Navigating LLM Ethics: Advancements, Challenges, and Future Directions

I. Introduction

• LLMs’ growing influence on communication and information dissemination heightens the need for accountability and ethical consideration, especially given their propensity for inaccurate content.Source: §I · p. 2
• The paper examines LLM ethics, including bias, societal impacts, model opacity, and prospective strategies such as dynamic audits, continuous monitoring, and explainability.Source: §I · p. 2
• The study aims to identify pressing ethical challenges and mitigation strategies, then discusses ethical principles and considerations before concluding with future prospects.Source: §I · pp. 2–3

2.1. Understanding Large Language Models

• Language models predict token probabilities autoregressively, using preceding tokens and the chain rule to estimate a sequence’s joint probability.The passage denotes tokens as u and describes conditional probabilities from the first token through token t.Source: §2.1 · p. 4
• LLMs scale up language models with many parameters and extensive, diverse training data, supporting coherent text generation and contextual language abilities.The paper describes their applications in areas including healthcare, education, and finance.Source: §2.1 · pp. 4–5

Review methods

• The review selected 192 papers from 456 relevant search results, applying journal-publication and LLM-ethics-topic criteria after removing duplicates.The authors report an English-language review using PRISMA, manual searching, and keyword-based Google Scholar searches.Source: §III · pp. 5–6

Ethical concerns

• The review’s coded concerns include bias and fairness, privacy, misinformation, accountability, intellectual property, inequality, public safety, and environmental impact.Four researchers conducted the qualitative analysis and developed eight major codes for categorizing the review results.Source: §IV · p. 7, Table 2 · p. 7

Ethical frameworks and platforms

• The framework themes span human-centered and user-centric approaches, trustworthiness, deontological ethics, autonomy, risk assessment, and Ethics by Design.The review also describes top-down frameworks grounded in moral theories and multidisciplinary risk research.Source: §IV · p. 7

Mitigation strategies

• Mitigation strategies address bias, privacy, and hallucination through approaches including dataset enhancement, adversarial learning, embedding purification, and output verification.Source: §IV · p. 8

Transparency in LLMs

• Transparency concerns include undocumented data and questionable legality, while the review identifies clarity, interpretability, open practices, and model reporting as related themes.Source: §IV · p. 8

Other ethical concerns

• The taxonomy also includes censorship concerns such as subjectivity and potential infringement of free speech, alongside intellectual-property issues such as copyright and ownership.Source: §IV · p. 8

• LLM-related ethical concerns include bias and fairness, privacy and data security, misinformation, accountability, intellectual property, and unequal access,, [14],,,,,,.Source: §4.1 · p. 8
• Training data can encode bias or sensitive information, while generated false content may manipulate public opinion and black-box systems can obscure responsibility, [14],,,,.The paper also links human-like text generation to plagiarism concerns and notes that unequal distribution of benefits may exacerbate inequality,,.Source: §4.1 · p. 8
• Prior work examines LLM ethics through surveys, experiments, reviews, and scenario-specific studies, including work on cheating advice and blending advertisements with search results,,,,.Source: §4.1 · p. 9
• Addressing these concerns involves technical strategies, privacy policy, ethical guidelines, and stakeholder engagement across industry, academia, and regulatory bodies,,,,,,,.Source: §4.1 · p. 9

• A proposed human-centered evaluation framework considers users’ mental models, use-case utility, and cognitive engagement in human–AI interaction.Source: §4.2 · p. 9
• Proposals include user-centric measures, trustworthiness categories, deontological ethics, and medical-ethics-informed monitoring and regulation for healthcare chatbots,,,.Source: §4.2 · p. 10
• Other approaches prioritize human autonomy, curated datasets, moral-theory-based top-down frameworks, and interdisciplinary risk research through RegNLP [31],,.Source: §4.2 · p. 10
• Ethics by Design approaches emphasize proactive planning, default ethical settings, embedded ethics, transparency, and respect for user values [34],,.Source: §4.2 · p. 10
• The discussion of GPT-3 and RoBERTa notes reliance on fine-tuning with specific data and the need for ethical-theory foundations.Source: §4.2 · p. 11

4.3.1. Different types of bias in LLM

• Social, language, and representation biases reflect disparities in social-group outputs, language performance, and whose perspectives are represented.The categories overlap, and bias can be multidimensional and content-specific rather than reducible to sentiment.Source: §4.3.1 · pp. 11–12
• Social biases are frequently flagged in public audits, language biases disadvantage low-resource languages, and representation biases are hardest to detect and fix.The passage also notes that mitigation can be difficult because of training-data scale, data scarcity, and costly, complex validation.Source: §4.3.1 · pp. 12–13

4.3.2. Fairness in LLM

• Fairness studies evaluate LLM disparities, including ChatGPT unfairness across sensitive attributes and group- and individual-fairness measures in high-stakes domains.Proposed approaches include demographic-aware prompt methods and reference-generation methods reported to improve fairness.Source: §4.3.2 · p. 13

