In the process of operationalising AI ethics principles conducted by AIOLIA industrial partners, several open issues emerged according to the specificities of the different types of AI technologies used, which correspond to the three AIOLIA research areas, namely, General-Purpose AI (GPAI), Emotional AI and Decision Support Systems. This is an overview of the key open issues in the governance of AI for General Purpose AI, which acknowledges both the role and limits of AIOLIA’s organisational measures in addressing these. For a complete analysis see AIOLIA Deliverable 3.3.
General-purpose AI (GPAI) is a category of AI models and systems that display significant generality, are capable of competently performing a wide range of distinct tasks and can be integrated into a variety of downstream applications. These advanced AI systems exhibit a significant degree of autonomy and the ability to generalise to new tasks and across domains the system has not been previously exposed to or intentionally trained to address. Despite GPAI systems being characterised by a high degree of operational opacity, their utility and ability to perform a wide range of tasks, identifying patterns across multimodal datasets, has motivated GPAI uptake across private and professional contexts.
As a research area, GPAI presents unique ethical challenges. On the one hand, the generality of tasks GPAI models can perform, both in professional and private contexts, means that there is no single, standard range of ethical principles and measures that would apply to this research area. On the other hand, the fact that GPAI models, as foundation models developed by third parties, often form the baseline architecture upon which different applications are built, fine-tuning the foundation model to perform in specific contexts, poses new challenges to the operationalisation of ethics-by-design in the contexts of model deployment. Below are some of the key open issues for the governance of GPAI faced by organisations and the ways in which AIOLIA’s organisational guidelines can support addressing these.
| Open Issue & Concerns | How Organisational Measures Help |
|---|---|
| Open issue #1: How can organisations govern emergent risks that arise from GPAI deployment rather than at the system development stage? | |
|
Concern:
GPAI systems are highly malleable and their behaviour evolves dynamically through human-AI interaction. This means risks cannot be fully anticipated or mitigated through ex-ante approaches alone, and may only materialise within specific deployment contexts, including non-technical harms such as deskilling, that fall outside conventional risk frameworks. Risks stem not just from technical robustness, but from the unpredictable, emergent dynamics of human-AI interaction. |
Organisations can define clear boundaries for intended use and monitor for drift or divergence from those boundaries. Internal ethics reviews for new features, combined with traceability documentation of system changes, allow organisations to detect and respond to emergent risks. Governance committees provide a structural safeguard against the organisational pressures that can accelerate deployment at the expense of safety. |
| Open issue #2: How should organisations address deskilling and the long-term erosion of human competencies resulting from GPAI use? | |
| Concern: The reliance on natural language interfaces shields users from cognitively engaging with tasks, and productivity gains from GPAI deployment may come at the cost of gradual erosion of skills. This risk compounds over time and may become embedded in organisational practices before it is recognised, undermining not only professional competence but the quality of human oversight itself and safety of AI-based workflows and decisions. | Organisations can establish governance committees that act as safeguards against unrealistic KPIs and workloads, implement role-based training programmes that preserve and update human competencies, and periodically reassess reliance patterns as systems evolve. Oversight responsibilities should be explicitly embedded in workflows so that meaningful human engagement is maintained by design rather than assumed. |
| Open issue #3: How can governance of GPAI be reoriented beyond safety- and compliance-centric frameworks to address individual and societal well-being? | |
| Concern: The dominant safety agenda in GPAI governance is largely framed around catastrophic or existential risks, leaving out a broad range of current and concrete harms, including algorithmic discrimination, cognitive impacts, and erosion of social trust. This framing risks making governance appear futile and out of reach outside the realm of frontier AI providers, narrowing the aperture of ethical deliberation in ways that exclude actual deployment contexts. | Organisations can adopt ethics-by-design approaches, including diversity and fairness reviews, impact assessments on societal well-being, and deliberative processes with civil society and impacted communities. Participation in broader research consortia and the publication of research findings contribute to widening the evidence base beyond capability-focused development. |
| Open issue #4: How can the absence of stable regulatory frameworks and harmonised standards be managed at the organisational level? | |
| Concern: Organisations operating in less standardised domains lack a clear baseline for translating ethical principles into practice. The AI Act’s obligations are largely directed at providers of high-risk systems, leaving deployers with significant discretion in determining what responsible deployment looks like. The absence of harmonised standards for GPAI, including in emerging areas such as Emotional AI, makes consistent and auditable governance difficult to achieve. | Organisations can adopt governance practices that go beyond legal compliance, effectively self-applying high-risk standards in the absence of a regulatory requirement to do so. Governance committees and independent ethics reviews can contribute to the development of sector-specific best practices, although these also require significant resources that might not be available among small organisations. Sharing safety criteria and research findings with professional and academic communities supports the emergence of shared norms. |
| Open issue #5: How should the environmental impact of GPAI be governed? | |
| Concern: There is growing evidence of the significant environmental costs of GPAI, including energy and water consumption and raw material use associated with model training and inference. However, environmental concerns have not emerged as a central issue in AIOLIA use cases, not because of a lack of awareness, but because organisations tend to prioritise the challenges they can directly tackle, such as safety, oversight, and skill preservation. Standard indicators for measuring environmental impact remain absent, and the trade-offs between competitiveness and environmental sustainability have yet to be clearly articulated at the policy level, leaving organisations without actionable guidance. | Environmental concerns are largely beyond the reach of individual organisations and require intervention at the regulatory and policy level, including the formulation of standard measurement indicators and clear sustainability priorities. Where possible, organisations can contribute by participating in broader research and policy communities and factoring environmental considerations into procurement and system design decisions. This underscores the limits of organisational governance and the need for complementary public and regulatory action. |
Source: AIOLIA deliverable 3.3