While AI ethics principles have become the bedrock of Trustworthy AI, organisations face significant challenges in applying and putting these theoretical ethical principles into practice. Addressing this issue, AIOLIA’s industrial partners engaged in the operationalisation of ethics principles, that is, in translating high level ethical principles into practical actions, tools, processes and governance structures that can guide and be applied throughout the lifecycle of AI systems. This process has resulted in a wealth of different approaches and measures that have been systematised in the development of context-enriched organisational guidelines for different types of AI systems and AI research areas, namely General-Purpose AI, Emotional AI and Decision Support Systems.
AIOLIA’s organisational guidelines provide a set of actionable and non-technical measures that can be implemented by organisations with the aim of 1) promoting responsible AI practices, 2) building trust in AI uptake, and 3) aligning organisational practices with legal requirements and standards, such as the EU AI Act, the GDPR and ISO/IEC42001. Organisational measures focus on how an organisation can incorporate and manage ethical AI practices through structures, policies and governance frameworks.
The intended audiences of these guidelines are:
- Organisations that deploy GPAI systems;
- Organisations that provide GPAI models or systems;
- Policymakers working in the field of Artificial Intelligence.
These guidelines are informed by the operationalisation of AI ethics principles across six European use cases, operating across areas as diverse as healthcare, automotive engineering, detection of harmful and illegal content, and personal companions. Given the different areas of deployment covered, the operationalisation process undertaken in the different industrial use cases has been informed by sectoral regulatory frameworks and standards, such as ISO automotive industry standards or the Medical Device Regulation, in addition to the AI Act, resulting in a rich guidance.
Below is the list of organisational measures identified for the three different AIOLIA research areas. To understand the measures in full, attending to the relevance, procedures and implementation challenges faced by organisations operating in the different research areas, consult our full deliverable here.
| Measure | Description | GPAI | Emotional AI | DSS |
|---|---|---|---|---|
| Context | ||||
| OM 01 | Define clear boundaries and guidance for intended use of the AI system, including when AI outputs should be questioned, verified, or supplemented with human input. | X | X | |
| OM 02 | Define clear rules on the type of content, behaviour, or expression that are restricted for AI use and why, including distinguishing legitimate behavioural influence from manipulative practices. | X | ||
| OM 03 | Define what constitutes vulnerability in users or patients, including static and emergent forms, set out identification protocols, and establish corresponding enhanced safety measures, interaction protocols and continuous monitoring of effects for identified vulnerability profiles. | X | ||
| OM 04 | Define multi-level safety criteria and categories of risks and harms. | X | ||
| OM 05 | Consider diversity criteria and their intersection across roles, locations, languages, or cultural characteristics. | X | X | |
| Leadership | ||||
| OM 06 | Promote a safety-first culture within the organisation. | X | X | |
| OM 07 | Establish a representative AI governance committee as a formal oversight structure within the organisation. | X | X | X |
| Planning | ||||
| OM 08 | Explicitly define and document responsibility for AI outputs. | X | X | |
| OM 09 | Clearly define oversight responsibilities and embed them into workflows, including checkpoints when active human professional engagement is required. | X | ||
| OM 10 | Establish a clear escalation path for the organisation hierarchy to decide on responsibility assignments during AI system design or operation. | X | ||
| OM 11 | Define a clear policy on high-risk or sensitive decisions that must involve human judgment. | X | ||
| OM 12 | Conduct a privacy impact assessment. | X | ||
| OM 13 | Involve mental health professionals in the design, monitoring and evaluation of the AI system. | X | ||
| OM 14 | Determine the permitted degree of autonomy of the AI system (if any) and accordingly, specify the conditions in which a professional in the loop is required, e.g. via human moderation. | X | ||
| OM 15 | Consider psychological, emotional, and behavioural influences in AI risk analysis, not only user or patient information or technical AI performance. | X | ||
| Support | ||||
| OM 16 | Conduct trainings to support users in understanding AI capabilities and limitations, as well as related human capabilities and limitations. | X | X | |
| OM 17 | Clearly communicate how the AI system may affect human individuals, next of kin and society at large. | X | ||
| Operation | ||||
| OM 18 | Put in place a mechanism for users to report, question, or contest AI behaviour, decisions and/or restrictions. | X | X | X |
| OM 19 | Design explanations that enable users to understand key behavioural factors, limitations, and uncertainties in their interaction with the AI system. | X | ||
| OM 20 | Document and enable traceability of changes to models, data or the GPAI system with sufficient detail to support internal reviews, external audits, or regulatory scrutiny. | X | ||
| OM 21 | Implement additional safeguards for sensitive data. | X | ||
| OM 22 | Ensure the AI system, if autonomous, applies behavioural nudging only in documented and auditable contexts with specified trigger thresholds and objectives. | X | ||
| OM 23 | Conduct internal ethics reviews for new features or changes of the AI system. | X | ||
| OM 24 | Design explanations that support meaningful review, contestation or justification of AI-supported decision. | X | ||
| Evaluation | ||||
| OM 25 | Periodically reassess reliance patterns as systems evolve or scale. | X | X | |
| OM 26 | Monitor situations when the purpose or actual use of the AI system drifts or diverges from the intended ones, and informs users when relevant. | X | X | |
| OM 27 | Put in place auditable Standard Operating Procedures for AI design and validation. | X | ||
| OM 28 | Conduct periodic independent ethics reviews of the AI system’s impact on wellbeing and autonomy addressing all stages of the AI lifecycle. | X | ||
| OM 29 | Conduct audits of informed consent mechanisms and document limitations, especially with regard to the adaptivity to context. | X | ||
| OM 30 | Review AI outputs for unintended disparity impact. | X | ||
| Improvement | ||||
| OM 31 | Translate audit findings into corrective design action. | X | X | |
| OM 32 | Identify and analyse unforeseen effects of the AI system on individual and societal well-being. | X | X | X |
| OM 33 | Assess and collect feedback from users of conversational systems regarding perceived honesty, non-coerciveness of AI interactions and impact on wellbeing. | X | ||
Source: AIOLIA deliverable 3.3