Algorithmic Fairness Testing for Inclusive Growth: Bridging Theory and Industrial Practice

 Corresponding author: Yuliia Baranetska; ORCID: 0009-0008-9412-0420 Copyright © 2025 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution License 4.0. Algorithmic Fairness Testing for Inclusive Growth: Bridging Theory and Industrial Practice Yuliia Baranetska * Independent researcher, Kyiv, Ukraine. Global Journal of Engineering and Technology Advances, 2025, 25(01), 215-228 Publication history: Received on 18 September 2025; revised on 25 October 2025; accepted on 27 October 2025 Article DOI: https://doi.org/10.30574/gjeta.2025.25.1.0315 Abstract The rapid growth of Artificial Intelligence systems across industries necessitates robust processes to ensure their ethical and fair use. A critical aspect of AI development is algorithmic fairness, which addresses potential biases that can lead to discrimination. This study bridges the gap between theoretical concepts of fairness and their practical application in the industry. It examines fairness measures, bias detection, and mitigation throughout the AI lifecycle, including data collection, model training, and ongoing monitoring. Additionally, it assesses the challenges and opportunities in implementing fairness testing in industrial pipelines, considering factors like computational costs and regulatory compliance. By integrating scholarly research with practical implementation plans, this paper offers actionable insights for researchers, practitioners, and policymakers to create equitable AI systems. It positions algorithmic fairness testing as a means for global prosperity by mitigating algorithmic harms in critical sectors. The study promotes inclusive growth, public trust, and equitable access to AI-enabled services through governance-aware practices tailored for various organizational contexts. Keywords: Algorithmic Fairness; Bias Audit; Equalized Odds; Inclusive Growth; Governance and Policy; Global Prosperity 1. Introduction Artificial Intelligence (AI) systems have gained a quick presence in the healthcare sector and the financial industry, offering automation, improved decision-making processes, and innovative solutions. Nonetheless, historical developments are associated with an imminent ethical dilemma: how to make AI systems fair and not contribute to the intensification of disparities in society. One of its fundamental issues is algorithmic fairness where AI models, which are typically trained based on past data, may replicate and increase their biases and yield discriminative results. This is especially a problem in high-stakes settings such as hiring, lending, and criminal justice, where AI-generated biases can have a significant impact on people’s lives [5]. The most important of these issues is the need of testing fairness in Artificial Intelligence, which can be defined as assessing, detecting, and reducing biases that may occur in the data, algorithm, or output. Although theoretical research on fairness has advanced significantly, there has been a gap in applying fairness theories in the real world [18]. The conceptualizations of fairness tend to take a variety of forms according to the theoretical frameworks based on fairness, whether in an individual form, where like persons should be treated like, or in a group form, where all demographic groups should be treated equally. However, these definitions do not always agree, and there is a trade-off in terms of implementation [8, 19]. Global Journal of Engineering and Technology Advances, 2025, 25(01), 215-228 216 This study aims to investigate the topography of fairness testing in AI, considering both the theory behind fairness and the issues that may arise when attempting to implement fairness in industries [13]. This paper starts by discussing the definitions of fairness and an overview of the ethical principles of fairness testing in AI. It also investigates fairness testing methodologies, such as adversarial testing, bias audits, and fairness-aware learning [17]. Finally, this paper discusses the obstacles to the implementation of fairness testing in the industry and how these issues can be overcome to promote its use in AI systems [28]. Table 1 provides a comparison of the two main fairness definitions used in AI systems. Table 1 Comparison of Fairness Definitions Definition Description Examples Key Metrics Individual Fairness Treat similar individuals similarly. AI in recruitment. Consistency, Similarity Group Fairness Treat groups equally in outcomes. AI in lending or hiring. Demographic Parity, Equalized Odds 2. Methodology This study employs a conceptual qualitative approach, informed by a recent literature review covering the years 2016 to 2025 [10]. We also analyze industrial case reports in the fields of healthcare, employment, and credit scoring [26]. Our work synthesizes various metrics, including demographic parity, equalized odds, and disparate impact, and evaluates testing methodologies such as adversarial techniques, audits, and fairness-aware learning [1]. Additionally, we examine governance practices to provide operational guidance [27]. This methodology is designed to be applicable across diverse organizational contexts and accommodates varying levels of regulatory maturity. 