions.
Second-level discrimination
It covers the “intentional uneven choice of standards in order to exclude certain cases” (Franke, 2022). An example of this type of discrimination could be the requirement, in a job offer, of having specific, expensive titles, degrees, or capacities (S1) that the job offeror knows that members of group B do not have but those of A do, so as to exclude Bs from applying. Hence, in this case, the focus is shifted onto the choice of the standard with the goal of achieving a certain result, namely the exclusion of a person or a group of people. Formally put: S1 → ¬B (use of S1 so as not to have B). Such a type of bias could occur, for example, in an AI context when “a prejudiced decisionmaker […] pick[s] proxies for protected classes with the intent of generating discriminatory results” (Kroll et al., 2017).
First-level arbitrariness
First-level arbitrariness covers “other uneven applications of existing standards” (Franke, 2022). In this case, the uneven application is voluntary and grants benefits or inclusion into a certain group – thus the opposite of discrimination. A clear example that Franke provides to illustrate such a concept is pardons. These are inherently arbitrary applications of standards: an inmate who, according to the standard, is in prison because he committed a given crime, is pardoned by the president based on an arbitrary, highly uneven application of standards (in this case, the possibility of releasing someone). Referencing back to the previous job offer example, if the standard is to require candidates to provide their criminal record (S1), the job offeror that forgoes this requirement for a specific candidate (let’s say a friend) is carrying out first-level arbitrariness. Formally put: A → ¬S1 (it should be S, but since you are A, let’s not apply S1).
Second-level arbitrariness
Finally, second-level arbitrariness arises when “a standard may be chosen without regard to all the relevant alternatives” (Franke, 2022). This could be formalised with S = {S1, S2, S3, … Sn} → Sx: among the possible standards S1, S2, S3, … Sn within the set S, a specific standard Sx is arbitrarily chosen, where Sx ∈ S. An example of such arbitrariness is when, in a job offer, the employer might arbitrarily decide, for no specific reasons, to require candidates to have a university degree from a specific famous institution (S1) without considering equally relevant alternatives, such as degrees from other accredited institutions (S2) or equivalent professional experience (S3). In this case, the choice of standard S1 over the equally valid standards (S2 and S3) is arbitrary because it excludes potential candidates who meet the job’s actual requirements but through different, legitimate pathways. This arbitrariness does not stem from an intention to discriminate against a particular group (as it was in second-level discrimination) but rather from a failure to weigh all reasonable options, resulting in an unjustified exclusion. As Franke points out, “arbitrariness is a matter of degree” (Franke, 2022); as such, second-level arbitrariness cannot be eliminated but reduced: for example, a “rational person will use some procedures to operate upon and correct other of her procedures, to correct biases in the sampling and evaluation of reasons and in the procedure of estimation based upon them” (Nozick, 1994). Tracing back to the example, the job offeror who discovers that there are other equally viable standards, may decrease his second-level arbitrariness by rewriting the job offer in a way that expands the accepted standards.
Addressing the different types of bias: the gap in the literature
Having identified the four different types of bias (first- and second-level discrimination and first- and second-level arbitrariness), we now examine how these are addressed (see Fig. 1).
[See PDF for image]
Fig. 1
AI governance levels and proposed expansion.
First- and second-level discrimination
When it comes to discrimination, be it first- or second-level, as previously explained, there is an intentional exclusion or different treatment of people based on some given features (race, gender, faith, etc.). Granted the intentionality and the nature of the exclusion, discrimination is typically addressed within the realm of the law. In a recent study for the Council of Europe, Zuiderveen Borgesius (2018) highlighted how “the most relevant legal tools to mitigate the risks of AI-driven discrimination are non-discrimination law and data protection law”.
First-level arbitrariness
The issue becomes more complex when the issue moves to addressing arbitrariness, as there is no intention to discriminate against a given group – although the actions might produce discrimination. In this context, pulse oximeters are an emblematic case of first-level arbitrariness (Franke, 2022). A recent paper by Sjoding et al. (2020) has underlined the presence of “racial bias in pulse oximetry measurement”: such instruments measure patients’ oxygen saturation and are commonly employed in triaging in intensive care units; however, as evidenced by the presence of racial bias (to be understood here as arbitrariness, as, prima facie, there seems to be no discriminatory intentions), due to the overestimation of blood saturation in black patients, these were put in risk of becoming hypoxemic, “i.e., an arterial oxygen saturation of <88% despite an oxygen saturation of 92 to 96% on pulse oximetry” (Sjoding et al., 2020). The arbitrariness stems from the fact that, “it is not that different oximeters are used for different groups of patients”, which otherwise amount to first-level discrimination, “but it is rather that applying the same oximeter, attempting to apply the same standard (blood-oxygen saturation, measured as a percentage), the machine nevertheless implements an uneven application of this standard” (Franke, 2022).
Most of AI governance has been developed to address this type of arbitrariness: it has been proposed to increase transparency (de Laat, 2018; Shin, 2021) and explainability (Barredo Arrieta et al., 2020; Das & Rad, 2020; Dwivedi et al., 2023; Fan, 2025; Ziosi et al., 2024) to make sure that people can “obtain a factual, direct, and clear explanation of the decision-making process, especially in the event of unwanted consequences” (Floridi et al., 2018). Hence, the rationale behind such an approach is to uncover issues whose creation might have been unintentional or unconscious and create awareness of the issue so as to correct it. It is according to this approach that “debiasing algorithms” (Peters, 2022) and instruments such as Amazon’s SageMaker Clarify, IBM’s AI Fairness, Google’s Fairness Indicator, or Microsoft’s Responsible ML are being developed (Franke, 2022). The peculiarity of such instruments is that they enable “developers to automatically and regularly apply statistical tests of first-level bias to their systems, to inspect models in order to explain their decisions and inner workings, and offer suggestions on how to rectify any unwanted behaviour detected” (Franke, 2022).
