Unbiased AI: Tackling Algorithmic Discrimination

Researcher, TRENDS Research & Advisory
January 29, 2024

Introduction

Artificial Intelligence (AI) has revolutionized many aspects of our lives, from healthcare to education, finance, and transportation. However, the emergence of algorithmic discrimination in AI presents a significant challenge to fairness, equality, and societal justice. Algorithmic discrimination occurs when AI systems perpetuate and amplify biases and discriminatory practices, leading to unequal treatment and outcomes for different groups. This insight delves into the multifaceted aspects of unbiased AI, exploring current challenges, solutions, and the future path towards ensuring that AI systems are equitable and non-discriminatory.

Specifically, this insight seeks to answer the following questions: What is algorithmic discrimination, and how is it identified? What are the societal risks associated with biased AI? How have recent cases highlighted the issue of AI fairness? What technical challenges exist in developing unbiased algorithms? How do data collection and processing contribute to bias? What role do legal and ethical frameworks play in perpetuating or preventing bias? What methodologies are being developed to detect and mitigate bias? How can diversity in AI development teams reduce the risk of discrimination? What are the predictions for the future of unbiased AI and its impact on society?

By addressing these questions, this insight contributes to the ongoing discussion on unbiased AI and provides information that can guide policymakers, developers, and users on how to create AI systems that are fair, transparent, and ethical.

The Emergence of Algorithmic Discrimination in AI

What is algorithmic discrimination and how is it identified?

Algorithmic discrimination emerges as a critical issue when automated systems produce unjust results, often as a result of biases inherent in their input data. The identification of such discrimination requires a multifaceted approach, beginning with a rigorous examination of the data sets used to train these algorithms.[1] This scrutiny is essential because biased, inaccurate, or unrepresentative training data is a known predecessor to algorithmic bias, which can directly lead to discriminatory outcomes.[2]  Moreover, the discrimination may not always be obvious; it can be subtle and embedded within the algorithm’s decision-making process, such as when a seemingly neutral feature like postal code indirectly correlates with a protected attribute, leading to unintended consequences in sectors like loan and insurance premium calculations.[3] The challenge is compounded by the fact that algorithms, particularly in machine learning (ML) and artificial intelligence (AI) operate on a statistical basis, meaning that discrimination can be statistically identified by analyzing outcome patterns related to protected attributes such as race, gender, or age.[4] Therefore, a clear understanding and identification of algorithmic discrimination require not only a thorough analysis of the data but also a study of the algorithm’s outcomes in relation to these protected attributes to uncover both direct and indirect discriminatory effects.[5]

What are the societal risks associated with biased AI?

Amidst the ongoing discourse on algorithmic fairness, the societal repercussions of biased AI systems are vast and multifaceted, with harmful impacts that extend to individual rights and social cohesion. Biased AI, resulting from flawed algorithm design or data collection processes, can lead to discriminatory decision-making where individuals are adversely affected based on their gender, race, sexual orientation, and other attributes.[6] This is not merely a theoretical concern but manifests in various sectors such as employment, finance, and law enforcement, where biased algorithms can perpetuate existing stereotypes, inadvertently reinforce social biases, and result in unfair decision-making.[7] Therefore, it is crucial to implement rigorous data quality control processes, including the renewal of data sets to ensure the accuracy of statistical patterns and relationships, and to conduct regular verification and audits of the AI process, which can serve as safeguards against the perpetuation of societal biases.[8]

How have recent cases highlighted the issue of AI fairness?

The pervasive influence of AI in decision-making processes has recently come under scrutiny in the light of several high-profile cases that have brought the issue of algorithmic fairness into sharp relief. These incidents illustrate the potential for AI applications to perpetuate or even exacerbate discrimination against legally protected groups, a reality supported by empirical evidence.[9] For instance, in the European context, these cases have raised complex legal challenges as to whether the current frameworks of anti-discrimination laws are apt to tackle the nuances of algorithmic decision-making.[10] Such challenges are compounded by the difficulty victims face in proving discrimination without access to the proprietary algorithms and datasets used to make these decisions.[11] To address these issues, experts are calling for a synergistic approach that combines the principles of anti-discrimination and data protection laws. This proposed integrated strategy underscores the need for mechanisms such as algorithmic audits and data protection impact assessments, which can shed light on the often opaque decision-making processes of AI systems.[12] By doing so, this approach advocates for a proactive stance where transparency serves as the bedrock for ensuring fairness and equity in the digital age.[13]

Current Challenges in Achieving Unbiased AI

What technical challenges exist in developing unbiased algorithms?