3. Theoretical Foundations of Fairness in AI 3.1. Defining Fairness in AI In the case of AI, the notion of fairness can be perceived differently based on the ethics one is focused on [25]. These definitions are based on various philosophical perspectives and may interfere with each other in practice. 3.1.1. Individual Fairness It is a model of fairness that states that an AI system should treat similar people in a similar manner. This is especially applicable in sensitive activities such as hiring or diagnosing a medical condition where the consideration of justice to every individual is essential [12]. The problem is that the concept of similarity is subjective and may be difficult to introduce in the case of big datasets. 3.1.2. Group Fairness It is a method based on treating various groups of people (racial or gender groups) equally. Group fairness has been applied across a variety of areas such as loan application or criminal justice, which ought to ensure that diverse demographic categories have an equal chance to utilize AI systems [5]. In this model, such metrics as demographic parity (the different groups must have equal acceptance in a particular process) or equalized odds (the different groups must be equal in terms of their performance in a given process) are frequently taken into account. These definitions provide divergent positions on fairness as an individual fairness focuses on how people are treated whereas group fairness lays more emphasis on what happens to demographic groups [9]. When these models are extended to real world data they tend to clash with each other and in this, practitioners are faced with hard decisions to make on which type of fairness to give precedence. 3.2. Fairness Metrics In order to assess fairness, there are a number of measures used to assess the ability of the AI systems to treat people or groups fairly [6]. The most widely applied measures of fairness are: Global Journal of Engineering and Technology Advances, 2025, 25(01), 215-228 217 3.2.1. Demographic Parity This is to guarantee that the percentage of positive results is the same among the various demographic groups (e.g., gender, race). Although this measure is easy and simple to calculate, it may also cause a lack of efficiency in the accuracy of the models since it may fail to capture the individual variance between cases [12]. 3.2.2. Equal Opportunity The measure is aimed at making sure that no particular groups have a disadvantage in terms of getting positive results. It specifically focuses on the true positive rate, i.e. people who belong to other groups must have equal opportunity to be picked to have a positive result [11]. 3.2.3. Equalized Odds In this statistic, the equal opportunity is extended in the sense that the true positive rate as well as the false positive rate are equal between groups. This will counter any systemic biasness in the model predictions. 3.2.4. Disparate Impact It is a metric most frequently used in legal practice but is used to select whether the AI model is unfair to a group that cannot be explained by the results. When the probability of a group of people to receive positive outcomes will be discovered as much lower than another group of people, having similar qualifications or data points, a disparate impact is identified [5]. Figure 1 illustrates the different fairness metrics and their applications in AI systems. Figure 1 Overview of Fairness Metrics in AI Systems A diagram illustrating different fairness metrics (e.g., Demographic Parity, Equal Opportunity, Equalized Odds, Disparate Impact) and their use cases in AI applications. 3.3. Ethical Foundations of Fairness in AI Justice and equity theories play a central role in determining the ethical foundations of fairness in AI [20]. The two philosophical approaches to this situation are: Global Journal of Engineering and Technology Advances, 2025, 25(01), 215-228 218 3.3.1. Theory of Justice as propounded by Rawls John Rawls developed principles of fairness that guarantee maximum good to the least privileged in the society a theory he described as the difference principle. This may imply the creation of systems that give equal opportunities to the marginalized communities at the cost of efficiency in certain situations in AI [22]. 