Second-level arbitrariness
While “debiasing algorithms” (Peters, 2022) represent a significant step forward, they cannot detect for second-level arbitrariness – thus making it particularly challenging to address this last form of bias. Due to these technical limitations and the nature of this type of bias itself, second-level arbitrariness has received far less attention in the literature despite its “quite common [presence] in the design and use of automated decision-making” (Franke, 2022).
Nevertheless, second-level arbitrariness plays a crucial role within the biased programmers’ approach as it covers unintended and unconscious biases introduced by programmers in their code by, for example, not knowing that the element X could have been assessed not only through the framework Y, but also Q, W and Z. Although this might appear farfetched, imagine the following hypothetical, revised, trolley problem: an AI system implemented into an autonomous vehicle is tasked to make a decision in a case where if the car continues on its path it will run into a car with five passengers, but if it diverts its course – say by running on the pavement – it will run into a mother with a stroller; no other option is available but these two. Let now assume that the programmer responsible for encoding the system’s decision-making process has never followed an ethics course in their life, but by reasoning with the numbers the programmer codes into the algorithm a utilitarian ethical perspective where the right action is the one that brings about the greatest happiness for the greatest number (hence, running over the mother with the baby as this creates fewer deaths than ramming into the car with five occupants (2 < 5)). By being unaware that there are alternative ethical frameworks to utilitarianism (such as deontology or virtue ethics), the programmer is unwittingly encoding his second-level arbitrariness into the algorithm – precisely as the job offeror that did not know that other university titles and programmes would have provided equally competent prospective employees, but arbitrarily set up a standard. As it is apparent, the choice of the system to be used is arbitrary and “contestable on both epistemic and normative grounds” (Binns, 2018) since it results from an omission or lack of consideration of alternative possible standards to be used (among the possible standards S1, S2, S3, … Sn within the set S, a specific standard Sx is arbitrarily chosen, where Sx ∈ S). Unlike first-level arbitrariness, which statistical tools can identify and mitigate, second-level arbitrariness is not a purely technical issue but, as Franke (2022) notes, it is a “normative question”.
Although such a type of bias cannot be fully eliminated, it can be mitigated by different means. Franke (2022) suggests three possible approaches to do so: (1) “articulating and scrutinising our choices of standards”; (2) “creating design situations with more diversity”; (3) “ensuring proper competition and consumer choice”. Building on this foundation, and with the goal of filling this gap in the literature on how to mitigate this type of arbitrariness, we propose a practical twofold strategy (the adoption of ADRSs and AIPSs) that integrates all three aspects. ADRSs address paths (1) and (2), encouraging programmers to explicitly reflect on their normative commitments and engage with diverse perspectives, while AIPSs address path (3) by fostering transparency regarding the epistemic and ethical positions – and inherent residual bias – embedded in AI systems, thereby enabling stakeholders to make informed choices.
Such a proposal should not be understood as overturning existing approaches to AI governance but rather as an expansion of them: as an additional layer (Fig. 1). Current frameworks largely concentrate on first- and second-level discrimination (particularly in the legal domain) and on first-level arbitrariness, while second-level arbitrariness remains largely neglected. This proposed expansion could be readily incorporated into explainable AI (XAI) governance (Ali et al., 2023; Barredo Arrieta et al., 2020; Das & Rad, 2020; Dwivedi et al., 2023; Gunning et al., 2019; Shin, 2021) by adopting broader conceptions of the notions of transparency, interpretability and explainability (Bernardo et al., 2023). Such an approach would not only require clarifying the technical operations of AI – the conventional effort to “open the black box” – but also address the social aspects, too: who constructed the box, why it was designed in a particular way, and what intended or unintended consequences those design choices may entail for end users. As Danks (2024) emphasises, “[s]uccessful governance of AI requires some knowledge of how, and more importantly why, the system functions as it does”. Yet this “why” should extend beyond purely technical dimensions; otherwise, governance risks reproducing the narrow lens that reduces bias to biased data alone.
Section II
Having defined the theoretical framework of the paper and the target of the proposed epistemic enhancer strategies, we now present the concepts of reflexivity statements and positionality statements.
Reflexivity statements and positionality statements in the social sciences
Reflexivity has been deemed a fundamental feature of qualitative research (Lazard & McAvoy, 2020; Sybing, 2022) and its theoretical foundations as a methodological tool in ethnography and anthropology trace back to the 1970s (Finlay, 2002). It consists of a critical, “disciplined” (Wilkinson, 1988) self-examination: an enquiry of “one’s own assumption, belief, and judgement systems” that requires “thinking carefully and critically about how these influence the research process” (Jamieson et al., 2023). In other words, it asks researchers to critically evaluate how their axiological, epistemological, and ontological assumptions, namely how they see the world, influence their research (Holmes, 2020). Fundamentally, it requires them to answer the question: “what is the research process and how am I influencing it?” (Lazard & McAvoy, 2020). As such, reflexivity amounts to an ongoing phenomenological self-reflection that questions the ethics and utility of what is studied as well as who, how, and why that research is carried out (Willig, 2013): a meticulous analysis that creates (i) a dialectical enquiry between the research and the researcher to understand how one influences the other (Wilkinson, 1988) and (ii) a disentanglement of “perceptions [from] interpretations from the phenomenon being studied” (Finlay, 2002).
However, as noted by Jamieson et al. (2023), reflexivity must not be confused with reflection – although the two are part of a continuum (Finlay, 2002; Shaw, 2010). This is because the former is an “active acknowledgement” of one’s worldview carried out “before, during, and after” the research; the latter, on the other hand, has a past-only temporal dimension, meaning that it is carried out “retrospectively and typically leads to insights about details that were ‘missed’ in the original research process” (Jamieson et al., 2023), fundamentally, it can be understood as a distanced “thinking about” (Finlay, 2002) something. As such, due to this continuous temporal nature, it has been remarked how reflexivity “has a greater potential to guide the research process, across all research epistemologies and methodologies” (Jamieson et al., 2023).