4.3.3. Bias and Fairness Detection Methods

• Bias detection methods include template-based probes, likelihood and score metrics, embedding-space analysis, and behavioral audits.Examples include StereoSet and CrowS-Pairs for stereotypical completions, identity-group toxicity comparisons, embedding association tests, and audits of real-world interactions.Source: §4.3.3 · p. 14

• LLMs can memorize sensitive training data and reveal it through targeted extraction attacks, risking exposure of personal, financial, medical, and private information.The passage notes that such information may have been included in training datasets without consent.Source: §4.4 · p. 14
• Inference-time risks include sensitive user queries, revealing conversation histories, inadvertent disclosure about others, and exposure across third-party integrations.These risks complicate monitoring and control and intersect with GDPR and CCPA compliance, including deletion, minimization, and consent requirements.Source: §4.4 · p. 14
• Researchers are exploring privacy-preserving methods, including differential privacy, discriminator use, and matrix-vector functions, to protect data while maintaining utility and accuracy.Some research focuses on integrating such techniques into LLM architectures and training processes.Source: §4.4 · p. 15

• Unintentional misinformation can arise from contaminated training data, hallucinated facts, outdated information presented as current, or accurate information used inappropriately.The passage distinguishes misinformation as inadvertently false information from disinformation as deliberately false information.Source: §4.5 · p. 15
• LLMs can be deliberately misused to generate fake news, misleading social media content, conspiracy theories, propaganda, and targeted opinion manipulation.Proposed responses include fact-checking outputs against reliable sources, tracking AI-generated content, real-time monitoring, and training models to handle factual information more discerningly.Source: §4.5 · p. 15
• LLM capabilities advance faster than detection technologies, creating ongoing challenges for information integrity in digital environments.The passage also notes unresolved challenges in evaluating truthfulness and contextual accuracy.Source: §4.5 · pp. 15–16

4.6. Accountability and Governance in LLM

• Accountability supports transparent, auditable decision-making and is presented as important for public trust and regulatory expectations.The passage highlights concerns about biased or unreliable recommendations and added burdens in healthcare.Source: §4.6 · p. 16
• Proposed accountability measures include dataset documentation, citation mechanisms, data-accountability frameworks, and external scrutiny through red-teaming and auditing.The cited discussion also describes CANGARU Guidelines for ethical accountability, reproducibility, disclosure, and reporting.Source: §4.6 · pp. 16–17

4.7. Case Studies in LLM ethics

• Healthcare case studies emphasize confidentiality, consent, bias, and the potential harms of medical misinformation, with extensive human oversight in this high-stakes setting.The passage notes that liability for incorrect AI medical advice remains unclear.Source: §4.7 · p. 17
• Academic and educational cases include false references, plagiarism, disclosure standards, unreliable AI-assisted evaluations, and uneven access across institutions.The discussion describes institutional policy as reactive and teacher training as lagging behind student adoption.Source: §4.7 · pp. 17–18
• Business cases raise concerns about bias in hiring, promotion, and customer service, while specialized domains are said to require tailored ethical frameworks.The passage also notes reduced decision-making transparency as AI complexity increases.Source: §4.7 · p. 18

4.8. Mitigation Strategies in LLM ethics

• Mitigation studies address social bias, privacy protection, and hallucination prevention through technical, ethical, and procedural approaches.Bias interventions target datasets, training objectives and methods, fine-tuning, and model architecture.Source: §4.8 · p. 18
• Adversarial pretraining, semantic-similarity fine-tuning, and social-group-agnostic methods are among the reviewed bias interventions.The social-group-agnostic approach showed comparable performance to group-specific methods on standard bias benchmarks, with greater theoretical flexibility and practical scalability.Source: §4.8 · pp. 18–19
• Knowledge unlearning offered a more robust empirical privacy guarantee when data vulnerable to extraction attacks were known beforehand.A separate embedding-purification approach significantly outperformed baseline backdoor-mitigation methods across five sentence-classification tasks.Source: §4.8 · p. 19
• 47.5% to 14.5%: logit-output detection, validation, and mitigation reduced average hallucinations in GPT-3.5.The approach identifies potential hallucinations, validates candidates, and mitigates confirmed ones before generation proceeds.Source: §4.8 · p. 20
• Participatory design elicited user-desired interface features intended to reduce hallucinations’ impact on users.Users were involved in designing features aimed at reducing that impact.Source: §4.8 · pp. 20–21

4.9. Transparency in LLM

• A 20-participant study found that chaining LLM steps elevated task-result quality and increased transparency, controllability, and perceived collaboration.In chaining, one step’s output becomes the next step’s input.Source: §4.9 · p. 21
• Transparency approaches include model reporting, publishing evaluation results, providing explanations, communicating uncertainty, and tools that attribute output importance to input features,, [194], [195].Captum attributes importance scores to input features; LLM-TT analyzes Transformer attention heads and feed-forward neurons.Source: §4.9 · p. 21
• Many projects claiming to be open source use undocumented data of questionable legality, and few disclose instruction-tuning processes.The passage identifies human annotation labor as a reason instruction-tuning processes matter.Source: §4.9 · p. 21

