Across the six European use cases, AIOLIA co-creation exercise resulted in 12 contextualized ethics principles, broken into 37 components which covered 30 unique aspects. Generally, the ethics principles in use cases align well with the overarching ALTAI framework but also show important patterns and deviations (see Table below). The direct comparison of ALTAI with the UC-specific ethics principles demonstrates that the bottom-up process conducted in AIOLIA foregrounds very similar ethics concerns as established frameworks. The high-level principled frameworks such as ALTAI thus seem well-reflected in practitioner discussions about AI ethics. However, we also observed that across the diverse contexts, different focus and emphasis was given to either overall ethics concerns (e.g., human oversight) or specific sub-aspects (e.g., auditability, deskilling, safety). Moreover, we found that the same ethics consideration could be seen as either an overarching principle or as an aspect of other ethics principles (e.g., transparency as part of non-bias), and that the same component could be linked to different ethics principles (e.g., accuracy appeared as component in human oversight and non-maleficence, while auditability emerged as a component in robustness/reliability, non-maleficence, and accountability). Together, these observations illustrate that in AI research and design perspectives the status of principles and the exact relations between the components may be much more fluid than in more abstract AI ethics frameworks.
| UC | Ethics principles in use cases | ALTAI Principles (Requirements) |
|---|---|---|
| UC-principles covered in ALTAI | ||
| UC2 | Robustness and reliability | Req #2, Technical robustness and safety |
| UC5 | Privacy and data protection | Req #3, Privacy and data governance |
| UC1, UC3 | Transparency and explainability | Req #4, Transparency |
| UC3, UC4 | Non-bias, fairness and non-discrimination | Req #5, Diversity, non-discrimination and fairness |
| UC1, UC4, UC6 | Accountability and responsibility | Req #7, Accountability |
| UC-principles address sub-parts of ALTAI principles: UCs each consider disparate aspects of Req #1 or sub-aspects as distinct ethics principle | ||
| UC2 | Human oversight | Req #1, Human agency and oversight |
| UC5, UC6 | Autonomy/User agency | Req #1, Human agency and oversight |
| UC2, UC3 | Over-reliance and deskilling | Req #1, Human agency and oversight |
| UC-principles named in different ways but addressing aspects similar to ALTAI | ||
| UC1, UC6 | Non-maleficence Focus: Covers important aspects within General Safety |
Req #2, Technical robustness and safety |
| UC4 | Freedom of expression and non-censorship Focus: Covers important aspects of Oversight |
Req #1, Human agency and oversight |
| UC-principles named in similar ways but addressing aspects different from ALTAI | ||
| UC5 | Safety/Human safety Difference: Addresses primarily safety of users rather than safety of AI system |
Req #2, Technical robustness and safety |
| UC5 | Human well-being Difference: Addresses individual well-being, rather than broader societal or environmental issues |
Req #6, Environmental and societal well-being |
Across AIOLIA use cases UC1 to UC6, 37 components emerged, covering 30 unique aspects. The table below lists the components identified for each of the ethics principles. The colour-coding illustrates the degree of inconsistency in positioning components within ethics principles. In the table, overlaps are indicated by identical colour, e.g., mentions of ‘Privacy, consent and data protection’ are highlighted in light blue, mentions of ‘Transparency and explainability’ are marked in light green.
| Ethics principle from the use cases | Components |
|---|---|
| Non-bias, fairness and non-discrimination | Diversity Representativeness/inclusivity Objectivity Non-stigmatising use / proportionality Equality and impartiality Transparency of criteria |
| Accountability and responsibility | Auditability Human oversight Liability (responsibility) Human agency and responsibility Professional competence Auditability and evaluation Responsiveness |
| Privacy, consent and data protection | User consent and transparency Data Minimisation, data use and storage Third-party sharing and compliance |
| Autonomy | Transparency and User understanding Privacy Risk of over-attachment and dependency Informed Consent System Customisation |
| Human oversight | Validity / accuracy Bias Privacy |
| Transparency and explainability | Accessibility / access to information Explainability Justifiability Openness Documentation, traceability, and auditability |
| Over-reliance and deskilling | Dependence Contestability and human oversight Preservation of human skill and expertise Feedback and learning loops for human adaptation Organisational policies for shared responsibility |
| Freedom of expression, non-censorship | Autonomy and agency Proportionality Non-discrimination |
| Robustness/reliability | Auditability Human oversight Liability (responsibility) |
| Non-maleficence | Subsidiarity and proportionality Effectiveness Societal well-being Validity / accuracy (from technical perspective) Bias Privacy |
| Safety/human safety | User protection Security Measures Human Oversight |
| Human well-being | Promotion of User’s health Scope boundaries Crisis Recognition |
Overlaps can be found for seven of the twelve ethics principles: (1) Non-bias, fairness and non-discrimination; (2) Accountability and responsibility; (3) Privacy, consent and data protection; (4) Autonomy; (5) Human oversight; (6) Transparency and explainability; (7) Over-reliance and deskilling.
The most noticeable inconsistencies emerged for ethics principles that were also listed as components:
- Human oversight, while being a principle itself, is being listed as component in four other principles: Accountability and responsibility; Over-reliance and deskilling; Robustness/reliability; Robustness/reliability
- Privacy is listed as component in the three principles: Autonomy; Non-maleficence; Safety/Human Safety
- Transparency is listed as a component in the three principles: Non-bias, fairness and non-discrimination; Privacy, consent and data protection; Autonomy
This clearly illustrates that the same ethics consideration (e.g., transparency) can be seen as either overarching principle or as aspect of other ethics considerations (e.g., transparency as part of non-bias).
Moreover, across use cases the same component could be linked to different ethics principles. For instance, accuracy appeared as component in human oversight and non-maleficence, while auditability emerged as a component in robustness/reliability, non-maleficence, and accountability. Together, these observations illustrate again that in end-user perspectives the status of principles versus components and the exact relations of components may be much more fluid than in more abstract representations within AI ethics frameworks.
Source: AIOLIA deliverable 3.1