Source-linked AI summary
AI-Ethics by Design. Evaluating Public Perception on the Importance of Ethical Design Principles of AI
Kimon Kieslich, Birte Keller, Christopher Starke
TL;DR
Public perceptions of ethical AI principles are under-researched despite the importance of human-centric design and unavoidable trade-offs. Using a conjoint survey of 1,099 Germans focused on tax-fraud detection, the paper compares seven principles and identifies aggregate and subgroup preference patterns. Respondents viewed the principles as broadly balanced overall, while clusters differed in preferred attributes and attribute importance; the findings are limited to the studied use case and German population.
Problem
Empirical evidence on how the public comparatively assesses ethical AI principles is limited, despite the human-centric goal of accounting for affected people’s views.
Method
A conjoint survey of 1,099 German respondents evaluated seven ethical principles in tax-fraud-detection AI, followed by k-means clustering of preference patterns.
Results
German respondents considered ethical principles broadly balanced overall, with accountability slightly highest and machine autonomy slightly lowest; clusters showed distinct preference models.
Takeaways & Limitations
Developers and organizations should not neglect particular ethical principles, while public preference profiles indicate that ethical design priorities differ across groups.
Takeaways & Limitations
The findings are limited to AI tax-fraud identification in Germany, and importance weights or preference profiles may differ across use cases and countries.
Abstract
from arXiv · showhide
Despite the immense societal importance of ethically designing artificial intelligence (AI), little research on the public perceptions of ethical AI principles exists. This becomes even more striking when considering that ethical AI development has the aim to be human-centric and of benefit for the whole society. In this study, we investigate how ethical principles (explainability, fairness, security, accountability, accuracy, privacy, machine autonomy) are weighted in comparison to each other. This is especially important, since simultaneously considering ethical principles is not only costly, but sometimes even impossible, as developers must make specific trade-off decisions. In this paper, we give first answers on the relative importance of ethical principles given a specific use case - the use of AI in tax fraud detection. The results of a large conjoint survey (n=1099) suggest that, by and large, German respondents found the ethical principles equally important. However, subsequent cluster analysis shows that different preference models for ethically designed systems exist among the German population. These clusters substantially differ not only in the preferred attributes, but also in the importance level of the attributes themselves. We further describe how these groups are constituted in terms of sociodemographics as well as opinions on AI. Societal implications as well as design challenges are discussed.
1 Introduction
Ethical AI is intended to be human-centric and beneficial to society, but its principles can conflict in practice. This makes affected publics’ opinions relevant to ethical design decisions.
- AI development is widely expected to be human-centric, trustworthy, and beneficial to the whole society.
- Conflicts between ethical principles, such as explainability and accuracy, can require concrete design trade-offs.
- Taking ethical AI seriously requires decision makers to consider the opinions of affected members of the public.
- The study compares seven ethical principles in AI tax-fraud detection through a conjoint survey of 1,099 participants.
2 Ethical Guidelines of AI Development
AI guidelines promote human-centered development, but their principles create practical trade-offs because ethical objectives cannot always be maximized simultaneously.
- Ethical AI guidelines seek to address societal challenges through human-centered or human-centric AI.
- Guideline reviews identify recurring principles including transparency, fairness, responsibility, privacy, safety, autonomy, and beneficence.
- The paper operationalizes seven prominent principles as explainability, fairness, security, accountability, accuracy, privacy, and limited machine autonomy.
- Security emphasizes technical protection against attacks, while autonomy concerns manipulation, monitoring, and human oversight in decisions.
- Ethical principles can conflict because designing systems that maximize all ethical aspects simultaneously is often infeasible.
- Higher accuracy may require complex models that are difficult for laypersons to understand, illustrating an accuracy–explainability trade-off.
3 Public Preferences for AI Ethics Guidelines
Research has found that people value ethical principles in AI, but comparative evidence is limited and produces mixed rankings. This motivates examining public trade-off patterns.
- Public concerns about AI include privacy violations, unfair outcomes, removing humans from crucial decisions, and failure to capture human complexity.
