Introduction
In high-stakes domains like child welfare, the development of "Ethical AI" often relies on abstract concepts of justice—such as fairness or equity—that are disconnected from the operational realities of the systems they serve. This paper argues that justice in AI should not be defined in a vacuum. Instead, it must be grounded in the "fine print" of administrative policy: the dense thicket of rules and regulations that dictate how decisions are made and how rights are protected in practice.
By analyzing over 23,000 child welfare policy documents, the researchers identify 13 distinct "justice principles" that are already codified in law. They propose that these existing values should serve as the primary source for Value Sensitive Design (VSD), ensuring that AI tools align with the institutionalized values of the human services landscape.
Methodology: A Computationally Book-Ended Approach
The researchers employed a mixed-methods case study to bridge the gap between high-level policy and actionable AI design. The process involved three key stages:
- Computational Grounded Theory: A Structural Topic Model (STM) was applied to a massive corpus of 23,298 policy documents from all 50 U.S. states, Washington D.C., and Puerto Rico. This identified broad thematic trends across 13 million words of legal text.
- Qualitative Coding: Human experts performed iterative coding on 379 policies from New York State to refine these themes into specific justice principles.
- Sentence Embeddings (SimCSE): To validate the findings, the team used the SimCSE model to perform similarity searches, identifying policy excerpts that matched the principles even when keywords differed.
Defining Justice Through Policy
The study surfaced 13 Justice Principles, ranging from well-known concepts like transparency to nuanced rights like the "Liberty of Faith" and "Proprietary Rights of Familial Relations." These principles act as institutionalized guardrails, defining how values are operationalized before a technical tool is even conceptualized.
These principles are categorized into four dimensions: defining justice (e.g., fairness), communicating justly (transparency), intangible justice (e.g., well-being, autonomy), and acting justly (e.g., protecting from harm).
From Policy to Practice: Operationalizing Values
To make these principles actionable for AI developers, the researchers mapped them to specific policy excerpts. This mapping demonstrates how abstract values like "Well-being" or "Relationships of Care" are translated into concrete requirements, such as providing a "family atmosphere of acceptance" or recognizing the "emotional relationship of siblings."
One standout insight is the principle of Protecting from Exposure to Harm. In child welfare, system interaction is often inherently traumatic. Consequently, AI tools should be designed to minimize these interactions, effectively placing a "burden of proof" on developers to demonstrate that their tool represents the least harmful course of action.
Design Guidelines for Human Services AI
Based on these findings, the authors derived eight design guidelines. A core tenet is that AI must treat data points not as static rows in a database, but as "real independent actors whose future is not dictated by their past." This shift requires moving away from purely predictive risk modeling toward tools that nurture and strengthen a child’s primary, safe relationships.
Limitations and Future Directions
While the national corpus provided a broad view, the deep qualitative analysis focused on New York State. Policy varies significantly by jurisdiction; for instance, some states prioritize outdoor youth programs while others focus on facility standards. Future work will expand this value source analysis to other states and incorporate the lived experiences of system-impacted individuals to identify potential conflicts between administrative policy and community needs.
Summary updated