Such a process of self-examination can be formalised into what is known as a reflexivity statement. We stipulatively define this as a written text that documents one’s engagement with their axiological, epistemological, and ontological systems, critically examining how these might be influencing their research. As such, a reflexivity statement is not merely a declaration but a dynamic report of an ongoing process over time. Some of its elements might be “regarded as fixed, for example gender, race, skin-colour, nationality” while others “such as political views, personal life-history, and experiences, are more fluid, subjective, and contextual” (Holmes, 2020). Given this fluidity, it is crucial to document how these elements evolve throughout the research process.
This formalisation has the benefit of uncovering or pinpointing possible conscious or unconscious biasing elements that might be voluntarily or involuntarily transposed into the research selection, question, framing, methodology, outputs, and conclusions (Finlay, 2002; Grix, 2018; Malterud, 2001; Rowe, 2014). This written formalisation is of fundamental importance in the scientific process as, by increasing the transparency of the research process (Sybing, 2022), it augments the research validity, rigour and trustworthiness (Busso & Leonardsen, 2019; Massoud, 2022), and provides the reader a reading key to interpret the research itself and its findings in light of the expressed world view.
Although reflexivity and positionality are often used interchangeably in the literature, we believe that they should be distinguished: reflexivity is the description of an ongoing process, whereas positionality represents a snapshot of that process at a given point in time. In line with Holmes (2020), we maintain that “reflexivity informs positionality” and, accordingly, argue that reflexivity is ontologically prior to positionality. Therefore, we stipulatively define a positionality statement as the formalised articulation of a reflexivity process that captures and freezes in time the researcher’s reflexivity process with which the research was undertaken at the moment of its documentation.
Transposing reflexivity and positionality statements to the AI field
Having defined the concept of reflexivity and positionality statements, we now move to defend a transposition to the field of AI in two forms: ADRSs (algorithm designers’ reflexivity statements) and AIPSs (AI positionality statements).
As highlighted in Section I, algorithms are written by programmers that inevitably embed their worldview into codes: “each programme probably encodes the cultural and cognitive biases of their creators” (Baeza-Yates, 2018). While biased training datasets can often be identified and corrected using technical and statistical tools, the biases introduced by programmers themselves are more subtle and difficult to detect. If we set aside cases of intentional discrimination – i.e., “intentional” bias “done in order to exclude certain cases” (Franke, 2022) – which fall under the realm of the law, the key challenge becomes identifying and addressing arbitrariness, namely the “unintentional [bias] […] not done in order to accomplish anything in particular, or if it is intentional, [it is] not done in order to exclude certain cases” (Franke, 2022), but to include them. As seen, particularly problematic is second-level arbitrariness, which occurs when “a standard may be chosen without regard to all the relevant alternatives” (Franke, 2022). This is because this type of arbitrariness, being unconscious and involuntary, manifests in seemingly arbitrary choices – such as the selection of model parameters, target variables, class labels, data processing methods, proxies, or even the dataset itself. Unlike first-level arbitrariness, second-level arbitrariness cannot be statistically measured and corrected (see Section I.3).
To address this specific type of bias, we propose transposing reflexivity statements – a widely used tool in social sciences (see Section II.1) – into algorithms and AI development. Just as qualitative researchers disclose their positionality to make a possible interference between observer and observation transparent, we defend that algorithm designers should engage in a similar practice. We argue that programmers should undertake a reflexive process where they exercise and train to explore the assumptions, perspectives, and cultural biases they hold and that they may unintentionally and unconsciously introduce into their codes and designs, and to formalise this self-enquiry into an individual ADRS. Fundamentally, we invite algorithm designers to “embrace their own humanness” (Walsh, 1995) and consider “their own involvement and fore-understandings” (Finlay, 2002) in an exercise of bias-awareness that aligns with Franke’s (2022) suggested strategy of “articulating and scrutinising […] choices of standards”.
This transparency exercise is expected to foster significant behavioural changes. To understand why, it is helpful to draw a parallel with Jeremy Bentham’s (1791) Panopticon theory, later interpreted by Michel Foucault in Discipline and Punish (1995). Bentham’s proposal for an “inspection house” – applicable to institutions such as hospitals, factories, prisons, and schools – centres on an architectural design in which individual rooms (for instance, prisoners’ cells) are arranged around the circumference of a circular building. At its centre stands a watchtower from which an observer can see into every room, while the occupants, blinded by the interplay of light and shadow, cannot tell whether they are being watched. This uncertainty triggers a self-regulatory mechanism referred to as the panopticon effect (Downing, 2010; Seele, 2016): inmates adjust their behaviour as if they are constantly under surveillance. As Foucault aptly observed, the inmate “becomes the principle of his own subjection” (Foucault, 1995). Likewise, but for reading the subjection in a negative connotation, it is possible to see how ADRS might foster a self-regulatory discipline: programmers, aware that their reflexivity or positionality statements may be read and scrutinised by others, actively moderate, reflect upon, and counter their own biases.
To put this into practice, a structured yet flexible process must be developed (see Fig. 2). The initial phase would involve defining the scope and objectives of the ADRS programme, developing internal guidelines, addressing possible resistance by algorithm designers, appointing supervisory bodies to oversee the process, and establishing data collection protocols to ensure consistency. Once this preparatory work is in place, algorithm designers would be asked to engage in a structured exercise of self-reflection – akin to that undertaken by social scientists – in which they examine their ontological, epistemological, and axiological perspectives and how these influence their coding and design choices. The results of these self-examinations would then be formalised into written individual ADRSs, i.e., written statements that critically assess worldviews and values programmers hold and how these may impact algorithmic decisions.
[See PDF for image]
Fig. 2
Diagram flow and process steps of an ADRS and AIPS strategy.
