Source-linked AI summary
Mapping the Ethics of Generative AI: A Comprehensive Scoping Review
Thilo Hagendorff
TL;DR
Recent generative AI applications have prompted ethical debates, but no meta-study had mapped their broad range of ethical issues. This scoping review systematically searched and coded the literature, finding a discourse dominated by risk concerns and often lacking empirical support.
Problem
Existing meta-studies covered related topics, but none targeted the broad assemblage of ethical issues associated with recent generative AI applications.
Source: §1 · p. 2Method
The authors conducted a scoping review using systematic literature searches and coding to synthesize the discourse and map normative concepts.
Source: §1 · p. 2Results
Ethical discussion is dominated by fairness, safety, harmful content, and hallucinations, while benefits receive less balanced attention and many concerns lack sufficient empirical or statistical backing.
Source: §6 · p. 11Takeaways & Limitations
The authors call for a critical mindset toward risk claims, noting that citation chains and popularity biases can amplify concerns lacking evidence of significance.
Source: §4 · p. 9Citation chaining was limited to checking paper titles in reference sections.
Source: §5 · p. 10
Abstract
from arXiv · showhide
The advent of generative artificial intelligence and the widespread adoption of it in society engendered intensive debates about its ethical implications and risks. These risks often differ from those associated with traditional discriminative machine learning. To synthesize the recent discourse and map its normative concepts, we conducted a scoping review on the ethics of generative artificial intelligence, including especially large language models and text-to-image models. Our analysis provides a taxonomy of 378 normative issues in 19 topic areas and ranks them according to their prevalence in the literature. The study 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. We discuss the results, evaluate imbalances in the literature, and explore unsubstantiated risk scenarios.
1 Introduction
The review addresses the absence of a meta-study covering the broad set of ethical issues associated with recent generative AI applications. It maps this discourse to provide a practical overview for relevant stakeholders.
1 Introduction
- AI ethics has shifted toward proactive attention to harms from misaligned AGI, while alignment and safety address technical and theoretical issues with normative grounding.The authors describe the field as having also turned toward practical implementation of principles and alternatives to principle-based approaches.Source: §1 · pp. 1–2
- Transformers and diffusion models improved text and image generation and enabled adaptation to many downstream tasks, helping generative models reach broader users through accessible interfaces.The paper links wider public uptake to natural-language prompts and easy-to-use browser interfaces.Source: §1 · p. 2
- The paper defines generative AI as large, foundation, or frontier models capable of transforming text, images, code, audio, video, or 3D content.The definition includes text-to-text, text-to-image, image-to-text, and other listed transformations.Source: §1 · p. 2
- No existing meta-study targeted the broad assemblage of ethical issues associated with the latest surge of generative AI applications.Earlier meta-studies covered topics such as policy guidelines, sociotechnical harms, and language-model risks.Source: §1 · p. 2
- The scoping review uses systematic searching and coding to synthesize the discourse, map normative concepts, and discuss imbalances.The authors aim to provide a basis for future research and technology governance.Source: §1 · p. 2
2 Methods
The authors searched for and screened literature on generative AI ethics, then inductively coded normative arguments to identify and organize recurring issues. They followed PRISMA and used multiple search sources, screening and follow-up procedures.
2 Methods
- The authors searched Google Scholar, arXiv, PhilPapers, and Elicit using 29 keywords and collected the first 25 or 50 results per search pass.They followed PRISMA and included papers from 2021 onward.Source: §2 · pp. 2–3
- The review included papers from 2021 onward and identified 179 documents for detailed content analysis after screening and additional-record searches.The initial search returned 1,674 results; 162 documents met the criteria for full-text screening, and 17 more were identified through citation chaining and monitoring.Source: §2 · p. 3
- They inductively coded normative arguments paragraph by paragraph, excluding purely descriptive or unrelated technical content and arguments about traditional machine learning.Coding continued until theoretical saturation; appendices were not annotated.Source: §2 · p. 3
- A second coding cycle checked consistency, resolved discrepancies, clustered similar codes, and organized the analysis into 378 distinct codes.The codes were grouped into high-level categories.Source: §2 · p. 3
3 Results
The review identifies 19 topic clusters, ranked by prevalence: fairness is discussed most, followed by safety. The clusters span technical risks and broader social, economic, environmental, and governance concerns.
