Every artificial intelligence (AI) system produces its results on the basis of criteria established by human beings. This piece traces who defines those parameters, which rules and technical standards govern them, and how adherence to them can be checked.
The human responsibility behind automated decisions
An AI system does not “think” in the way a person does: it processes vast quantities of data and identifies correlations. The criteria by which it selects, weighs and delivers a result, however, derive from human choices. Someone establishes which datasets to use for training, which objectives to optimise and which errors to treat as acceptable. Even a statistical calculation thus reflects priorities and assumptions defined upstream.
Linked to this is the issue of bias, or systematic distortion. When the historical records used in training reflect inequalities or prejudices, the model tends to reproduce them, sometimes amplifying them, with an appearance of objectivity that makes them harder to recognise. A tool that screens CVs, assesses creditworthiness or supports a diagnosis is no more impartial than the information and the choices that shaped it. Responsibility for automated outcomes ultimately remains human and identifiable.
The European rules: transparency and human oversight
For years, AI ethics remained a set of general principles, such as fairness, transparency and accountability. With the European regulation on artificial intelligence (AI Act), Regulation (EU) 2024/1689 of 13 June 2024, the European Union gave those principles legal form. It is the first comprehensive legislative framework dedicated to AI, and it adopts a risk-based approach: the greater a system’s potential impact on people’s rights and safety, the more stringent the obligations.
Two provisions bear directly on this subject. Transparency (Article 13) requires high-risk systems to function in a sufficiently intelligible way, so that those who use them can interpret their results. Human oversight (Article 14) requires that such applications be designed to be genuinely supervised by natural persons, capable of grasping their limitations and anomalies and, where necessary, of stepping in.
The technical standards: from principles to verifiable practices
Rules define the goals; technical standards show how to reach them. Drawn up by international bodies such as the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC), they are not legal norms but shared references that turn abstract principles into verifiable procedures, frequently used by organisations to demonstrate conformity with what the legislation demands.
The ISO/IEC 42001:2023 standard sets out the requirements for an artificial intelligence management system: allocation of roles, documentation of choices, and monitoring of risks throughout the system’s life cycle. ISO/IEC 42005:2025, by contrast, addresses impact assessment, namely the method for analysing how an AI system may affect individuals, groups and society. Both contribute to explainability: documenting the underlying data, criteria and responsibilities is what makes an automated choice reconstructible and, where necessary, open to challenge.
Outlook
The AI Act entered into force in 2024, and its provisions apply on a staggered basis over the following years, with obligations becoming progressively operational for high-risk systems. On the technical side, the standards supporting the regulation are still largely under development within European and international bodies. What remains to be defined, in particular, are the harmonised references that will connect the legal requirements to technical specifications, and the ways in which transparency, human oversight and risk management will be verified in practice.
In dialogue with the 2030 Agenda
- Goal 16: Peace, justice and strong institutions. Clear rules, transparency and human oversight make automated choices reconstructible and open to challenge, reinforcing accountability.
- Goal 9: Industry, innovation and infrastructure. Shared technical standards turn ethical principles into verifiable practices, supporting responsible innovation in AI.