One of the fundamental challenges in the quest for unbiased algorithms lies in the mathematical formulation of fairness and causality. The philosophical concept of causality, pivotal for fair decision-making, must be translated into precise mathematical language that information systems (IS) can utilize, which is a substantial hurdle for both researchers and practitioners.[14] Furthermore, despite regulatory efforts like the General Data Protection Regulation (GDPR) mandating transparent algorithms, the reconciliation of such transparency with the implementation of fair AI remains an intricate research area.[15] Regulatory initiatives advocate for transparency but do not necessarily dictate the methodology for achieving it, thus underscoring the need for a proactive approach to developing fair algorithms.[16] This task is compounded by the necessity to measure fairness through a meticulous analysis of prediction performance, including a critical examination of error rates among different subgroups, which can reveal hidden biases within the system.[17] Consequently, the technical work of developing unbiased algorithms is not merely about crafting code but involves a nuanced understanding of the underlying data, the societal context, and the legal framework that governs the deployment of such algorithms.

How do data collection and processing contribute to bias?

The insidious nature of bias in data collection and processing is multifaceted and can manifest in various forms that ultimately contribute to the unfair outcomes observed in algorithmic discrimination. For instance, selection bias is a significant concern; it occurs when the data are not fully representative of the population they are meant to model, leading to systematic deviations in the estimated parameters from their true values.[18] It is akin to the biases encountered in behavioral experiments, where the selection of participants can skew results.[19] This type of bias is often perpetuated when data is annotated in a way that reinforces the annotators’ preconceived notions, thereby recycling and amplifying existing prejudices.[20] Furthermore, Mehrabi’s survey categorizes these as historical and representational biases, the former pertaining to historical data that may be outdated or tainted with past discriminatory practices, and the latter concerning how data represents different groups within the sample.[21] These biases in the initial stages of data handling are not self-contained; they are replicated and magnified in subsequent analyses and model training, which can lead to discriminatory decisions by algorithmic systems.[22] Thus, acknowledging the pivotal role of data collection and processing in introducing biases into AI systems is crucial for developing more equitable technologies.[23]

What role do legal and ethical frameworks play in perpetuating or preventing bias?

The pivotal role that legal and ethical frameworks play in curbing the perpetuation of bias in AI systems is multi-faceted and critical to ensuring equitable decision-making processes. In light of the societal risks associated with biased AI, such frameworks emerge as necessary tools to both recognize and address the ethical and social dimensions of bias, discrimination, and fairness.[24] The necessity of these frameworks is underscored by the argument posed by Fjeld et al., who contend that engineering methods alone are insufficient to safeguard fundamental rights against the unintended consequences of AI technologies.[25] This assertion is further supported by the obligation of fairness outlined by the Access Now Organization and the Public Voice Coalition, which provide benchmarks for defining bias in AI, indicating the need for a structured approach to evaluating these systems.[26] Furthermore, the Principled AI International Framework exemplifies a global initiative to establish comprehensive policies and guidelines, informed by ethical principles, to steer the design of AI towards the avoidance of biased decisions.[27] These efforts highlight the indispensable nature of legal and ethical frameworks in not only preventing bias but also in fostering trustworthiness and accountability within AI systems, thereby mitigating the societal risks that biased decision-making algorithms could precipitate.

Solutions and Future Directions for Equitable AI Systems

What methodologies are being developed to detect and mitigate bias?