- Prior evidence links accuracy, fairness, and human oversight to public evaluations of AI systems, trust, legitimacy, or institutional consequences.
- Few studies compare ethical indicators directly, and existing findings disagree about whether fairness, transparency, accountability, or explainability matters most.
- The study asks how ethical-principle configurations influence prioritization and which preference patterns exist among the German public.
- Because public preferences may vary across social settings and beliefs, the study also examines characteristics associated with specific ethical design preferences.
4 Method
The study uses a German online conjoint survey to evaluate preferences across seven ethical attributes in AI tax-fraud detection. Respondents rated randomized system configurations, and related AI attitudes were measured separately.
- 4.2 Procedure: The conjoint experiment presented eight randomized tax-fraud-detection system cards whose configurations varied across seven ethical principles.
- 4.2 Procedure: The seven attributes were explainability, fairness, security, accountability, accuracy, privacy, and limited machine autonomy.
- 4.2 Procedure: Conjoint analysis forces choices among imperfect ethical configurations and supports preference estimation for combinations not all directly rated.
- 4.2 Procedure: A fractional factorial design generated balanced attribute combinations for the system cards.
5 Results
The study estimates respondents’ ethical-design preferences by fitting respondent-level regressions to conjoint ratings and averaging the resulting part-worth values.
- 1,099 respondent-level regression models estimated ethical-design preferences from conjoint card ratings using dummy-coded attributes.The regression coefficients, called part-worth values, were averaged to characterize the German population’s preferences.
5.2 Part-Worth of Attributes
Respondents rated accountability highest on average, while the seven ethical attributes were broadly balanced overall and machine autonomy ranked lowest.
- All ethical principles except limited machine autonomy positively influenced AI-system approval, with accountability having the largest part-worth at bAccountability=0.8.The reported part-worths were bAccuracy=0.64, bExplainability=0.57, bFairness=0.66, bAutonomy=0.32, bPrivacy=0.66, and bSecurity=0.66.
- Accountability ranked highest, while fairness, security, privacy, and accuracy were equally important on average.
- Machine autonomy was least important, explainability was slightly less important, and aggregate attribute weights remained relatively balanced.
5.3 Preference Patterns Among the Public
Cluster analysis identified five public preference profiles that differed in which ethical principles they valued and how strongly those principles affected system approval.
- 5.3 Preference Patterns Among the Public: The five clusters comprised Human in the Loop, Ethically Concerned, Safety Concerned, Fairness Concerned, and Indifferent preference profiles.The profiles respectively emphasized human control, all principles, safety-related principles, fairness and accuracy, or limited concern for ethical design.
- 5.3 Preference Patterns Among the Public: 345 respondents (31.39%) formed the largest Ethically Concerned cluster, treating all ethical principles as highly important and giving low approval when systems failed to satisfy them.
- 5.3 Preference Patterns Among the Public: 267 respondents (24.29%) formed the second-largest Indifferent cluster, whose approval was only slightly affected by ethical design and remained medium across systems.
- 5.3 Preference Patterns Among the Public: The Safety Concerned cluster comprised 167 respondents (15.2%) who prioritized safety, privacy, and accountability over fairness, accuracy, and explainability.Their approval ratings across presented systems were low to medium.
- 5.3 Preference Patterns Among the Public: 166 respondents (15.1%) were Fairness Concerned, emphasizing fairness and accuracy, while privacy and security had little effect on positive ratings.
- 5.3 Preference Patterns Among the Public: 154 respondents (14.01%) formed the Human in the Loop cluster, for whom opposition to machine autonomy made human control particularly relevant.Approval of the presented systems was medium on average.
5.4 Cluster Description
The five clusters differ significantly across sociodemographic characteristics and opinions on AI, revealing distinct profiles of ethical-AI preferences.
- 5.4 Cluster Description: All analyzed characteristics significantly differ across clusters, with mean scores and group comparisons reported in Table 3.The overall MANOVA was significant, followed by ANOVAs and Tukey-HSD tests for each dependent variable.