Two points warrant further clarification. First, this paper does not propose a standardised format for such self-assessment, since such an endeavour would fall beyond its scope. Furthermore, as the literature underscores, “there is no ‘one size fits all’ […] authors should feel able to share as much (or as little) of themselves as they feel safe and comfortable with” (Jamieson et al., 2023). Nevertheless, we recommend that future research focus on identifying key guiding questions for producing such ADRSs that carefully account for the specificities of the tech field and the potential challenges posed by overly demanding disclosure requirements. Second, one could argue – drawing on a rich body of literature (see, e.g., Savolainen et al., 2023) – that such introspection is fraught with difficulties, thereby raising doubts about the capacity of programmers to reliably identify their own biases. While not trivialising these practical hindrances which de facto complicate the identification of biases, we contend, however, that the process ought to be understood in a ‘philosophy of science’ sense. In this respect, the proposed self-analysis resonates with the tradition (Bartlett, 1992; Collingwood, 1994) that calls for “honesty” and “vulnerability” (Erden, 2021) by the researcher (in this case, a programmer researching one’s biases) and turns the self-analysis into a “serious effort to understand […] the conditioning factors which influence and control the organisation of experience into a philosophy” (Gamertsfelder, 1928). To paraphrase Gamertsfelder (1928), it is, in essence, a research of the researcher: an effort to trace back to one’s own source code and identify the hidden bugs – the implicit biases and blind spots – that quietly structure how reality is interpreted and organised.
Once collected, ADRSs should be handled with strict confidentiality. Given the sensitivity of these statements, companies should implement safeguards to ensure anonymity, either by allowing programmers to submit ADRSs anonymously or by anonymising them post-collection. The HR department, or an independent ethics committee, should oversee this process to guarantee that these statements are used exclusively for internal purposes and do not lead to any form of discrimination. The risk of self-disclosure mechanisms being misused or leading to bias has been noted in the literature (Oswald, 2024) and would create a paradoxical situation where, to decrease bias, discrimination occurs.
Following their collection, ADRSs would need to be systematically analysed using both qualitative and quantitative methods. These analyses would identify patterns, clusters, and underlying epistemological or axiological trends at working groups or organisational level, thereby making second-level arbitrariness more visible. Visualisation tools, such as network graphs or heatmaps, could be employed to provide a clearer picture of how specific perspectives influence algorithmic design. The findings from these self-awareness exercises and analyses would then be compiled into an internal Algorithm designers’ reflexivity statements Report (ADRSs Report). Such an idea of merging multiple reflexivity statements into one single positionality statement is not novel, but it is quite common in social sciences articles written by multiple authors (Jamieson et al., 2023); as such, any argument about the possible infeasibility of such condensation should be pre-emptively silenced. Similarly, practical issues consisting of how to merge multiple (in certain cases hundreds or thousands) reflexivity statements could be solved by adapting the canonical techniques, such as consensus sessions (Jamieson et al., 2023), proper to the social sciences. The Report would synthesise key patterns, highlight critical biases, and propose corrective actions, such as affirmative design policies, diversity-oriented interventions (Floridi, 2024b; Peters, 2022), or modifications to internal design protocols. The ADRSs Report should then be shared with key internal stakeholders to facilitate discussions on bias awareness, encourage a feedback loop to refine the reflexivity process, and inform future iterations of the ADRS framework. As reflexivity is an ongoing process, scheduled reviews of ADRSs should be conducted to assess whether changes in company policies have led to shifts in ADRSs patterns and a debiasing of algorithms and AI systems.
However, this process would be incomplete if its results remained solely internal. To ensure transparency and accountability, we propose that organisations publish a public version of the ADRSs Report in the form of an AI positionality statement (AIPS). This statement would disclose to users any inherited, unavoidable and residual bias that emerged in the ADRSs Report or any other bias analyses (such as biased training data analyses or debiasing programmes). The openness about the algorithms or AI’s biases brought by an AIPS, is of fundamental importance for the users as they would be given a reading key to interpret – in light of the disclosed biases – the output provided by the algorithm or the AI: a strategy that aligns with recent calls to contextualise and “communicate the boundaries and context of [AIs] expertise” (Mahajan, 2025). Moreover, this openness also adds integrity, rigour, transparency and trustworthiness to the AI outputs – similar to the use of reflexivity statements in social sciences research (Jamieson et al., 2023).
Through the implementations of ADRSs, ADRSs Reports, and AIPSs, we believe tech organisations could take a significant step towards increasing bias awareness and detection. By adapting the well-established reflexivity practices from the social sciences, an ADRS would become an exercise or training of bias-awareness aimed at identifying and reducing inherent biases among programmers (Franke’s first and second strategy paths), but also introduce practical tools for mitigating bias at various stages of algorithm design. This, in turn, would result in AI systems that are more rigorous, transparent, and trustworthy – thus enabling their diffusion and adoption (Afroogh et al., 2024). Finally, by making these biases explicit to users, in the form of AIPSs, tech companies would empower users to make informed decisions (as per Franke’s (2022) third strategy path, i.e., “ensuring proper competition and consumer choice”) about the systems with which they interact.
Yet, words of caution are necessary. First, bias-transparency statements should not foster a false sense of security regarding bias mitigation. Disclosure does not equate to resolution, i.e., the mere acknowledgement of biases does not imply that they have been adequately addressed. Rather, such statements should be situated within a broader framework of bias mitigation and regarded as an initial step in a comprehensive strategy aimed at minimising (from the company side) and adapting (from the user side) to bias. Second, leaving aside the possibility of organisational resistance to such disclosures, it is possible to conceive a scenario where if bias transparency were to become a brand differentiator and a sought-after feature among users (as is the case with environmental performances (Delmas & Burbano, 2011)), a phenomenon akin to greenwashing could emerge, whereby companies capitalise on the hype without implementing substantive changes – thus adding to the growing list of washing phenomena (Gatti et al., 2025). This misalignment between “talk” and “walk” (Walker & Wan, 2012), which might be described as bias-transparency-washing, would involve companies disclosing AIPS primarily as an image-enhancing strategy – a form of window-dressing designed to signal bias awareness without embedding it in organisational practices and policies. Third, a process similar to the trivialisation of ethics, often criticised as “ethicswashing” (Metzinger, 2019) and “ethics bashing” (Bietti, 2020), could also materialise: here, bias transparency (through AIPS) would be valued not intrinsically but merely instrumentally, in terms of the benefits it yields for the company. Such developments risk reducing a philosophical endeavour aimed at fostering more rigorous and less biased algorithms and AIs into a narrow compliance exercise designed primarily to legally protect the company (for example, by shifting responsibility from organisations towards developers).
