Secondary Bounded Rationality: A Theory of How Algorithms Reproduce Structural Inequality in AI Hiring

Based on the provided abstract, here is a detailed technical summary of the paper.

Technical Summary: Secondary Bounded Rationality

1. The Problem: The Illusion of Algorithmic Objectivity

The paper addresses a critical failure in AI-driven recruitment: the gap between the promised "objectivity" of automated systems and their actual tendency to perpetuate systemic inequality. The core problem is that while AI is designed to optimize for efficiency, it does not operate in a vacuum. Instead, it inherits human cognitive biases and structural social disparities.

The authors identify a specific mechanism where AI systems transform social and cultural capital (e.g., prestigious degrees, specific linguistic styles, or professional networks) into mathematical proxies for competence. This creates a "black box" of inequality where discriminatory outcomes are rebranded as meritocratic, data-driven decisions.

2. Theoretical Framework & Methodology

The paper introduces a novel theoretical synthesis to explain this phenomenon, combining two distinct academic disciplines:

• Simon’s Bounded Rationality (Decision Science): The authors adapt Herbert Simon’s concept to the digital age. While traditional bounded rationality refers to human cognitive limits, "Secondary Bounded Rationality" refers to the constraints placed on AI. Even though AI has massive computational power, its "rationality" is bounded by the quality, scope, and inherent biases of the training data and the specific optimization functions defined by its developers.
• Bourdieusian Capital Theory (Sociology): The authors use Pierre Bourdieu’s framework to categorize the types of data AI processes. They argue that algorithms do not just look at "skills"; they look for Cultural Capital (credentials, mannerisms) and Social Capital (network connections/homophily).
• Multimodal Analysis: The methodology involves analyzing "multimodal recruitment frameworks"—systems that process diverse data types, such as text (resumes), audio/visual (video interviews), and relational data (network graphs).

3. Key Contributions

• The Theory of Secondary Bounded Rationality: This is the paper's primary conceptual contribution. It posits that AI's decision-making is limited by a recursive loop: the AI optimizes for "legible" data (data that is easy for a machine to quantify), but because social inequality makes certain groups more "legible" (e.g., those with elite credentials), the AI inadvertently optimizes for privilege.
• Identification of the Recursive Inequality Cycle: The paper maps how historical inequalities are converted into technical features, which are then used to train future models, creating a self-reinforcing loop of exclusion.
• Taxonomy of Algorithmic Bias in Hiring: It distinguishes between different modes of bias, such as credential privileging (over-weighting elite institutions) and network homophily (favoring candidates who "look" or "act" like the existing successful workforce).

4. Results and Findings

The analysis demonstrates that:

• Proxy Transformation: AI systems successfully transform non-meritocratic social advantages into "ostensibly meritocratic" outcomes. For example, a specific dialect or a gap in employment (often tied to socioeconomic status) is mathematically interpreted as a lack of "competence" or "cultural fit."
• Optimization Paradox: The more an algorithm optimizes for "efficiency" and "predictive accuracy" based on historical data, the more it entrenches existing social hierarchies. The "accuracy" the AI achieves is actually just the accurate replication of historical human bias.

5. Significance and Proposed Mitigations

The paper is significant because it moves the conversation from "fixing data" to "rethinking the logic of optimization." The authors argue that simply adding more diverse data is insufficient if the underlying logic remains bounded by traditional definitions of merit.

Proposed Mitigation Strategies:

• Counterfactual Fairness Testing: Testing how an algorithm would decide if a candidate's protected attributes (race, gender, class) were changed, ensuring the decision remains stable.
• Capital-Aware Auditing: Developing auditing tools specifically designed to detect when an algorithm is using "cultural capital" as a proxy for "job performance."
• Regulatory Interventions: Moving beyond voluntary ethical guidelines toward formal legal requirements for algorithmic transparency and social impact assessments in the recruitment sector.