First Published on 5 July 2023.

Introduction:

Welcome back! Data science has revolutionized numerous industries, providing valuable insights and enabling data-driven decision-making. However, as algorithms become increasingly pervasive, it is crucial to examine and address the issue of bias that can be embedded within them. In this blog post, we will explore a specific lens of bias in data science, focusing on the intersection of gender and facial recognition algorithms.

Part 1: Examples of Algorithmic Biases

To understand the impact of bias in facial recognition algorithms, let’s examine some real-world examples:

1. Bias in Online Recruitment Tools: Amazon, a renowned online retailer, faced gender bias issues in its recruiting algorithm. The algorithm, trained on historical data predominantly from male applicants, downgraded resumes that included the word “women’s” and penalized candidates who attended women’s colleges. This resulted in gender bias and hindered the advancement of female applicants.
2. Bias in Word Associations: Princeton University researchers found that machine learning algorithms, when analyzing word associations, displayed biases such as perceiving European names as more pleasant than African-American names. Additionally, the association of “woman” and “girl” with arts rather than science and math perpetuated gender stereotypes.
3. Bias in Online Ads: Research conducted by Harvard’s Latanya Sweeney revealed that online search queries for African-American names were more likely to result in ads related to arrest records and high-interest credit cards. This biased targeting reinforces societal stereotypes and perpetuates discriminatory practices.
4. Bias in Facial Recognition Technology: MIT researcher Joy Buolamwini discovered that commercial facial recognition software exhibited lower accuracy rates in recognizing darker-skinned faces, particularly darker-skinned women. This bias stemmed from the underrepresentation of diverse faces in the training data.
5. Bias in Criminal Justice Algorithms: The COMPAS algorithm, used by judges to predict defendants’ risk of reoffending, was found to be biased against African-Americans, leading to longer periods of detention for individuals from this group. This bias highlights the potential for algorithms to perpetuate discriminatory practices within the criminal justice system.

Part 2: Causes of Bias

Understanding the causes of bias in facial recognition algorithms is crucial for developing effective mitigation strategies. Two primary causes of bias are:

1. Historical Human Biases: Biases entrenched in historical human practices, such as systemic racism or gender discrimination, can inadvertently find their way into algorithms. If training data reflect these biases, the algorithm may replicate and perpetuate discriminatory outcomes.
2. Incomplete or Unrepresentative Training Data: Insufficient training data that lacks diversity can contribute to biased algorithmic outcomes. For example, the underrepresentation of diverse faces in facial recognition training data leads to lower accuracy rates for certain groups.

Part 3: Mitigation Proposals

To address gender bias in facial recognition algorithms, we propose the following mitigation strategies:

1. Updating Nondiscrimination and Civil Rights Laws: Existing laws should be updated to encompass online practices and ensure that algorithms do not contribute to discriminatory outcomes. This would provide legal frameworks to hold algorithm operators accountable for biases.
2. Implementing Bias Impact Statements: Operators of algorithms should develop bias impact statements that assess potential biases throughout the algorithm’s design, implementation, and monitoring phases. These statements would help identify and mitigate biases proactively.
3. Diversity-in-Design: To combat bias, diversity should be integrated into the algorithm design process. Employing a diverse workforce, inclusive spaces, and cultural sensitivity in decision-making processes can help address biases upfront.