Fairness - Bias
- Fairness is the literature’s most discussed issue, with concerns spanning biased outputs, concentrated power, unequal access, and culturally imposed values.Reported harms include stereotyping, racism, sexism, ideological leanings, and marginalization of minorities.Source: §3 · p. 5
Safety
- Safety ranks second, with literature addressing catastrophic risks from human-level or superhuman models, malicious misuse, and measures such as oversight and red teaming.Concerns include weaponization and biological-agent planning, alongside proposals for more safety research and safety cultures.Source: §3 · pp. 5–6
Harmful content; Cybercrime; Copyright - Authorship
- The literature discusses harmful and deceptive outputs, high-stakes advice, social-engineering abuses, and copyright disputes over training data.Examples include disinformation, fake news, deepfakes, and unauthorized collection of text or image training data.Source: §3 · pp. 6–8
Hallucinations; Privacy
- Hallucinations and privacy risks include false information or code, sensitive-data leakage, surveillance, and proposals to sanitize training data or use synthetic data.Some sources call for manual validation and fact-checking of model outputs.Source: §3 · p. 6
Interaction risks; Security - Robustness; Alignment
- Interaction, security, and alignment concerns cover excessive trust, manipulation, jailbreaking, guardrail circumvention, and uncertainty over whose values should guide models.Alignment discussions also address deceptive alignment, reward hacking, proxy gaming, and goal misgeneralization.Source: §3 · pp. 6–7
Education - Learning; Writing - Research
- Education discussions include cheating and personalized-learning opportunities, while writing and research concerns include academic integrity and calls for synthetic-text detectors.Some sources also worry about reduced learner effort and low-quality LLM-generated manuscripts.Source: §3 · pp. 7, 9
Governance - Regulation; Transparency - Explainability; Evaluation - Auditing
- Governance proposals include regulation and oversight, while transparency and evaluation discussions call for organizational openness, impact assessment, and third-party audits.The literature also notes a lack of safety benchmarking for languages other than English.Source: §3 · pp. 7–8
Labor displacement; Sustainability; Art - Creativity; Copyright - Authorship; Miscellaneous
- Other clusters address labor displacement, environmental resource costs, financial harms to artists, and infrequently or vaguely specified issues such as accountability and trustworthiness.Copyright discussions also cover model memorization or plagiarism and unresolved ownership of outputs.Source: §3 · pp. 8–9
4 Discussion
The generative AI ethics literature tends to emphasize harms over benefits, and some frequently repeated risks lack strong empirical support. The discussion also identifies gaps in the literature’s scope and cautions against letting speculative future risks skew assessments of present risks.
4 Discussion
- The literature gives greater or exclusive weight to risks and harms, while benefits are often overshadowed and some risks gain traction through citation chains despite limited evidence.The authors caution that this negativity bias can distort decision-making and that some concerns circulate without evidence of their significance.Source: §4 · p. 9
- The repeated claim that LLMs can aid pathogen creation [54,65,68,74,75] rests on only a handful of experiments, insufficient to show greater assistance than traditional search engines [155,156].The authors traced the claim through citation chains to assess its evidential basis.Source: §4 · pp. 9–10
- Evidence shows LLMs are notably ineffective at linking personal information to specific individuals [35,72,88,102,137,100], while emissions are lower than humans’ for equivalent tasks [25,35,57,139,157,158].The review notes that linking personal information to an individual would be necessary to establish an actual privacy violation.Source: §4 · p. 10
- The discourse often echoes earlier publications without empirical support, is anthropocentric, and rarely addresses multimodal models or agents [159–165].The authors also note that the literature concentrates on LLMs and text-to-image models, despite the ethical implications of multimodal models and agents for mediating human communication.Source: §4 · p. 10
- When discussion moves beyond LLMs and text-to-image models, it often turns to speculative AGI risks, which should not skew assessments of present and near-future risks.Examples include existential risks, deceptive alignment, power-seeking behavior, and shutdown evasion [36,54,66,67].Source: §4 · p. 10
5 Limitations
The review’s evidence base and findings are limited by its inclusion of some non-peer-reviewed preprints, constraints on citation chaining, and search choices. As a scoping review, it also does not trace changing ethical debates over time or resolve conflicts between positions.