Conclusions
Despite the widespread prejudice that machines are (more) objective and unbiased, it is now widely established that algorithms-based systems like in AI (like large language models) do not only produce “incorrect, hallucinated and intentionally misleading content” (Nature Machine Intelligence, 2024) but also perpetuate and even amplify human biases and prejudices such as those related to race and gender (Hofmann et al., 2024) – a phenomenon often referred to as algorithmic bias. Given the increased presence of such systems in all aspects of life and research, “the need for curation and safeguarding of high-quality data and information is more crucial than ever” (Nature Machine Intelligence, 2024).
The prevailing explanations for bias and incorrect content in algorithms identify the training data as one source of the problem. The “biased training data” (Cowgill et al., 2020) framework asserts that “[t]he decisions made by AI are shaped by the initial data it receives. If the underlying data is unfair, the resulting algorithms can perpetuate bias, incompleteness, or discrimination”(Chen, 2023). As a result, current quality-securing strategies have been focusing on creating fairer datasets (Mittal et al., 2024), documenting their provenance (N. Vincent, 2024), conducting algorithmic audits (Morewedge et al., 2023), or establishing ethical governance frameworks for AI development. However, focusing solely on the problem of biased training data overlooks a critical dimension of the problem: bias in algorithms also stems from “biased programmers” (Cowgill et al., 2020). This perspective, which has unfortunately received comparatively less attention in the academic discourse, emphasises the fact that algorithms are written by people who – intentionally or unintentionally, consciously or unconsciously (Stinson, 2022) – inevitably embed their axiological, epistemological, and ontological assumptions (namely, their worldview) into codes.
In this paper, we addressed this overlooked perspective. Building on Franke’s (2022) ontological matrix of bias that distinguishes among first- and second-level discrimination, and arbitrariness, we analysed the mitigation practices in place to address these types of biases. After identifying a literature gap in how to address second-level arbitrariness, we proposed to adapt a well-established tool used in the social sciences to the field of AI. This is because qualitative research (especially ethnography and anthropology) has long acknowledged and grappled with the influence of researchers’ presence on their research: “nothing can be accomplished without subjectivity, so its elimination is not the solution. Rather how the subject is present is what matters, and objectivity itself is an achievement of subjectivity” (Giorgi, 1994). To address this issue, the field has developed reflexivity statements, i.e., declarations that document the long, continuous, rigorous introspective analysis in which researchers explore their “role and assumption[s] in knowledge production” (Knott et al., 2022). The epistemological benefit of reflexivity statements is that readers are provided with the tools necessary to contextualise the findings of the research in light of the declared potential biases. This approach fosters the urgent need for transparency (Floridi, 2024b) and enhances the quality of research by making explicit any factor that might have – unintentionally but inevitably – influenced the work. Thus, it increases the reliability and validity of inherently biased research that otherwise would not be possible to be conducted, as social realities are messy and biased. Reflexivity statements as instruments of methodological rigour, therefore, also add integrity and ethicality to research outcomes.
Inspired by this approach and considering the urgent “need for curation and safeguarding of high-quality data and information” (Nature Machine Intelligence, 2024), we proposed to adapt reflexivity and positionality statements in the form of ADRSs (algorithm designers’ reflexivity statements) and AIPSs (AI positionality statements).
Just as qualitative researchers disclose their positionality, we proposed that algorithm designers should engage in a reflexive process where they explore the assumptions, perspectives, and cultural biases they hold and that they may unintentionally and unconsciously introduce into their codes and designs. This introspective analysis should be articulated into a written personal statement – namely, an ADRS – where relevant axiological, epistemological, and ontological assumptions are disclosed. These statements are intended exclusively for internal use of awareness-creation and should be treated with confidentiality: this is to ensure a candid and proper self-assessment and avoid outing designers for their world views. We then suggested that once collected, the ADRS would need to be analysed to identify patterns and clusters of worldviews and values that might have skewed or could influence algorithms or their designs. The results of these analyses would then be summarised into an ADRSs Report that would be made available to relevant internal stakeholders to codevelop mitigating or correcting measures (such as affirmative design policies, diversity-oriented interventions, or modifications to internal design protocols).
Noting that this process would be incomplete if its results remained solely internal, to ensure transparency and accountability also to users of AI and algorithm systems – thereby countering the sense of opacity and incomprehension often experienced by “ordinary users who lack technical knowledge” (Shin, 2021) in relation to these increasingly complex ‘black boxes’ – we propose that companies publish a public version of the ADRSs Report in the form of an AI positionality statement (AIPS). This statement would disclose to users any inherited, unavoidable and residual bias that emerged in the ADRSs Report or any other bias analyses. AIPSs serve as a crucial interpretive tool for users to better understand and, more importantly, contextualise AI outputs in light of unintended residual biases.
Aware that ADRSs and AIPSs cannot fully eliminate bias from algorithms, by ensuring transparency, they can, nonetheless, represent a first meaningful step toward building “assurance and confidence” in AI (Shin, 2021) and mitigating second-level arbitrariness by making the invisible visible. These tools could be easily incorporated into current XAI governance as an expansion of the notions of transparency, interpretability and explainability from their technical aspect to a social one.
In an era where algorithm-based systems are increasingly influential and ubiquitous, solely addressing training data bias leaves aside the complexity of the human factor (coder bias) and its amplifying interactions. Hence, pairing current quality-securing training data-focused strategies with a people-centred strategy is not only advisable but necessary for the integrity and ethicality of AI. In this article, we proposed to fill this gap in the literature by bringing reflexivity into codes through algorithm designers’ reflexivity statements and by transposing positionality statements to AIs.
