Below we summarize the fifth risk framework included in the AI Risk Repository: “A framework for ethical AI at the United Nations” by Thilo Hagendorff
This paper conducts a scoping review on the ethics of general artificial intelligence, including large language models and text-to-image models. This analysis provides a taxonomy of 378 normative issues in 19 topic areas and ranks them according to their prevalence in the literature. This taxonomy is available online: https://www.thilo-hagendorff.info/ethics-of-generative-ai/tree.html.
These 19 topic areas are:
Fairness - Bias
Generative AI risks perpetuating biases like racism and sexism, leading to unequal access and marginalization.
Safety
AI safety focuses on preventing catastrophic risks like deceptive behavior and ensuring responsible AI development.
Harmful Content - Toxicity
Generative AI can create harmful content, including disinformation, deepfakes, and violent material.
Hallucinations
LLMs sometimes generate false or misleading information, posing risks in critical areas like healthcare.
Privacy
Generative AI threatens privacy through data leaks and misuse of sensitive information from training data.
Interaction Risks
Human interactions with AI can lead to over-trust, manipulation, and negative impacts on mental health.
Security - Robustness
Generative AI is vulnerable to security threats like prompt injection and adversarial attacks that bypass safeguards.
Education - Learning
Generative AI enhances personalized learning but raises concerns about cheating and academic integrity.
Alignment
Ensuring AI aligns with human values is challenging due to the ambiguity around whose values should guide it.
Cybercrime
Generative AI is used for cybercrime, enabling social engineering, phishing, and creating malicious code.
Governance - Regulation
There is a need for international AI regulation, but overregulation may hinder innovation.
Labor Displacement - Economic Impact
Generative AI could displace jobs across sectors, potentially increasing socioeconomic inequality.
Transparency - Explainability
AI transparency requires explaining model decisions and organizational openness about AI use.
Evaluation - Auditing
Auditing AI systems is essential for ensuring fairness, safety, and preventing harm.
Sustainability
Generative AI has a significant environmental impact, driving calls for more sustainable practices.
Art - Creativity
Generative AI disrupts human creativity and raises concerns about the financial impact on artists.
Copyright - Authorship
Generative AI challenges copyright norms, raising questions about ownership and ethical data use.
Writing - Research
AI affects writing quality and research integrity, with concerns about AI-generated papers.
Miscellaneous
Some less-discussed topics include trustworthiness, accountability, and AI’s role in socio-political instability.
Offers a comprehensive overview for scholars, practitioners, or policymakers, condensing the ethical debates surrounding fairness, safety, harmful content, hallucinations, privacy, interaction risks, security, alignment, societal impacts, and others
Evaluates imbalances in the literature, such as the disproportionate focus on risks, and explores unsubstantiated risk scenarios
Hagendorff, T. (2024). Mapping the ethics of generative ai: A comprehensive scoping review. arXiv preprint arXiv:2402.08323.