What is the unintended source of bias in artificial intelligence (Al) algorithms?

Human developers are a primary, though often unintended, source of bias in AI algorithms. Their personal, cognitive, and societal biases can be unintentionally embedded at every stage of the AI development lifecycle. This includes the initial problem formulation, the selection and labeling of training data, the choice of features and algorithms, and the interpretation of results. Because developers' worldviews and assumptions shape the system's design and the data it learns from, their inherent biases become encoded in the AI's logic, leading to skewed or unfair outcomes.

A. Tainted examples: This is a direct cause of bias, but it is often a symptom or result of human bias in the data collection and selection process, rather than the ultimate source.

B. Excessive validation sample sizes: A large validation set is generally beneficial for assessing a model's generalization capabilities and does not inherently introduce bias. Small or unrepresentative samples are the problem.

D. Excessive training sample sizes: A large quantity of training data is typically advantageous. The source of bias is the quality and representativeness of the data, not its sheer volume.

1. Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., & Galstyan, A. (2021). A Survey on Bias and Fairness in Machine Learning. ACM Computing Surveys, 54(6), 1–35. In Section 2.2, "Human Bias," the authors state, "Bias can arise from a human’s cognitive, personal, and institutional biases... This can influence the data generation process, the algorithm design, and the evaluation process." This identifies the human element as a core source of bias that permeates the entire system. (DOI: https://doi.org/10.1145/3457607)

2. Suresh, H., & Guttag, J. V. (2021). A Framework for Understanding Sources of Harm throughout the Machine Learning Life Cycle. In Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO '21). This framework details how human choices and biases introduce harm at every stage, including data collection (leading to "tainted examples") and model development. It positions human-driven processes as the origin of these issues. (DOI: https://doi.org/10.1145/3465416.3483305)

3. Friedman, B., & Nissenbaum, H. (1996). Bias in Computer Systems. ACM Transactions on Information Systems, 14(3), 330–347. This foundational paper identifies three categories of bias, including "preexisting bias," which is rooted in social institutions, practices, and attitudes. Human developers, as members of society, carry and can embed these biases into systems. (DOI: https://doi.org/10.1145/230538.230561)