← Saleh Afroogh · Featured Projects
AI and Ethics (Springer) · 2025

Navigating LLM Ethics

Advancements, challenges, and future directions

Junfeng Jiao, Saleh Afroogh, Yiming Xu, Connor Phillips
PaperarXiv
192papers selected for review
1,147search results screened
3challenges unique to LLMs
200+citations on Google Scholar

Overview

LLMs inherit the ethical problems of AI in general, such as privacy and fairness, and add new ones of their own. This PRISMA-based systematic review separates the two. It examines challenges common to LLMs and other AI systems, and challenges that arise uniquely from LLMs: hallucination, verifiable accountability, intellectual property and the complexity of decoding censorship. For each, it reviews mitigation strategies and argues for domain-specific ethical frameworks, dynamic auditing and interdisciplinary collaboration.

Highlights

Framework

Fig. 2: how many reviewed studies address each ethical topic. Mitigation strategies, case studies, and bias and fairness dominate; censorship is barely studied.
Fig. 2: how many reviewed studies address each ethical topic. Mitigation strategies, case studies, and bias and fairness dominate; censorship is barely studied.
Table 6: ethical concerns common to LLMs and other AI, those exacerbated in LLMs, and those unique to LLMs.
Table 6: ethical concerns common to LLMs and other AI, those exacerbated in LLMs, and those unique to LLMs.
Table 3: ethical frameworks and platforms proposed for LLMs.
Table 3: ethical frameworks and platforms proposed for LLMs.
Table 4: major biases identified in LLMs: social, language and representation biases.
Table 4: major biases identified in LLMs: social, language and representation biases.
Fig. 1: PRISMA flow diagram for the review.
Fig. 1: PRISMA flow diagram for the review.
Common challenges

Shared with other AI

  • Privacy and data security
  • Fairness and bias
  • Breaking the black box
  • Ethical evaluation and dynamic audits
New challenges

Unique to LLMs

  • Hallucination, fabrication and misinformation
  • Intellectual property, copyright and plagiarism
  • Censorship complexity and dilemmas
Examples reviewed

Mitigations

  • Knowledge unlearning and differential privacy
  • Debiasing in data, loss functions and fine-tuning
  • Output verification against hallucination
  • Watermarking and citation mechanisms
Recommendations

Future directions

  • Domain-specific ethical frameworks
  • Dynamic, continuous audit systems
  • Interdisciplinary collaboration
  • Adaptive, transparent censorship policies
Summary of the review structure: challenges, mitigations and future directions.

How it works

Resources

Paper (AI and Ethics)Paper ↗Preprint (arXiv)arXiv ↗

Citation

Jiao, J., Afroogh, S., Xu, Y., & Phillips, C. (2025). Navigating LLM ethics: Advancements, challenges, and future directions. AI and Ethics, 1–25.