Human control necessary to avoid discriminatory impacts

The third paragraph of Article 11 of the Law 132/2025 lays down an important principle: artificial intelligence employed in the organisation and management of work cannot generate discrimination and must be governed by effective human guarantees.

The rule expressly recalls respect for the inviolable rights of the person and prohibits differentiated treatment on the basis of sex, age, ethnic origin, religion, orientation, opinions or personal and social conditions. This gives rise to a very important employer responsibility: the employer must assess ex ante the impartiality and fairness of the systems, oversee the logic and metrics used and intervene when the outcomes show discriminatory effects, even if unintentional.

The adoption of algorithms in selection, evaluation and task allocation shifts the risk from human behaviour to decision-making models. Biases in data or calculation rules can produce distorting outcomes: systematic exclusion of certain candidates, age or maternity-related penalties, performance attributions based on spurious signals. In this perspective, human control is not a formal frill, but the operational antidote that makes it possible to read the context, correct model errors and document the traceability of decisions.

The anti-discrimination principle of paragraph 3 is part of an already robust national framework: Article 3 of the Constitution, Article 15 of the Workers' Statute, Legislative Decrees 215 and 216/2003 (race/ethnicity, religion, disability, age, orientation) and Legislative Decree 198/2006 (Equal Opportunities Code). The protections cover all stages of the relationship: access, contractual conditions, training, progressions and remuneration.

At the European level, the picture is being further consolidated. The EU directive 2023/970 on pay transparency shifts the focus to measurable outcomes (gaps of more than 5% between salaries will have to be corrected), while the EU regulation 2024/1689 (Ai Act) classifies the systems used for recruitment and personnel management as 'high risk', imposing data quality, traceability and human supervision. The basic idea is the same as in subsection 3: what counts is not the intention, but the objective effect of automated decisions.