3.3.2. Utilitarianism Utilitarian ethics on the other hand aim to maximize the total benefits of the societal good, which can be against the fairness of the individual since the result of maximizing the benefits of the majority might end up in marginalization of certain groups [3]. This moral theory can be used in the development of AI tools that aim to achieve the utmost social good but at the same time cause inequitable results to the small underprivileged groups (Bentham, 1789). These ethical frameworks provide guiding principles for analyzing AI fairness. Whereas the theory introduced by Rawls emphasizes equity, utilitarianism prioritizes efficiency, posing a conflict of equities. Table 2 summarizes the relationship between fairness metrics and their ethical foundation. Table 2 Comparison of Fairness Metrics and Ethical Foundations Fairness Metric Focus Ethical Framework Application Example Demographic Parity Equal representation of groups Rawls' Theory of Justice Employment recruitment Equal Opportunity Equal chances of positive outcomes Rawls' Theory of Justice Loan approvals Equalized Odds Equal true/false positive rates across groups Utilitarianism (maximize overall good) Criminal justice sentencing Disparate Impact Equal outcomes (selection rates) across groups Anti-discrimination / deontic fairness Hiring, lending, education 4. Fairness testing methodologies One of the most important steps towards ensuring that AI systems do not reinforce harmful biases is fairness testing, both deliberately and accidentally [14]. It entails various methodologies that evaluate the extent of observance of an AI system to fairness principles, including individual fairness and group fairness. Such testing strategies are referred to as adversarial testing, bias audits, fairness-conscious learning, or the application of fairness measures to evaluate and reduce bias at any stage of the AI lifecycle [2]. 4.1. Tests in Adversarial Fairness Adversarial fairness testing is a method used to test whether an Artificial Intelligence system is affected by sensitive factors, such as race, gender, or age, in its prediction in a disproportionate manner. This is aimed at testing the worstcase scenarios, to identify undiscovered biases in the AI models that might not be apparent in real life. Perturbations are introduced to the input data, and the built predictions are observed in the framework of adversarial testing. This is a popular testing technique in image classification models and models of natural language processing [23]. To illustrate, during facial recognition, adversarial inputs may be to manipulate images of people of other races to determine whether the model is biased against some population. One of them is the adversarial de-biasing technique, which involves adversarial training that eliminates bias in the model prediction (Zhao et al., 2018). The procedure is effective in that it produces counter-example scenarios that expose biased behavior and trains the model to counteract bias in learning. Figure 2 presents the framework used for adversarial fairness testing. Global Journal of Engineering and Technology Advances, 2025, 25(01), 215-228 219 Figure 2 Adversarial Fairness Testing Framework Diagram illustrating the adversarial fairness testing framework, showing input manipulation, model prediction, and evaluation for bias detection. 4.2. Bias Audits Bias audits are a methodological procedure that assesses AI models with a view to uncovering and estimating biases at various phases of the AI lifecycle [4]. Such audits also assist in discovering whether AI system is discriminating or benefiting some groups unintentionally depending on the sensitive characteristics. Audits of bias usually include three following steps: 4.2.1. Data Audit The initial one is to evaluate the data that will be utilized in the training of the AI model. A bias audit begins by checking the data set on whether there are any imbalances or representational problems. To illustrate, in the case where a dataset of the recruitment algorithm disproportionately represents female candidates in highly skilled job categories, it is more likely to give preference to the male candidates. 4.2.2. Model Audit The second stage will concern the analysis of the behaviour of the trained AI model when predicting the behaviour of various demographics. This may include evaluating the model using a set of subpopulations (e.g., race, gender) in terms of accuracy, precision, recall, and fairness. 4.2.3. Outcome Audit Lastly, the outcome audit is an assessment of the end decisions or projections of the AI system. This step verifies the disproportionate effect of the output of the model to some groups. An example is an outcome audit to determine whether an AI-based credit scoring system is discriminatory against the applicants of a particular race or ethnic group. Bias audits are effective mechanisms of identifying problems of fairness, although they are also costly to resources and may be a matter of privacy concerns because they frequently demand access to sensitive demographic data. Furthermore, audits cannot always be used to give solutions to do away with biases, they just assist in revealing them. Table 3 outlines the three stages of bias auditing in AI systems. Table 3 Steps in Bias Auditing AI Systems Audit Stage Focus Objective Example Data Audit Evaluate dataset for representational bias Ensure diversity and inclusivity