As the literature on algorithmic bias grows, distinct methodologies for both detecting and mitigating such biases are being developed to address the multifaceted nature of unfairness in machine learning systems. One of the emerging approaches involves the integration of constraint-based methods in the design of recommender systems. These methods are predicated on the inclusion of explicit constraints derived from multi-objective optimization techniques that strive to balance various aspects of fairness.[28] For instance, fairness in algorithmic systems is increasingly being treated as an optimization constraint, a mathematical expression of the desired fairness, ensuring it is given due consideration during the model’s learning process.[29] This approach is particularly evident in the creation of recommender systems, where researchers are actively incorporating fairness characteristics into the algorithmic design. By doing so, they aim to prevent the reinforcement of existing biases or skewed distributions in the underlying data that could perpetuate unfairness.[30] To achieve this, constraint-based optimization problems are formulated to ensure equitable representation across various sub-groups, thereby allowing for a more balanced sampling that does not adversely affect gradient-based learning algorithms.[31] Such innovative methods represent a departure from traditional associative or causal inference techniques, offering a new paradigm that emphasizes fairness-of-exposure across the board.[32] These developments highlight a growing recognition of the importance of technical interventions in ensuring algorithmic fairness and represent a significant step forward in the ongoing effort to create more equitable AI systems.

How can diversity in AI development teams reduce the risk of discrimination?

The imperative to mitigate algorithmic discrimination leads directly to the composition of AI development teams. A homogenous group of developers may inadvertently encode their own biases into the AI systems they create, as they may overlook or fail to recognize discriminatory patterns that could emerge from the technology.[33] For instance, an AI system developed predominantly by one demographic could be less effective or fair when deployed in urban or public services, as it may not accurately reflect the needs and circumstances of a diverse population.[34] By prioritizing diversity within these teams, organizations can bring a wide range of perspectives to the table, which is crucial in identifying and correcting biases that might otherwise go unnoticed. This is not merely a theoretical benefit; practical implementations have shown that diverse teams are more adept at anticipating the varied ways in which different demographic groups interact with technology, thereby reducing the risk of a biased AI system failing to serve a significant portion of the community effectively.[35] Diversity within AI development teams, therefore, acts as a safeguard against the perpetuation of existing inequalities through new technological mediums, ensuring that AI systems serve the public with greater fairness and competence.

What are the predictions for the future of unbiased AI and its impact on society?

Given the societal risks that biased AI can impose, as evidenced by its potential to lead to discriminatory decision-making, there is a clear imperative to address fairness within AI systems. The burgeoning field of fairness in AI has seen a significant rise in scholarly attention, with a pronounced uptick in research on fairness in recommender systems since the mid-2010s.[36] This surge in research underscores a growing consensus that fairness is not merely an optional feature but a core tenet of responsible AI development.[37] However, it is becoming increasingly evident that market forces alone are insufficient to guarantee the development of AI systems that fairly serve the interests of the entire population.[38] As the impact of AI continues to expand, there is a pressing need for legal frameworks that can ensure AI systems are designed and used in a manner that benefits society at large. This is where legal regulation steps in as an indispensable element. Predictions for the future of AI suggest that, much like the regulatory paths taken with sharing economy platforms such as Uber and Airbnb, the field of AI will also undergo a similar trajectory of government oversight.[39] The aim of such regulation would be to mitigate the potential harms that have been increasingly spotlighted in recent discussions, ensuring that AI-powered systems are harnessed for the greater good of society.[40]

Conclusion

The emergence of algorithmic discrimination as a critical issue in the use of AI systems has raised concerns about the potential for biased decision-making and unfair outcomes. This insight explored the multifaceted approach that is required to tackle algorithmic discrimination, beginning with a rigorous examination of the data sets used to train these algorithms. Biased, inaccurate, or unrepresentative training data is a known precursor to algorithmic bias, which can directly lead to discriminatory outcomes.

The challenge is compounded by the fact that algorithms operate on a statistical basis, meaning that discrimination can be statistically identified by analyzing outcome patterns related to protected attributes such as race, gender, or age. Moreover, the discrimination may not always be overt; it can be subtle and embedded within the algorithm’s decision-making process. Therefore, it is crucial to implement rigorous data quality control processes, including the renewal of data sets to ensure the accuracy of statistical patterns and relationships, and to conduct regular verification and audits of the AI process, which can serve as safeguards against the perpetuation of societal biases.