Limitations and future research
This paper advances a theoretical proposal to adapt, mutatis mutandis, established methodologies from the social sciences – specifically reflexivity and positionality – into the field of AI as a way to address the problem of programmer bias. The argument is, however, primarily conceptual. Empirical validation through pilot studies and systematic research will be required to assess its practical feasibility. One possible design could involve two coding teams: one employing ADRS and AIPS, and a control team that does not. Bias checks and public perception metrics could then be used to examine whether these tools reduce programmer bias or, at the very least, increase public confidence in AI trustworthiness. Such a study, however, would need to be carefully designed, given the non-replicable composition of human teams (and as such of their biases) and the difficulty of establishing reliable measures for normative issues. Further research should also explore whether users would engage meaningfully with algorithmic positionality statements (AIPS) in their evaluations and decision-making. Finally, the growing use of automated tools in AI development calls for an investigation into whether – and how – such reflexive practices can and should be applied in contexts where AI systems themselves increasingly generate or modify code as well as into the potential applicability of positionality statements for code authored by AI systems.
Acknowledgements
The authors are grateful to Jacopo M. Conti for the visual renderings of the two figures used in this article. The authors declare no funding.
Author contributions
LGC and PS contributed equally to this work. LGC and PS jointly conceived the research idea, developed the theoretical framework, conducted the analysis, and wrote the manuscript. LGC and PS have both read and approved the final version of the manuscript.
Data availability
No datasets were generated or analysed during the current study.
Competing interests
The authors declare no competing interests.
Ethical approval
This article does not contain any studies with human participants performed by any of the authors.
Informed consent
This article does not contain any studies with human participants performed by any of the authors.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
Afroogh, S; Akbari, A; Malone, E; Kargar, M; Alambeigi, H. Trust in AI: Progress, challenges, and future directions. Hum Soc Sci Commun; 2024; 11,
Ali, S; Abuhmed, T; El-Sappagh, S; Muhammad, K; Alonso-Moral, JM; Confalonieri, R; Guidotti, R; Del Ser, J; Díaz-Rodríguez, N; Herrera, F. Explainable Artificial Intelligence (XAI): What we know and what is left to attain Trustworthy Artificial Intelligence. Inf Fusion; 2023; 99, [DOI: https://dx.doi.org/10.1016/j.inffus.2023.101805] 101805.
Amini A, Soleimany AP, Schwarting W, Bhatia SN, Rus, D (2019) Uncovering and Mitigating Algorithmic Bias through Learned Latent Structure. Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, 289–295. https://doi.org/10.1145/3306618.3314243
Baeza-Yates, R. Bias on the web. Commun ACM; 2018; 61,
Barocas S, Selbst AD (2016) Big Data’s Disparate Impact (SSRN Scholarly Paper No. 2477899). https://doi.org/10.2139/ssrn.2477899
Barredo Arrieta, A; Díaz-Rodríguez, N; Del Ser, J; Bennetot, A; Tabik, S; Barbado, A; Garcia, S; Gil-Lopez, S; Molina, D; Benjamins, R; Chatila, R; Herrera, F. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Inf Fusion; 2020; 58, pp. 82-115. [DOI: https://dx.doi.org/10.1016/j.inffus.2019.12.012]
Bartlett SJ (Ed.) (1992) Reflexivity: A Source-Book in Self-Reference. In Reflexivity: A source-book in self-reference. North-Holland
Bellas, F; Naya-Varela, M; Mallo, A; Paz-Lopez, A. Education in the AI era: A long-term classroom technology based on intelligent robotics. Hum Soc Sci Commun; 2024; 11,
Bentham J (1791) Panopticon; or, The inspection house (http://books.google.com/books?id=NM4TAAAAQAAJ&oe=UTF-8). T.Payne; Internet archive. https://archive.org/details/panopticonorins00bentgoog/page/n2/mode/2up
Bernardo, V, Attoresi, M, Lareo, X, Velasco, L, & European Data Protection Supervisor (Eds). (2023). TechDispatch: Explainable artificial intelligence; #2/2023. Publications Office. https://doi.org/10.2804/802043
Bietti, E (2020). From ethics washing to ethics bashing: A view on tech ethics from within moral philosophy. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, 210–219. https://doi.org/10.1145/3351095.3372860
Binns, R. Algorithmic accountability and public reason. Philos Technol; 2018; 31,
Bjerring, JC; Busch, J. Artificial intelligence and patient-centered decision-making. Philos Technol; 2021; 34,
Bogen M (2019). All the Ways Hiring Algorithms Can Introduce Bias. Harvard Business Review. https://buy.tinypass.com/checkout/template/cacheableShow?aid=mbs77qtzUz&templateId=OTZUP68WMV81&templateVariantId=OTVOTAY63MK7N&offerId=fakeOfferId&experienceId=EXNL7T7RQ8ZR&iframeId=offer_f16488d3075278ec3421-0&displayMode=inline&debug=false&widget=template&url=https%3A%2F%2Fhbr.org
Bueter, A. Bias as an epistemic notion. Stud Hist Philos Sci; 2022; 91, pp. 307-315. [DOI: https://dx.doi.org/10.1016/j.shpsa.2021.12.002] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34955230]
Busso, LD; Leonardsen, A-CL. Qualitative research as reflexive process: A word limitation challenge in qualitative medical research publications?. Nordisk Sygeplejeforskning; 2019; 9,
Chen, Z. Ethics and discrimination in artificial intelligence-enabled recruitment practices. Hum Soc Sci Commun; 2023; 10,