- 5.4 Cluster Description: The Human in the Loop group demands human control, opposes machine autonomy, and combines older age and lower education with low AI acceptance and trust.This group also reports comparatively high risk awareness and low perceived benefits of AI.
- 5.4 Cluster Description: The ethically concerned group is comparatively young and well educated, with high AI interest and trust, high acceptance and benefit perception, and low risk perception.The supplied passages describe this group as demanding high standards across all ethical principles.
- 5.4 Cluster Description: The Safety Concerned group occupies an intermediate profile, with average age and education and moderate AI interest, acceptance, risk awareness, benefit perception, and trust.Its opinions on AI are described as between those of the other groups.
- 5.4 Cluster Description: The Fairness Concerned group is young and well educated, and reports the highest trust, acceptance, and perceived benefits alongside the lowest perceived risks.This group is concerned with AI accuracy and fairness and has among the highest levels of AI interest.
- 5.4 Cluster Description: The Indifferent group, like Human in the Loop, has relatively negative AI opinions, including low acceptance, benefits, and trust but high risk awareness and little interest.This group does not demand ethically designed systems.
6 Discussion
The discussion argues that public perceptions should inform ethical AI design and legal frameworks, while recognizing substantial preference differences and important scope boundaries.
- 6.1 From Ethical Guidelines to Legal Frameworks?: German respondents broadly valued all ethical principles similarly, with accountability slightly preferred and limited machine autonomy slightly less important.The aggregate pattern suggests balanced expectations rather than a strong preference for one principle.
- 6.2 Ethical Design and Demands of Potential Stakeholder Groups: Five preference groups differed considerably, showing that no universal balance of ethical principles exists across the German population.The groups also differed in their preferred attributes and overall importance levels.
- 6.2 Ethical Design and Demands of Potential Stakeholder Groups: 24% of respondents were Indifferent, showing little concern for ethical design and potentially creating weak public pressure for companies to implement it rigorously.This group nevertheless had the highest average approval scores for the presented systems.
- 6.2 Ethical Design and Demands of Potential Stakeholder Groups: The largest group equally valued all principles, setting high standards that may make ethical trade-offs difficult and potentially reduce AI acceptance.This group was characterized by higher education, younger age, stronger AI interest, and higher AI acceptance.
- 6.2 Ethical Design and Demands of Potential Stakeholder Groups: Public preferences should complement rather than dominate other AI-development perspectives because people may favor understandable systems over more substantively fair complex systems.The discussion therefore highlights public input while warning that public judgments may not always identify the most ethically thorough design.
- 6.3 Limitations: The findings are limited because the study used one artificial tax-fraud use case and surveyed only German respondents, while AI perceptions are context dependent.The authors call for simultaneous comparisons across use cases and replication in other countries.
7 Conclusion
The conclusion presents ethical AI as a major societal challenge and reports that ethical requirements matter for most German citizens, although a notable minority does not demand them.
- 7 Conclusion: Ethical AI is a major societal challenge, and taking ethical requirements seriously matters for gaining broad acceptance among German citizens.The conclusion also identifies public interest in AI as a possible way to raise demands for higher ethical standards.
- 7 Conclusion: A notable portion of the German population does not demand ethical AI implementation, which the authors regard as critical because implementation depends partly on public demand.The conclusion links this concern to the expense of ethical AI development.
- 7 Conclusion: People demanding high ethical standards were also interested in AI and aware of its risks, suggesting that raising public interest could strengthen expectations for ethical AI.This is presented as a promising direction rather than an established causal effect.
Funding
The study was conducted within the Meinungsmonitor Künstliche Intelligenz project, funded by North Rhine-Westphalia’s Ministry of Culture and Science.
- The study was conducted as part of the Meinungsmonitor Künstliche Intelligenz project.
- The project was funded by the Ministry of Culture and Science of the State of North Rhine-Westphalia.
- The funding ministry is located in North Rhine-Westphalia, Germany.
Appendix
The appendix documents the conjoint-analysis orthoplan and the wording of survey items.
- The appendix includes the orthoplan used for the conjoint analysis.
- The appendix provides the wording of the survey items.