in data Gender balance in hiring dataset Model Audit Assess model's performance across groups Identify disparities in accuracy or fairness Loan approval model's fairness to different racial groups Outcome Audit Evaluate model predictions' impacts Ensure fair outcomes for all groups Discriminatory impact predictive policing system Global Journal of Engineering and Technology Advances, 2025, 25(01), 215-228 220 4.3. Fairness-Aware Learning Fairness-conscious learning aims to adjust machine learning processes or algorithms so that the output model can meet requirements of fairness [15]. In contrast to the classical machine learning that focuses on optimizing the accuracy, the fairness-conscious learning is directly aimed at fairness constraints in the learning, in the way that fairness is not compromised at the price of the model performance. 4.3.1. Fairness-aware learning may be done in three major ways Pre-processing: This method involves the alteration of the data prior to the training of the model. This may include reweighting the underrepresented groups, over-Sampling of data of specific groups, or data augmentation methods that minimize bias prior to training. As an illustration, credit scoring could have a bias in the way the data is handled, whereby historically marginalized populations have a more proportional representation. In-processing: It involves the inclusion of fairness constraints in the model training. The algorithm applied in the learning process is effectively meant to reduce bias in the optimization. A typical method is to add a term of fairness to the loss objective, thus punishing a model that makes biased predictions. The model can also be trained to achieve accuracy and fairness at the same time, in certain cases. Post processing: Once the model has been trained the predictions made by the model are modified to provide fairness. This method is applied in cases when the model itself can not be changed. The post-processing may include the use of thresholds to make the decision, e.g., recalibration of the model, so that the proportions of positive predictions within various demographic groups should remain balanced. Techniques that are fairness-conscious are capable of minimizing bias, however, they are frequently based on a tradeoff. Although they can help increase fairness, they can also lead to higher inaccuracy especially in a situation where the fairness constraints are hard to balance with model performance [24]. Figure 3 depicts the complete fairness-aware learning pipeline with all three processing stages. Figure 3 Fairness-Aware Learning Pipeline Diagram illustrating the fairness-aware learning pipeline, showing pre-processing, in-processing, and post-processing stages to ensure fairness in AI systems. 4.4. Fairness Metrics in Practice The fairness measure is a vital decision in a fairness testing approach. The metrics yield different information and their choice will depend on ethical objectives of the system, and the circumstances of implementing the metrics. The most widespread measures of fairness are demographic parity, equalized odds, and disparate impact, each of which has a different objective to maintain fairness. To illustrate, the situation would need demographic parity, which could be the most applicable theory when the objective is to make sure that various groups (e.g., gender, race) are equally represented when making hiring choices. Equality of odds, conversely, might be more suitable where it is desired to provide equal true positive rates between groups, as in credit scoring or medical diagnostics. Global Journal of Engineering and Technology Advances, 2025, 25(01), 215-228 221 These metrics are necessary but they are normally conflicting with each other. It is also possible that the maximization of demographic parity would result into inaccuracy or prejudice of individual fairness, whereas equalized odds can result in a compromise of other fairness objectives. Thus, one must make sure that the ethical trade-offs are taken into consideration when choosing the appropriate metric. Table 4 provides a detailed comparison of the most common fairness metrics used in practice. Table 4 Comparison of Common Fairness Metrics Metric Goal Common Use Case Strengths Weaknesses Demographic Parity Equal representation across groups Hiring, Loan Approval Simple to calculate, easy to implement Can reduce model accuracy Equalized Odds Equal true positive and false positive rates Criminal Justice, Healthcare Reduces bias in prediction errors Conflicts with other fairness metrics Disparate Impact Equal outcomes across groups Employment, Education Easy to measure and understand May overlook group-specific nuances 5. Industry Adoption of Fairness Testing Although it is evident that there are ethical requirements of inequity in AI, industries still encounter numerous challenges in implementing fairness testing in AI systems [7]. The mismatch between theory and practice can result in creating a divide of uncertainty in companies because of concerns relating to costs, regulation, and the difficulty of adding fairness to current AI pipelines. In this section, the researcher will delve into the issues that industries encounter when embracing fairness testing as well as the strategies that have been implemented in addressing the problems. 