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

Abstract

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.

1. Artificial Intelligence and the legislative process: from efficiency tool to constitutional actor

The relationship between technology and law has always been one of productive tension. Each successive wave of technological innovation has forced legal systems to adapt—not merely by regulating the new phenomenon, but by reconfiguring the very instruments through which regulation is produced. The printing press democratized access to legal texts; the telegraph accelerated the transnational harmonization of commercial law; the computer transformed the management of statutory databases and the codification of regulatory information. Artificial intelligence, however, presents a qualitatively different challenge. Unlike its predecessors, AI does not merely alter the medium through which law is communicated or stored: it intervenes in the intellectual process through which law is conceived, refined, and scrutinized. This is not a question of efficiency alone. It is, at its core, a constitutional question.

The deployment of AI systems in the legislative process—hereafter referred to, for the reasons explained below, as AI-assisted legislative drafting or AALD—is no longer a speculative prospect confined to academic literature or science-fiction governance scenarios. It is an institutional reality. Parliaments and legislatures across the world are integrating machine learning tools, large language models (LLMs), and symbolic AI systems into their working procedures with a speed that has outpaced both scholarly reflection and regulatory response. The most recent comparative survey confirms the scale of the phenomenon: of 115 parliaments and chambers consulted across 86 countries, more than two-thirds have already adopted multi-year digital strategies, 73 per cent operate formal modernization programmes, and 29 per cent were already using AI in some form by 2024, against a markedly smaller share in 2020.¹

This trajectory is not without a rationale, and the rationale is not without merit. The phenomenon of legislative inflation—the exponential growth in the quantity of statutory and regulatory output—has been extensively documented in public law scholarship. Scholars such as Xanthaki and Mousmouti have long argued that the crisis of modern legislation is not merely quantitative but qualitative: the proliferation of norms produces incoherence, uncertainty, and ultimately a weakening of the rule of law itself.

Xanthaki’s methodology emphasizes that effective legislation requires genuine engagement with the underlying policy problem—not merely technical compliance with drafting conventions.² It follows that any delegation of the drafting function to an actor—human or artificial—that lacks the contextual understanding and democratic accountability inherent in the legislative process risks compromising the very quality the drafting enterprise is designed to achieve. Mousmouti’s concept of the ‘effectiveness test’ in legislative design reinforces this point, insisting that the quality of legislation must be measured by its capacity to achieve its regulatory objectives.³

If AI tools can genuinely assist drafters and parliamentary staff in navigating this complexity—by identifying redundancies, flagging constitutional inconsistencies, or summarizing voluminous amendment packages—their contribution to legislative quality is a legitimate and serious claim that deserves careful scholarly attention.

Yet, legislating is not a neutral, technocratic exercise. As Cohen argued in his foundational work,⁴ it is an act of translation—the translation of political will into legal form—that necessarily involves choices of emphasis, inclusion and exclusion, linguistic framing and normative prioritization. These choices are political in the deepest sense: they reflect power relations and shape the space of future judicial interpretation. To introduce AI into this process is not simply to automate a task: it is to alter the locus of agency in a process that democratic theory has always assigned to elected representatives and their expert human staff.

It is precisely here that the angle of this article must be stated with precision, because the constitutional stakes of AI in the legislative process are frequently obscured by the assumption that the relevant function is drafting. In strict constitutional terms, parliaments do not, as a rule, draft: they scrutinize. The government, or an individual member, proposes; the chambers examine, amend, and either consent to or reject. This is emphatically not to diminish drafting proper—the expert composition of legislative text, which a substantial body of scholarship has shown to be a demanding craft of judgment rather than a mechanical exercise, and which lies largely with specialist drafters and the executive rather than with the chambers themselves. It is rather to locate the constitutionally distinctive activity of the legislature in the critical examination of the texts placed before it—and, in a system such as Italy’s, that examination is conducted overwhelmingly through the medium of amendments.

This article accordingly takes parliamentary scrutiny, rather than autonomous text generation, as its unit of analysis, and asks a single organizing question: how does the introduction of AI into the amendment phase affect the legislature’s capacity to scrutinize legislation, and the accountability, transparency, and inter-institutional balance that such scrutiny is meant to protect? The working hypothesis is that AI tools simultaneously enhance the capacity for scrutiny—by rendering tractable volumes of amendments that would otherwise overwhelm human processing—and place new pressure on its quality, by interposing an opaque cognitive layer between the parliamentarian and the text. Whether the safeguard on which Italy has settled—human oversight—is adequate to reconcile the two is the question that animates the analysis that follows.