5 Limitations
- Some non-peer-reviewed preprints, primarily from arXiv, were included despite being considered poor quality, so poorly researched claims may enter the analysis.They were retained because they met the inclusion criteria.Source: §5 · p. 10
- Citation chaining checked reference-list titles rather than all corresponding abstracts, so relevant sources may have been missed.Reading the abstracts would have made the search more thorough but was considered too labor-intensive.Source: §5 · p. 10
- Screening only the first 25 results per search term, or 50 for Elicit, could omit later-listed sources; search strategies and terms also affect topic proportions and quantitative results.The review notes that its collection choices influence the distribution of topics it reports.Source: §5 · p. 10
- The scoping review cannot depict ethical debates unfolding over time, and resolving conflicts between normative positions is beyond its scope.Source: §5 · p. 11
6 Conclusion
The review maps 378 normative issues across 19 topic areas, with fairness, safety, harmful content, and hallucinations prominent in the discourse. It also notes that positive impacts receive less balanced consideration and calls for more empirically grounded analysis.
6 Conclusion
- 378 normative issues across 19 topic areas are mapped, with fairness, safety, harmful content, and hallucinations prominent.The review makes its complete taxonomy available.Source: §6 · p. 11
- The discourse often gives less balanced consideration to generative AI’s potential benefits in creativity, productivity, education, and other fields.The authors note that concerns can recur without sufficient empirical or statistical backing.Source: §6 · p. 11
- More empirical grounding and balanced analysis are described as essential to understanding the ethics landscape.The authors emphasize that this critique does not diminish the importance of ethics research in responsible engagement with generative AI’s potential.Source: §6 · p. 11
Appendix A
Appendix A documents search dates, keyword combinations for Google Scholar, arXiv, and PhilPapers, and tailored Elicit prompts. The listed queries cover ethics and risk terms alongside generative-AI, language-model, and image-generation terminology.
Appendix A
- Searches ran on October 30, 2023 for Google Scholar, arXiv, and Elicit, and on December 21, 2023 for PhilPapers; literature monitoring continued until January 22.The searches avoided combining AND and OR in a single term because search engines handle Boolean precedence differently.Source: Appendix A · p. 22
- Risk queries pair “risk*” with generative AI, language models, image generation, text-to-image, image synthesis, and Stable Diffusion.The appendix presents the keyword combinations alongside a separate set of tailored Elicit prompts.Source: Appendix A · p. 22
- Ethics queries pair terms such as “ethic*” with generative AI, AI-generated content, language models, image generation, text-to-image, image synthesis, and Stable Diffusion.These keyword combinations are listed for Google Scholar, arXiv, and PhilPapers.Source: Appendix A · p. 22, Table 2 · p. 23
Appendix B
Appendix B defines which scientific papers were eligible for the review, focusing on substantive ethical analysis of generative AI. It excludes purely technical or non-generative-AI ethical work, as well as several other formats and subjects.
Appendix B
- Eligible papers had to be in English and analyze ethical implications or normative dimensions of generative AI or its subfields in multiple paragraphs.Papers that merely touched on ethical topics did not meet this criterion.Source: Appendix B · p. 23
- The review excluded purely technical works, ethics or philosophy papers without a generative-AI reference, and non-scientific publication formats.The excluded formats included books, websites, policy and white papers, theses, interviews, and editor messages.Source: Appendix B · p. 23
- Papers about moral reasoning in AI systems, surveys of public opinion on generative AI, or material unavailable online were excluded..Source: Appendix B · p. 23
Appendix C
Appendix C identifies additional records from citation chaining and from monitoring the literature after the initial search. The supplied passages describe these two retrieval routes but do not provide the records’ details.
Appendix C
- Table 4 is titled as a list of additional records identified through citation chaining.The review’s citation-chaining process sifted reference lists until no additional publication could be identified.Source: Table 4 · p. 24
- Table 5 is titled as a list of additional records identified by monitoring literature after the initial search..Source: Table 5 · p. 24
Appendix D
Appendix D points readers to the complete list of references included in the content analysis.
Appendix D
- The complete list of references included in the content analysis is available here.Source: Appendix D · p. 25