5.1. Barriers to Adoption Some of the difficulties involved in the implementation of fairness testing in AI systems are not insignificant. The unwillingness of industries to have such practices embraced within them is attributed to a number of factors. 5.1.1. Absence of standardization The absence of standard measures and procedures is one of the major obstacles to the implementation of fairness testing. No single framework or set of best practices can be used by the companies to make their AI models fair. As it was mentioned in the previous section, various fairness metrics tend to contradict each other and the correct choice of a metric is dependent on the context that the given application will be implemented. There is no standardized methodology thus companies can hardly put fairness testing into effect. 5.1.2. Computational Costs Fairness testing particularly when applied in large scale systems may be computationally expensive. The process of developing a test may be complicated and expensive because it may require large amounts of computational resources to implement testing of fairness in multiple metrics and groups. These extra costs may prove to be a strong discouragement to businesses whose main goal is to make the best use out of the resources and to save as much as possible. 5.1.3. Regulatory Compliance Various industries have regulatory standards mandating transparency and accountability in systems based on AI, however, these standards are commonly ambiguous or undeveloped as far as fairness is concerned. The absence of explicit legal provisions on the issue of fairness in AI complicates the situation of companies operating in response to the law. Such ambiguity may contribute to reluctance to observe the practice of fairness testing since a company is afraid of being sued or failing to comply with new rules. Global Journal of Engineering and Technology Advances, 2025, 25(01), 215-228 222 5.1.4. Resistance in the Organization The organizational resistance to fairness testing could be connected to the trade-offs between fairness and performance as perceived. Organizations tend to focus on the model accuracy and efficiency particularly in competitive fields such as finance and health care. Fairness constraints may decrease the accuracy or may consume more time in model training, and is not readily marketable to profit maximizing companies. Moreover, the introduction of fairness testing can be costly in workflow reengineering, and that can be organizational counter-intuitive. 5.2. Regulatory Considerations and Frameworks Regulatory authorities have been establishing guidelines and frameworks on AI systems as ethical and fairness concerns about AI have increased. Nevertheless, the laws on fairness in AI are in development. The important advances of AI regulation are: 5.2.1. The EU AI Act The European Union has been on the forefront to regulate AI ethics through the AI Act which includes provisions on systems of high risk, transparency, accountability, and fairness. Another point made by the Act is that AI systems must be assessed in terms of risks and monitored, as well as fairness and nondiscrimination. It however does not give a complete framework of testing fairness in itself hence leaves to interpretation. 5.2.2. U.S. Federal Guidelines The fairness of AI has increasingly become a topic of concern in the U.S., especially in criminal justice and the health care industry. In 2019, the Algorithmic Accountability Act was introduced, which suggests that organizations should reveal information about the automated decision-making systems and assess the effects they are likely to have on fairness, privacy, and bias. The bill however is still pending and there is still no standard regulatory framework of fairness testing in AI systems throughout the U.S. 5.2.3. The ISO Standards on AI ISO, the International Organization on Standards, is busy formulating standards of AI, including fairness. These principles seek to help industries develop more equitable AI, but these are still being developed and are yet to be enforced. Regulatory ambiguity regarding fairness in AI has been a problem even with these efforts that are aimed at adopting the industry. It is possible that companies will not invest in fairness testing because they do not know whether such testing will be required in the future. 5.3. Real-World Applications and Case Studies Even with the obstacles, certain industries have gone a long way in ensuring that fairness testing is incorporated in their AI systems. This section examines a few practical instances where fairness testing has achieved its adoption. 