The hand on the pen: generative AI and the Italian parliamentary experience

The integration of artificial intelligence into the legislative process is among the most significant—and least examined—constitutional developments of the decade. This article’s concern is not the composition of legislative text ex novo, but the function constitutional theory assigns to parliaments: the scrutiny of the texts and amendments tabled before them. The research question is accordingly how the deployment of AI in the amendment phase affects the capacity of parliaments to scrutinize legislation—and the accountability, transparency, and separation of powers it secures—and whether human oversight suffices to preserve it. The inquiry centres on two Italian systems at opposite ends of a governance spectrum. GEM (Gestore EMendamenti), the Senate’s amendment-management ecosystem, operational since 2016, ranks among the most sophisticated parliamentary AI tools in use; its functions concern the processing, clustering, ordering, and prediction of amendments, not autonomous drafting—a narrow but constitutionally consequential scope. GENAI4LEX-B, a hybrid architecture combining symbolic reasoning with generative models, has been selected for the Chamber of Deputies but not yet deployed; its value is not empirical but architectural, as a counter-model in which safeguards were designed ex ante rather than emerging reactively from practice. The article advances two arguments. First, it identifies four constitutional fault lines in AI-assisted amendment processing—consequential algorithmic error, deliberate disruption through AI-generated flooding, accountability opacity, and epistemic homogenization across branches—while recognizing the countervailing potential to strengthen scrutiny; it then tests human oversight against the empirical literature on automation bias, concluding that nominal oversight is necessary but not sufficient. Secondly, it contends that the European Union AI Act leaves a significant gap for the legislative process, and proposes a five-element governance framework to ensure that the distinction between algorithmic assistance and algorithmic authorship is maintained through enforceable standards rather than self-regulation alone.