One of the fundamental challenges in the quest for unbiased algorithms lies in the mathematical formulation of fairness and causality, which requires a synergistic approach that combines the principles of anti-discrimination and data protection laws. The pivotal role that legal and ethical frameworks play in curbing the perpetuation of bias in AI systems is multifaceted and critical to ensuring equitable decision-making processes. The literature on algorithmic bias has spurred distinct methodologies for both detecting and mitigating such biases, including the integration of constraint-based methods in the design of recommender systems. As the impact of AI continues to expand, there is a pressing need for legal frameworks that can ensure AI systems are designed and used in a manner that benefits society at large. Overall, the research underscores a growing consensus that fairness is not merely an optional feature but a core tenet of responsible AI development and that diversity within AI development teams acts as a safeguard against the perpetuation of existing inequalities through new technological mediums, ensuring that AI systems serve the public with greater fairness and competence.


[1] Köchling A., Wehner M. C. Discriminated by an algorithm: a systematic review of discrimination and fairness by algorithmic decision-making in the context of HR recruitment and HR development // Business Research. 2020. Vol. 13. URL: link.springer.com/article/10.1007/s40685-020-00134-w.

[2] Ibid.

[3] Varona D., Suárez J. L. Discrimination, Bias, Fairness, and Trustworthy AI // Applied Sciences. 2022. URL: www.mdpi.com/2076-3417/12/12/5826.

[4] Ibid.

[5] Ibid.

[6] Ibid.

[7] Ibid.

[8] Köchling A., Wehner M. C., op. cit.

[9] Hacker P. Teaching fairness to artificial intelligence: Existing and novel strategies against algorithmic discrimination under EU law // Common Market Law Review. 2018. Vol. 55, no. 4. URL: https://kluwerlawonline.com/journalarticle/Common+Market+Law+Review/55.4/COLA2018095.

[10] Ibid.

[11] Ibid.

[12] Ibid.

[13] Ibid.

[14] Feuerriegel S., Dolata M., Schwabe G. Fair AI // Business & Information System Engineering. 2020. Vol. 62. URL: link.springer.com/article/10.1007/s12599-020-00650-3.

[15] Ibid.

[16] Varona D., Suárez J. L., op. cit.

[17] Feuerriegel, S., Dolata, M. & Schwabe, G., op. cit.

[18] Ibid.

[19] Ibid.

[20] Ibid.

[21] Varona D., Suárez J. L., op. cit.

[22] Feuerriegel S., Dolata M., Schwabe G., op. cit.

[23] Varona D., Suárez J. L., op. cit.  

[24] Ibid.

[25] Ibid.

[26] Ibid.

[27] Ibid.

[28] Deldjoo Y., Jannach D., Bellogin A., Difonzo A., Zanzonelli, D. Fairness in recommender systems: research landscape and future directions // User Modeling and User-Adapted Interaction. 2023. URL: link.springer.com/article/10.1007/s11257-023-09364-z.  

[29] Ibid.

[30] Ibid.

[31] Ibid.

[32] Ibid.

[33] Yigitcanlar T., Mehmood R., Corchado J. M. Green Artificial Intelligence: Towards an Efficient, Sustainable and Equitable Technology for Smart Cities and Futures // Sustainability. 2021. URL: www.mdpi.com/2071-1050/13/16/8952.

[34] Ibid.

[35] Ibid.

[36] Deldjoo Y., Jannach D., Bellogin A., Difonzo A., Zanzonelli D., op. cit.

[37] Ibid.

[38] Yigitcanlar T., Mehmood R., Corchado J. M., op. cit.

[39] Ibid.

[40] Deldjoo Y., Jannach D., Bellogin A., Difonzo A., Zanzonelli D., op. cit.

Key words: AI; Global Security

ET

E16/MIN – 24/01/29