Ethics of AI: A Systematic Literature Review of Principles and Challenges

Arif Ali Khan, Sher Badshah, Peng Liang, Muhammad Waseem, Bilal Khan, Aakash Ahmad, Mahdi Fahmideh, Mahmood Niazi, Muhammad Azeem AkbarEASE 2021 · DOI

Introduction

As Artificial Intelligence (AI) transitions from theoretical research to pervasive integration in sectors like healthcare, finance, and critical infrastructure, the urgency for ethical oversight has intensified. While international bodies have established numerous ethical guidelines, a significant disconnect remains between these high-level ideals and their practical application in software engineering. This paper addresses this "implementation gap" by identifying core ethical principles and the specific barriers preventing their adoption in the development lifecycle.

Methodology

The researchers conducted a Systematic Literature Review (SLR) following the Kitchenham and Charters guidelines to ensure a rigorous, unbiased meta-analysis. The research was structured into three distinct phases: planning, conducting, and reporting.

The three-phase SLR research process: Planning, Conducting, and Reporting.

The study selection process was exhaustive. Initially, 811 studies were retrieved from five major digital repositories, including ACM Digital Library and IEEE Xplore. Through a multi-stage filtering process—including title/abstract screening, full-text review, and backward snowballing—the pool was narrowed to 27 high-quality primary studies.

The study selection flow, illustrating the reduction from 811 initial results to 27 final primary studies.

To ensure robustness, each selected study underwent a quality assessment based on six criteria (QA1-QA6), evaluating research objectives, methodology, and the systematic discussion of results.

Core Ethical Principles

The review identified 21 distinct ethical principles. Analysis revealed a global consensus on a "Big Four" set of values:

  • Transparency: Cited in 17 studies.
  • Privacy: Cited in 16 studies.
  • Accountability: Cited in 15 studies.
  • Fairness: Cited in 14 studies.

Other principles, such as sustainability and solidarity, were significantly less prevalent, suggesting that current ethical discourse is heavily focused on individual rights and system oversight.

Challenges to Ethical Adoption

The study identified 15 major challenges hindering the integration of ethics into AI engineering. These can be categorized into three primary areas:

  1. The Knowledge Gap: A lack of ethical awareness among both management and technical staff. Practitioners often lack the training to recognize or address moral complexities.
  2. Conceptual Vagueness: Guidelines are frequently too abstract. For instance, "fairness" is rarely defined in technical terms that a developer can translate into code or testing suites.
  3. Stakeholder Conflict: Different actors (e.g., developers vs. end-users) often have incompatible interpretations of ethical duties, leading to friction during implementation.

Conclusion and Future Directions

The rapid growth in AI ethics research—evidenced by the surge from 2 publications in 2018 to 19 in 2020—highlights the field's increasing importance. However, the current "infancy" of the field means that many findings are still being refined. The authors propose the development of a Capability Maturity Model for AI Ethics, similar to the CMMI used in traditional software engineering. This model would provide a structured roadmap for organizations to assess and improve their ethical practices, moving beyond abstract principles toward measurable, engineering-grade ethical standards.

Summary updated

BibTeX
@inproceedings{khan2021ethics,
  author = {Arif Ali Khan and Sher Badshah and Peng Liang and Muhammad Waseem and Bilal Khan and Aakash Ahmad and Mahdi Fahmideh and Mahmood Niazi and Muhammad Azeem Akbar},
  title = {{Ethics of AI: A Systematic Literature Review of Principles and Challenges}},
  booktitle = {EASE},
  year = {2021},
  doi = {10.1145/3530019.3531329},
}
Figure