4.10. Censorship in LLM

• Censorship decisions lack universally agreed standards, and cultural, political, and contextual differences can prompt bias concerns and suppress legitimate discussion.Limited transparency about blocked prompts, filters, and decision processes can also undermine user trust.Source: §4.10 · p. 22
• Strategic prompt engineering can circumvent restrictions, while RLHF may cause mode collapse and overly cautious responses that impair informativeness and user experience.The section also frames censorship governance as a contested balance among expression, safety, fairness, and inclusivity.Source: §4.10 · p. 22

4.11. Intellectual Property and Plagiarism in LLM

• LLM training datasets include copyrighted books, subscription-based papers, news, creative works, blogs, and social-media content, often without explicit authorization.The passage notes that such use also raises privacy concerns.Source: §4.11 · p. 23
• Current practice assumes fair use without clear legal precedent; opt-out systems are inconsistently respected, and creators receive no compensation for work contributing commercial value.The passage describes potential licensing losses for publishers and authors and competition from AI-generated articles for news organizations.Source: §4.11 · pp. 23–24

4.12. Abusive LLM, hate speech and cyber-bullying

• Studies used Llama 2 to identify online sexual predatory chats and abusive language, measured LGBTQ+ bias, and found ChatGPT-3 effective for cyberbullying detection.The Llama 2 approach showed high proficiency and consistency across three datasets and five experiments.Source: §4.12 · p. 24

4.13. Auditing LLM

• Reviewed LLM audit approaches include a ChatGPT-powered causal auditor, discrete optimization, iterative in-context learning, deep learning, and human-AI interaction principles.Audits examine decision processes, training-data biases, and outputs for accuracy, toxicity, and misinformation risks.Source: §4.13 · p. 24
• Auditing requires resources and technical expertise, while evolving models and diverse cultural contexts make static, one-size-fits-all evaluation inadequate.The section calls for dynamic, periodically recalibrated systems involving interdisciplinary teams and community stakeholders.Source: §4.13 · p. 25

5.1. The Advantages of Pre-Trained Models in Integrating Normative Ethics in LLM

• Pre-trained models may better incorporate ethical norms through diverse data and fine-tuning, while addressing LLM ethics also requires perspectives beyond engineering.The discussion calls for collaboration among ethicists, social scientists, psychologists, legal experts, and technical specialists.Source: §5.1 · p. 26, §5.2 · pp. 26–27

5.3. Protecting Privacy in Language Models amidst Growing Data Concerns

• Privacy risks include sensitive information in training data and personal data collected or retained in deployment, calling for technical, legal, and ethical responses.Retention and deletion have implications for privacy-law compliance, secure deployment, and user trust.Source: §5.3 · p. 27

5.4. Ethical Uniqueness: Specialized Considerations for LLMS

• LLM ethics includes concerns shared with other AI, concerns amplified by LLMs, and challenges described as unique, including hallucination, intellectual property, and censorship.Opacity complicates accountability; unequal access and abusive language are also amplified concerns.Source: §5.4 · pp. 27–28

5.5. Hallucination and Distorted Realities in LLMs

• Hallucinations range from false facts and fabricated sources to contextually inappropriate, misleading combinations, and outdated or temporally confused information.The discussion argues that prevention requires targeted approaches; unchecked hallucinations can undermine reliability and spread misinformation.Source: §5.5 · pp. 28–29

5.6. Verifiable Accountability and Citation Integrity System in LLMs

• Citation and reference systems can trace generated information to sources, supporting transparency, accountability, intellectual-property concerns, and verifiability.The discussion also calls for comprehensive documentation and accurate acknowledgment of contributors.Source: §5.6 · p. 29

5.7. Diversifying Case Studies in LLMs

• Case studies across healthcare, academia, education, management, training, and religion can inform domain-specific mitigation strategies and policymaking.The discussion argues that sectors present distinct ethical challenges.Source: §5.7 · p. 29

5.8. Decoding Censorship Complexity in LLMs

• Censorship must balance protection from harmful content with free speech, address possible training-data bias, and adapt to changing norms and regional laws.The discussion also identifies attempts to circumvent censorship as a source of insight into evasion strategies and countermeasures.Source: §5.8 · pp. 29–30

5.9. Breaking the Black Box and Crafting Dynamic Audit Tools for LLMs

• Case studies, mitigation, and accountability/governance receive greater scholarly attention, while censorship, transparency, and intellectual property/plagiarism show lower publication metrics.The discussion calls for more balanced exploration across ethical frameworks.Source: Figure 2 · p. 31, §VI · p. 31