The Myth of the Objective Algorithm: Datafication, Hidden Classifications, and Digital Discrimination

In this deep-dive episode, the hosts explore the inner mechanics of artificial intelligence and data-driven decision-making systems. The episode deconstructs the widespread assumption that algorithms serve as neutral, objective mirrors of reality, revealing how AI functions in practice as human-constructed classification systems that embed cultural assumptions, political priorities, and historical biases directly into code.

The hosts examine foundational concepts by distinguishing between strong and weak AI and highlighting the hidden consequences of classification theory, including how reducing social complexity creates invisible residual categories for groups excluded from digital data collection. The episode dismantles the myth of Ground Truth, explaining why historical data cannot serve as an objective benchmark when it lacks tacit human knowledge and inevitably accumulates errors over time.

Through a systematic walkthrough of the five lifecycle stages of an AI model—task definition, data collection, data cleaning, model selection, and training and deployment—the hosts demonstrate how human choices and selective exclusions influence model outcomes at every step. Systemic human biases manifest in distinct forms, including modeling bias, training bias, and application bias. Real-world examples illustrate these dynamics, such as Amazon's automated hiring tool that filtered out female applicants, Microsoft's chatbot Tay, Facebook's image recognition, and Cambridge Analytica's behavioral profiling. A historical line is also drawn from nineteenth-century craniometry to modern computer vision datasets like ImageNet and UTKFace to show how technical taxonomies have long reproduced societal inequalities.

Furthermore, the episode explores technical and methodological frameworks for auditing algorithmic bias. The hosts analyze procedural and relational approaches to explainable AI, examining tools like LIME, counterfactual analyses, group fairness, and predictive parity. They explain why fairness through blindness fails when confronted with proxy variables like postal codes, and why effective governance demands a commitment to zero-sense discrimination, intersectionality, discrimination-aware AI, and value-accountability-by-design. The episode concludes with a perspective on the 10/100 principle, illustrating how combining core domain expertise with advanced AI capabilities drives responsible and high-impact innovation.