Making Power Explicable in AI: Analyzing, Understanding, and Redirecting Power to Operationalize Ethics in AI Technical Practice

TL;DR

This paper argues that operationalizing AI ethics in technical practice is impeded by dysfunctional power structures within the sociotechnical system of AI. Through four empirical XAI cases, it diagnoses power as the latent factor that distorts scope, design, evaluation, and objectives away from public-interest ethics. It then offers three interventional pillars—making power explicable, reframing AI/AI ethics toward justice, and encoding ethics as technical methods including critical limitation analyses—to realign AI practices with ethical aims. Grounded in STS and feminist epistemology, the work emphasizes adversarial, ongoing scrutiny of power, narratives, and values to enable ethical and responsible AI deployment with tangible societal benefits.

Abstract

The operationalization of ethics in the technical practices of artificial intelligence (AI) is facing significant challenges. To address the problem of ineffective implementation of AI ethics, we present our diagnosis, analysis, and interventional recommendations from a unique perspective of the real-world implementation of AI ethics through explainable AI (XAI) techniques. We first describe the phenomenon (i.e., the "symptoms") of ineffective implementation of AI ethics in explainable AI using four empirical cases. From the "symptoms", we diagnose the root cause (i.e., the "disease") being the dysfunction and imbalance of power structures in the sociotechnical system of AI. The power structures are dominated by unjust and unchecked power that does not represent the benefits and interests of the public and the most impacted communities, and cannot be countervailed by ethical power. Based on the understanding of power mechanisms, we propose three interventional recommendations to tackle the root cause, including: 1) Making power explicable and checked, 2) Reframing the narratives and assumptions of AI and AI ethics to check unjust power and reflect the values and benefits of the public, and 3) Uniting the efforts of ethical and scientific conduct of AI to encode ethical values as technical standards, norms, and methods, including conducting critical examinations and limitation analyses of AI technical practices. We hope that our diagnosis and interventional recommendations can be a useful input to the AI community and civil society's ongoing discussion and implementation of ethics in AI for ethical and responsible AI practice.

Figures (4)

• Figure 1: The scope of AI ethics in our paper. The problem of ethics implementation in AI technical practices focuses on II. practical AI ethics that implements I. AI ethics theories, principles, and values into the AI technical professionals’ day-to-day practices and activities in the whole pipeline of AI research, development, deployment and maintenance, etc.
• Figure 2: The feature attribution explanation on a natural image task of insect classification (panel A) and on a medical image tasks of brain tumor grading from MRI (panel B). The input image to an AI model and the AI prediction is indicated. The feature attribution explanation use a colored mask (in red) to highlight the important image pixels for the AI model’s prediction. The insect image is photographied by Jack Dykinga from USDA Agricultural Research Service (https://www.ars.usda.gov/oc/images/photos/may00/k5400-1/), and the brain MRI is from BraST 2020 dataset Menze2015 with data ID BraTS20_Training_221. Both images are publicly available. The medical image only shows the MRI T1CE modality for illustrative purpose and other MRI modalities are not shown.
• Figure 3: The neurosurgeon user study result in Case 4. It shows the 35 neurosurgeon participants’ medical image reading task accuracies in an experimental setting. The task was carried out in three conditions shown by the three box plots : when doctors performed the task alone (DR), with the prediction assistance of a black-box AI (DR+AI), and with the assistance of AI prediction and explanation (DR+XAI). Details about the study is in JIN2024102751. The result shows a repetitive pattern that can be seen in other similar AI-assisted studies, where with the help of a superior AI (the AI model’s accuracy 0.88 is visualized by the dotted line), human performance improved but not exceeding AI performance (*: statistically significant; ns: not significant). The same result may support distinct actions depending on different perspectives: it can either provide empirical evidence to support the replacement of humans with a black-box AI; or the empirical evidence can reveal flaws in the existing XAI approaches to improve both XAI and task performance.
• Figure 4: Illustration of the underlying mechanism on why it is difficult to implement ethics in AI in \ref{['sec:analyze']}. It shows that AI is embedded in the sociotechnical system (iceberg surrounded by sea and air). AI (iceberg) and the social system (sea) coproduce each other. The technical aspect of AI (tip of the iceberg) is built upon the social and cultural aspects of the AI community (the underwater part of the iceberg). Power can be understood as particles that are diffusive in the sociotechnical system. It can be polluted by unjust power, illustrated by the black color. The red circled numbers highlight key points to intervene in the roots, which is the dysfunction in power mechanisms, detailed in \ref{['sec:framework']}: 1 by making power visible (\ref{['sec:visible']}); 2 by reflecting and resisting the framing of narratives and assumptions from unjust power (\ref{['sec:frame']}); and 3 by embedding ethical power (white particles) into the social aspect of AI (the underwater part of the iceberg, \ref{['sec:method']}).