Photo by Umberto on Unsplash
National, supranational, and professional institutional actors, such as the EU, the OECD, and engineering, legal, and medical associations, shape what successful AI means in practice. FORSEE researchers analysed over 80 of their documents to understand their views. They share the assumption that economic benefits can be secured while protecting the public from harm.
However, the documents also reveal where that assumption runs into difficulty, and what it quietly leaves out. The clearest place to see this is in how European AI regulation has changed its understanding of trustworthy AI.
The regulatory turn: from moral reflection to procedural compliance
Between 2018 and 2025, European AI governance converted aspirations for trustworthy technology into procedural compliance. The EU’s approach now shapes AI regulation beyond Europe’s borders.
The 2019 Ethics Guidelines for Trustworthy AI, published by the EU’s High-Level Expert Group, were deliberately inclusive. Human agency, fairness, transparency, and societal well-being were framed as foundations for responsible innovation. Other supranational bodies, including UNESCO and the OECD, similarly grounded their guidelines in human rights, environmental protection, and sustainable development.
The EU AI Act, adopted in 2024, changed this. Many of the same terms appeared, but now situated within a risk-based regulatory framework modelled on existing EU product health and safety law. The AI Act turned many ethical principles into concrete regulatory obligations. In doing so, it also narrowed the range of questions regulation directly addresses. Notably, sustainable development, democracy, and social outcomes played a much smaller role in the AI Act than in previous documents.
“Sustainable development, democracy, and social outcomes had a much smaller role in the AI Act compared to previous documents.”
Technical standards add another layer to AI governance. ISO/IEC’s dedicated AI standardisation committee, JTC 1/SC 42, has developed standards covering areas ranging from AI management systems and risk management to data quality, explainability, impact assessment and environmental sustainability. FORSEE’s analysis found that, despite this broad scope, they primarily frame AI success at the organisational level: organisations are expected to identify risks, document processes and put governance mechanisms in place. This approach is strong on how organisations should manage risks, but deliberately weaker on a prior question: which risks should society accept in the first place, and in whose interests? Environmental sustainability appears in the standards landscape, but primarily in a non-normative Technical Report, reflecting an emerging consensus rather than requirements against which organisations can claim compliance.
This matters because the organisations asked to manage AI’s impacts often do not control the infrastructure that produces them. US companies dominate advanced AI chip design and much of the cloud infrastructure European AI systems use. When the EU delegates environmental compliance to individual organisations, it might delegate responsibility without giving them the power to act on it.
Professional associations complicate the picture
What does trustworthy AI look like in a courtroom, or what does it mean for a doctor to remain responsible for a diagnosis an AI system helped generate?
Professional associations take a more conditional view of AI success: AI succeeds only if it supports the profession’s core responsibilities and values. The Council of Bars and Law Societies of Europe specifies that lawyers must understand how AI tools function, communicate transparently with clients, and exercise professional judgement regardless of AI output. In healthcare, patient outcomes, doctor autonomy, and the patient-doctor relationship serve as the primary anchors for what counts as success, not accuracy or efficiency alone.
By repeatedly foregrounding human oversight, judgement, and institutional accountability, professional associations push back against the idea that AI success can be reduced to technical performance and formal risk controls.
Engineering associations are the partial exception. Their documents show greater internal diversity than in law or healthcare: some emphasise certification, explainability, risk, responsibility, bias, and transparency. Others frame AI success primarily in terms of availability, efficiency, and deployment readiness, treating adoption itself as a measure of progress.
The ethical governance themes, prevalent in legal and healthcare documents, appear in some engineering texts, but they don’t take precedence over performance and innovation criteria.
Beyond risk: what should AI achieve?
FORSEE’s analysis found broad agreement on risk management, transparency, and accountability across institutions. There is far less agreement on the social, democratic and ecological outcomes AI should produce. Risk management, transparency, accountability, and fundamental rights protections are shared, but environmental impact, democratic governance, and socio-ecological outcomes aren’t, at least not to the same extent.
Should the EU take a stronger approach to what successful AI outcomes should look like in the first place? FORSEE’s analysis on social media discourse on AI shows that concerns about concentrated power and working conditions are already part of public discussion around AI, and experts repeatedly call attention to the environmental and democratic effects of the technology.
Defining AI success primarily through risk management leaves a more fundamental question unanswered: success towards what?
For detailed analysis and methodological notes, read the full FORSEE reports:
AI Success criteria by national standards and ISO regulatory bodies by Marta Lasek-Markey, Delaram Golpayegani, Manushresth Mahesh, Victoria Wiegand, Arjumand Younus, Yuening Li, Aphra Kerr, and Dave Lewis.
Evolution of supranational institutional success criteria in post-2018 AI guidelines by Delaram Golpayegani, Marta Lasek-Markey, Arjumand Younus, Aphra Kerr, Dave Lewis, and Alexandros Minotakis.
Success criteria set by ACM and IEEE by Victoria Wiegand, Delaram Golpayegani, Marta Lasek-Markey, Arjumand Younus, Monique Munarini, Yuening Li, Aphra Kerr, and Dave Lewis.
Success criteria by professional associations in the EU by Linnet Taylor, Merve Öner Kabadayi, Princy Marimuthu, Delaram Golpayegani, Marta Lasek-Markey, Arjumand Younus, Aphra Kerr, and Dave Lewis.