5.3.1. Healthcare AI applications in healthcare like diagnostic algorithms and patient outcome predictive tools have been questioned with regards to possible biases in data and future outcomes. To this end, healthcare providers and companies are adding fairness audits to their AI models to guarantee that they do not discriminate against minority groups. Indicatively, scholars have created fairness measures to assess AI models applied to medical imaging, which guarantees that the models are equally effective when used on various ethnic populations [19]. 5.3.2. Hiring and Recruitment AI is also becoming a common tool to filter job applicants, yet the models may continue with biases in the hiring process. Fairness testing has been introduced in recruitment tools using AI in companies such as HireVue and Unilever. Such companies have fairness-conscious learning methods that maintain that their algorithms are not discriminating against applicants on the basis of gender, ethnicity, or socioeconomic status. Also, they do bias audits on a regular basis and make sure that the results of their recruitment algorithms are fair. Global Journal of Engineering and Technology Advances, 2025, 25(01), 215-228 223 5.3.3. Criminal Justice AI systems have also been applied to criminal justice, including the COMPAS system which was used to estimate the risk of recidivism. Nevertheless, issues of racial bias in these models have been raised and this has prompted moves to administer bias audit and enhance equity in such systems [21]. Different organizations and governments have started using fairness testing features within risk assessment algorithms in order to avoid such situations so that the marginalized populations are not disproportionately harmed by them. 5.4. Overcoming Organizational Resistance Organizational resistance is one of the biggest problems to overcome in implementing fairness testing in AI. A high profitability and efficiency are valued by many companies, and fairness testing is considered as an extra burden that may affect the accuracy of a model or slack down the process of its development. To be able to overcome this resistance, there has to be leadership buy-in and a communication of the long-term benefits of fairness [19]. 5.4.1. Leadership Buy-in The achievement of fairness testing is usually unsuccessful unless the top management is strongly supportive. Whenever the organizational leaders make fairness one of their AI strategies, it creates a strong message that AI systems should be ethically acceptable. An example is Microsoft, which has introduced a fairness testing under its corporate ethics program, which shows it as a responsible AI company. 5.4.2. Training of the Employees Training of their teams on the significance of fairness in AI should be done by the companies. This involves educating data scientists, AI developers and ethicists on means of bringing fairness into their practice. One of the necessary stages on the way to the establishment of a culture of fair AI development is the provision of workshops and resources related to the methodology of fair testing. 5.4.3. Rewarding Fairness By making fairness a central performance indicator among AI development teams, it may be possible to incentivize the practice of fairness testing. Rewards or recognition on attaining fairness milestones can add incentive to the teams to be more concerned about ensuring fairness in their work, despite making trade-offs with accuracy. 5.5. Policy and Governance Implications for Global Prosperity We translate fairness tooling into governance routines organizations can operationalize irrespective of jurisdictional maturity: (i) lightweight bias audits aligned with public‑interest outcomes (e.g., equal opportunity in approvals), (ii) SME‑ready fairness‑by‑design checklists spanning pre‑/in‑/post‑processing with run‑time dashboards, (iii) continuous monitoring with thresholds tied to stakeholder harm and escalation playbooks, and (iv) transparent reporting that supports non‑discrimination goals and public trust. This package lowers adoption cost while improving access and accountability in high‑stakes services, contributing to inclusive growth. 6. Challenges and Limitations of Fairness Testing Fairness testing is crucial for identifying and reducing bias in AI systems but faces several challenges, including data quality issues, ethical conflicts, trade-offs between fairness and performance, and privacy concerns [16]. Addressing these issues is essential for creating AI systems that are both effective and fair. 6.1. Data Bias and Data Quality Data bias poses a major challenge in fairness testing. AI models learn from the data they are trained on, and if this data contains past biases or systemic disparities, the model will replicate these issues. For instance, if the training data for an AI hiring tool is skewed towards males or a specific racial group, the system may favor these groups, leading to discriminatory hiring practices. The bias of data may be in a number of forms