Collingwood RG (1994) The Idea of History (van der Dussen, J Ed.; Revised edition). Oxford University Press
Courtland, R. Bias detectives: The researchers striving to make algorithms fair. Nature; 2018; 558,
Cowgill B, Dell’Acqua F, Deng S, Hsu D, Verma N, Chaintreau A (2020) Biased Programmers? Or Biased Data? A Field Experiment in Operationalizing AI Ethics (SSRN Scholarly Paper No. 3615404). https://doi.org/10.2139/ssrn.3615404
Crawford, K; Calo, R. There is a blind spot in AI research. Nature; 2016; 538,
Danks D (2024) Governance via explainability. In Bullock, JB, Chen, Y-C, Himmelreich, J, Hudson, VM, Korinek, A, Young, M & Zhang B (Eds), The Oxford handbook of AI governance (1st edn, pp. 183–197). Oxford University Press
Das A, Rad P (2020) Opportunities and Challenges in Explainable Artificial Intelligence (XAI): A Survey (No. arXiv:2006.11371). arXiv. https://doi.org/10.48550/arXiv.2006.11371
de Laat, PB. Algorithmic decision-making based on machine learning from big data: can transparency restore accountability?. Philos Technol; 2018; 31,
Delmas, MA; Burbano, VC. The drivers of greenwashing. Calif Manag Rev; 2011; 54,
Downing L (2010) The Cambridge introduction to Michel Foucault (3. print). Cambridge University Press
Dwivedi, R; Dave, D; Naik, H; Singhal, S; Omer, R; Patel, P; Qian, B; Wen, Z; Shah, T; Morgan, G; Ranjan, R. Explainable AI (XAI): Core Ideas, Techniques, and Solutions. ACM Comput Surv; 2023; 55,
Erden, YJ. Identity and bias in philosophy: what philosophers can learn from STEM subjects. Think; 2021; 20,
Fan, N. Explainability is necessary for AI’s trustworthiness. Philos Technol; 2025; 38,
Finlay, L. “Outing” the researcher: the provenance, process, and practice of reflexivity. Qualit Health Res; 2002; 12,
Floridi, L. Introduction to the special issues: the ethics of artificial intelligence: exacerbated problems, renewed problems, unprecedented problems. Am Philos Q; 2024; 61,
Floridi L (2024b) The Ethics of Artificial Intelligence: Exacerbated problems, renewed problems, unprecedented problems - Introduction to the Special Issue of the American Philosophical Quarterly dedicated to The Ethics of AI (SSRN Scholarly Paper No. 4801799). https://doi.org/10.2139/ssrn.4801799
Floridi, L; Cowls, J; Beltrametti, M; Chatila, R; Chazerand, P; Dignum, V; Luetge, C; Madelin, R; Pagallo, U; Rossi, F; Schafer, B; Valcke, P; Vayena, E. AI4People—An Ethical Framework for a Good AI Society: opportunities, risks, principles, and recommendations. Minds Mach; 2018; 28,
Foucault, M (1995). Discipline and punish: The birth of the prison (A. Sheridan, Trans.; Second Vintage Books edition). Vintage Books, a division of Random House, Inc
Franke, U. First- and second-level bias in automated decision-making. Philos Technol; 2022; 35,
Gamertsfelder, WS. The nature of philosophical impartiality. Philos Rev; 1928; 37,
Gatti L, Conti LG, Seele P (2025) A Comprehensive Guide to Greenwashing Phenomena, Contexts, and Trends: The Mean, Lean Washing Machine (1st edn). Edward Elgar Publishing. https://www.e-elgar.com/shop/gbp/a-comprehensive-guide-to-greenwashing-phenomena-contexts-and-trends-9781035328239.html?srsltid=AfmBOopv2AfqrKeISpb1Qk_7707u0HxfQo61vWpyFn4Lahwq5hnIcXeA
Gebru T, Morgenstern J, Vecchione B, Vaughan JW, Wallach H, III, HD, & Crawford K (2021) Datasheets for Datasets (No. arXiv:1803.09010). arXiv. https://doi.org/10.48550/arXiv.1803.09010
Gelman A, Loken E (2013) The garden of forking paths: Why multiple comparisons can be a problem, even when there is no “fishing expedition” or “p-hacking” and the research hypothesis was posited ahead of time
Gichoya, JW; Thomas, K; Celi, LA; Safdar, N; Banerjee, I; Banja, JD; Seyyed-Kalantari, L; Trivedi, H; Purkayastha, S. AI pitfalls and what not to do: Mitigating bias in AI. Br J Radiol; 2023; 96,
Giorgi, A. A phenomenological perspective on certain qualitative research methods. J Phenom Psychol; 1994; 25,
Grix J (2018) The Foundations of Research (3rd edition). Red Globe Pr
Gunning, D; Stefik, M; Choi, J; Miller, T; Stumpf, S; Yang, G-Z. XAI—Explainable artificial intelligence. Sci Robot; 2019; 4,
Hofmann, V; Kalluri, PR; Jurafsky, D; King, S. AI generates covertly racist decisions about people based on their dialect. Nature; 2024; 633,
Holmes, AGD. Researcher Positionality—A consideration of its influence and place in qualitative research—A New Researcher Guide. Shanlax Int J Educ; 2020; 8,
Jamieson, MK; Govaart, GH; Pownall, M. Reflexivity in quantitative research: A rationale and beginner’s guide. Soc Personal Psychol Compass; 2023; 17,
Johansen, J; Pedersen, T; Johansen, C. Studying human-to-computer bias transference. AI SOCIETY; 2023; 38,
Jordan, MI; Mitchell, TM. Machine learning: Trends, perspectives, and prospects. Science; 2015; 349,
Kelly T (2022) Bias: A Philosophical Study (1st edn). Oxford University PressOxford. https://doi.org/10.1093/oso/9780192842954.001.0001
Kirikkaleli, D; Aad, S; Kirikkaleli, NO. Sustainable development and investment in artificial intelligence in the USA. Hum Soc Sci Commun; 2025; 12,
Knott, E; Rao, AH; Summers, K; Teeger, C. Interviews in the social sciences. Nat Rev Methods Prim; 2022; 2,
Kroll, J; Huey, J; Barocas, S; Felten, E; Reidenberg, J; Robinson, D; Yu, H. Accountable Algorithms. Univ Pa Law Rev; 2017; 165,
Lazard, L; McAvoy, J. Doing reflexivity in psychological research: What’s the point? What’s the practice?. Qualit Res Psychol; 2020; 17,
Lum, K; Isaac, W. To predict and serve?. Significance; 2016; 13,
