TL;DR
The "gap" between AI ethics principles and actual practice isn't just a failure of implementation or "ethics washing." By studying AI-enabled anti-trafficking systems, this paper reveals that ethical failures often stem from how different professional and institutional contexts shape what developers even perceive as a problem. To fix this, we need "epistemic infrastructures"—durable institutional arrangements that ensure the knowledge of affected communities actually reaches and influences the design process.
What Problem Does This Paper Address?
Responsible AI (RAI) scholarship has long noted a persistent "principle-practice gap": we have high-level values (like fairness and transparency), but they rarely manifest in finished systems. Usually, this is blamed on a lack of awareness, technical difficulty, or corporate "ethics washing."
The authors argue this explanation is too simple. They suggest that ethical principles aren't stable "inputs" that get lost during coding. Instead, ethical responsibility is a situated practice. In institutionally complex fields like anti-trafficking, a police officer, a human rights researcher, and a software engineer might all be "committed to ethics," but their professional "moral economies" lead them to define "harm" and "evidence" in fundamentally different, often conflicting ways. The paper seeks to understand how these institutional logics determine which ethical concerns get built into a system and which are "attrited" (lost) along the way.
Methods
This is a qualitative, interpretive study rather than a quantitative benchmark. The researchers conducted 14 semi-structured interviews (between Oct 2025 and Feb 2026) with a diverse range of actors in the anti-trafficking space, including:
- Technology developers and software engineers.
- Law enforcement and intelligence practitioners.
- NGO and civil society leaders.
- Academic researchers.
The authors used an abductive analysis approach, tracing how participants justified design choices. They developed a framework of four recursive moments to track how an ethical concern moves from a thought to a feature:
- Perception: What is recognized as an issue?
- Judgement: How is that issue prioritized against competing goals (e.g., privacy vs. detection)?
- Situated Action: Does the actor have the mandate or resources to do something about it?
- Socio-technical Materialisation: Does the decision actually survive in the code, the UI, or the organizational workflow?
Datasets
As a socio-technical study, the paper does not introduce or use a training/evaluation dataset in the machine learning sense. Instead, it analyzes the types of data used in the systems discussed by participants, such as:
- Administrative and financial records for investigation.
- Dark-web intelligence and open-source (OSINT) data.
- Worker-generated reporting from supply chains.
- Victim-assistance data from NGOs.
Key Insights and Ideas
- Epistemic Choices are Ethical Choices: Decisions that seem purely technical—like how to label a "vulnerability" or which data proxies to use for "risk"—are the primary sites where ethics happen. If a system for police defines "victim" through a narrow legal lens, it may "erase" male victims or migrant workers at the data level before an ethical review even begins.
- The Four Moral Economies: The authors identified four distinct "registers" that dictate what actors value:
- Exposure & Safety: Focus on protecting the individual from the system (common in NGOs).
- Evidence & Actionability: Focus on linking data to catch traffickers (common in law enforcement).
- Risk & Compliance: Focus on corporate liability and ESG metrics (common in the private sector).
- Epistemic Authority: Focus on the validity and representation of the knowledge itself.
- Points of Attrition: Ethical concerns are often lost not because of malice, but because of "constrained action." For example, a researcher might want to include survivor input but lacks the "discretionary time" or the organizational mandate to do so.
- Displaced Responsibility: Developers often "route" ethical concerns away, arguing that ethics belongs to the "applicator" (the user) rather than the "builder," ignoring how their upstream technical choices (like data linkage) constrain what the user can eventually do.
Key Results
The paper provides a qualitative map (Table 1) of how ethical concerns are lost at different stages:
- Perception stage: Harms like "racialized profiling" often go unperceived by those in investigative roles who view biased data simply as "noise."
- Judgement stage: Privacy is often explicitly traded off for "detection" without a clear framework for the sacrifice.
- Materialisation stage: Even when ethical features are designed (like a "quick-exit" button for safety), they often fail to reach deployment because the project loses funding or the "institutional slack" disappears.
The core result is the argument for Epistemic Infrastructure. The authors conclude that RAI cannot rely on individual "virtue" or checklists. It requires durable institutional paths—like dedicated resources for survivor engagement and joint processes for examining problem definitions—to make marginalized knowledge "actionable" in design.
Limitations and Future Work
- Stakeholder Accounts vs. Observation: The study relies on what practitioners say they do, rather than direct observation of code-writing or system deployment.
- Lack of Survivor Voice: The study focuses on the "institutional actors" (developers, police, NGOs). The authors explicitly note that the lived experiences of survivors and affected communities are not captured here.
- Future Directions: The authors call for ethnographic and longitudinal studies that follow an AI tool from its first "epistemic choice" to its long-term impact on the communities it intended to help.
Summary updated