Mac R (2021). Facebook Apologizes After A.I. Puts ‘Primates’ Label on Video of Black Men. The New York Times. https://www.nytimes.com/2021/09/03/technology/facebook-ai-race-primates.html
Mahajan, S. The democratization dilemma: When everyone is an expert, who do we trust?. Hum Soc Sci Commun; 2025; 12,
Malterud, K. Qualitative research: Standards, challenges, and guidelines. Lancet; 2001; 358,
Massoud, MF. The price of positionality: Assessing the benefits and burdens of self-identification in research methods. J Law Soc; 2022; 49,
Metzinger T (2019). EU guidelines: Ethics washing made in Europe. Der Tagesspiegel. https://www.tagesspiegel.de/politik/eu-guidelines-ethics-washing-made-in-europe/24195496.html
Mittal, S; Thakral, K; Singh, R; Vatsa, M; Glaser, T; Canton Ferrer, C; Hassner, T. On responsible machine learning datasets emphasizing fairness, privacy and regulatory norms with examples in biometrics and healthcare. Nat Mach Intell; 2024; 6,
Morewedge, CK; Mullainathan, S; Naushan, HF; Sunstein, CR; Kleinberg, J; Raghavan, M; Ludwig, JO. Human bias in algorithm design. Nat Hum Behav; 2023; 7,
Nature. More accountability for big-data algorithms. Nature; 2016; 537,
Nature Machine Intelligence. Pick your AI poison. Nat Mach Intell; 2024; 6,
Nozick R (1994) The Nature of Rationality (First Paperback Printing edition). Princeton University Press
Obermeyer, Z; Powers, B; Vogeli, C; Mullainathan, S. Dissecting racial bias in an algorithm used to manage the health of populations. Science; 2019; 366,
Oswald F (2024) Positionality statements should not force us to ‘out’ ourselves. Nat Hum Behav, 8. https://doi.org/10.1038/s41562-023-01812-5
Peters, U. Algorithmic political bias in artificial intelligence systems. Philos Technol; 2022; 35,
Prasad, KDV; De, T. Generative AI as a catalyst for HRM practices: Mediating effects of trust. Hum Soc Sci Commun; 2024; 11,
Roberts, M; Driggs, D; Thorpe, M; Gilbey, J; Yeung, M; Ursprung, S; Aviles-Rivero, AI; Etmann, C; McCague, C; Beer, L; Weir-McCall, JR; Teng, Z; Gkrania-Klotsas, E; Rudd, JHF; Sala, E; Schönlieb, C-B. Common pitfalls and recommendations for using machine learning to detect and prognosticate for COVID-19 using chest radiographs and CT scans. Nat Mach Intell; 2021; 3,
Rowe WE (2014) Positionality. In Coghlan D & Brydon-Miller M (Eds), The SAGE encyclopedia of action research (pp. 627–628). SAGE Publ
Rudin C, Wang C, Coker B (2020) The age of secrecy and unfairness in Recidivism Prediction. Harvard Data Sci Rev, 2(1). https://doi.org/10.1162/99608f92.6ed64b30
Savolainen, J; Casey, PJ; McBrayer, JP; Schwerdtle, PN. Positionality and its problems: questioning the value of reflexivity statements in research. Perspect Psychol Sci; 2023; 18,
Seele, P. Envisioning the digital sustainability panopticon: A thought experiment of how big data may help advancing sustainability in the digital age. Sustain Sci; 2016; 11,
Shaw, R. Embedding reflexivity within experiential qualitative psychology. Qualit Res Psychol; 2010; 7,
Shin, D. The effects of explainability and causability on perception, trust, and acceptance: Implications for explainable AI. Int J Hum -Comput Stud; 2021; 146, [DOI: https://dx.doi.org/10.1016/j.ijhcs.2020.102551] 102551.
Sjoding, MW; Dickson, RP; Iwashyna, TJ; Gay, SE; Valley, TS. Racial bias in pulse oximetry measurement. N. Engl J Med; 2020; 383,
Stinson, C. Algorithms are not neutral. AI Ethics; 2022; 2,
Sybing, R. Dead reckoning: A framework for analyzing positionality statements in ethnographic research reporting. Writ Commun; 2022; 39,
Vasilogambros M, Journal, N (2015). Will Your Facebook Friends Make You a Credit Risk? The Atlantic. https://www.theatlantic.com/politics/archive/2015/08/will-your-facebook-friends-make-you-a-credit-risk/432504/
Vincent J (2018). Amazon reportedly scraps internal AI recruiting tool that was biased against women. The Verge. https://www.theverge.com/2018/10/10/17958784/ai-recruiting-tool-bias-amazon-report
Vincent, N. A step forward in tracing and documenting dataset provenance. Nat Mach Intell; 2024; 6,
Walker, K; Wan, F. The harm of symbolic actions and green-washing: corporate actions and communications on environmental performance and their financial implications. J Bus Ethics; 2012; 109,
Walsh, RA. The approach of the human science researcher: Implications for the practice of qualitative research. Humanist Psychol; 1995; 23,
Wang, Q; Sun, T; Li, R. Does Artificial Intelligence (AI) enhance green economy efficiency? The role of green finance, trade openness, and R&D investment. Hum Soc Sci Commun; 2025; 12,
Wang, X; Xu, X; Zhang, Y; Hao, S; Jie, W. Exploring the impact of artificial intelligence application in personalized learning environments: Thematic analysis of undergraduates’ perceptions in China. Hum Soc Sci Commun; 2024; 11,
Wilkinson, S. The role of reflexivity in feminist psychology. Women’s Stud Int Forum; 1988; 11,
Willig C (2013) Introducing Qualitative Research In Psychology. McGraw-Hill Education (UK)
Zhao, H-G; Li, X-Z; Kang, X. Development of an artificial intelligence curriculum design for children in Taiwan and its impact on learning outcomes. Hum Soc Sci Commun; 2024; 11,
Ziosi, M; Watson, D; Floridi, L. A genealogical approach to algorithmic bias. Minds Mach; 2024; 34,
Zuiderveen Borgesius F (2018) Discrimination, Artificial Intelligence, and Algorithmic Decision-Making (p. 91). Directorate General of Democracy, Council of Europe
© The Author(s) 2025. This work is published under http://creativecommons.org/licenses/by-nc-nd/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.