A terminological clarification is in order. This article distinguishes three categories of AI involvement in the legislative process that are often conflated in public discourse. The first is AI-assisted processing and drafting, in which AI tools support human actors by performing tasks such as summarization, clustering, consistency checking, comparative analysis, and the ordering of amendments, while final authorship and responsibility remain unambiguously human. The second is AI-generated drafting, in which an AI system produces a first or full draft of a legislative text that human actors then review and modify. The third, and most constitutionally fraught, is what might be called AI-delegated governance, in which AI systems are accorded decision-making authority without a clear or enforceable chain of human accountability. Both Italian systems examined here fall, by design, within the first category; the analytical interest lies in whether they can be kept there.

A second clarification concerns the scope of the systems under examination, and it must be made at the outset because it conditions everything that follows. Neither GEM nor GENAI4LEX-B drafts legislation in the sense of composing original normative text autonomously. Their functions are concentrated in the amendment phase: the classification, clustering, ordering, summarization, and predictive projection of amendments. A reasonable objection—one that must be confronted directly rather than assumed away—is that processing, organizing, and classifying amendments is not the same as drafting them, and that systems performing the former should not be presented as instances of AI-assisted drafting at all.

The objection is well taken as a matter of vocabulary, and this article does not elide the distinction; on the contrary, it treats the narrowness of the systems’ remit as analytically central. The reason the amendment phase nonetheless lies at the heart of the legislative enterprise in the Italian context is structural rather than rhetorical. Under the Italian Constitution, the legislative function is exercised collectively by the two chambers⁵, and that collective exercise is given effect precisely through the power to amend the text under discussion. The final text of a statute is not the base text proposed: it is the product of the interaction between that base text and the amendments adopted during its parliamentary passage. Systems that cluster, order, summarizes, or predict the effect of amendments therefore intervene directly in the production of primary legislation, not at its periphery. The scope is narrow; its constitutional consequences are not. This article uses the term ‘legislative drafting’ in that extended, amendment-inclusive sense throughout, and signals the distinction wherever the difference between text generation and amendment processing bears on the argument.

Italy is a particularly instructive case: it is among the first EU Member States to have moved toward institutional deployment of generative AI in parliamentary workflows; its civil law system and high legislative volume make the pressures of legislative inflation especially acute; and its constitutional framework—with its emphasis on parliamentary primacy and symmetrical bicameralism—makes the implications for scrutiny especially vivid. Italy operates, moreover, within the regulatory perimeter of the EU AI Act (Regulation (EU) 2024/1689), whose provisions on high-risk AI systems are directly relevant to the deployment under consideration.

The article proceeds as follows. Section 2 examines the Italian case study, analysing the Senate’s GEM ecosystem as the empirically mature core of the inquiry, the GENAI4LEX-B project at the Chamber of Deputies as an ex ante governance counter-model, and the statutory framework established by Italy’s Law No. 132/2025, with particular attention to the ‘Human in Command’ principle. Section 3 identifies the constitutional challenges and opportunities that AI-assisted amendment processing presents. Section 4 subjects the prevailing solution—human oversight—to critical scrutiny against the empirical literature, and asks directly whether it has altered parliamentary practice, the capacity of members to scrutinize legislation, and the quality of that scrutiny. Section 5 draws these threads together to propose a five-element regulatory framework, and concludes by assessing the Italian experience as both a model and a warning.

2. The Italian case study: GEM, GENAI4LEX-B, and the architecture of AI-assisted amendment processing

To understand the significance of Italy’s experiment, it is necessary first to appreciate the institutional environment in which it has been conducted. Italy’s Parliament is a bicameral legislature of unusual symmetry: under the architecture established by the 1948 Constitution,⁶ and preserved in its essential structure following the 2020 reform that reduced the number of parliamentarians, both the Chamber of Deputies and the Senate of the Republic exercise full legislative functions, with each bill requiring approval in identical text by both chambers before it can become law.

This system of equal bicameralism—increasingly regarded as an anomaly in comparative terms, and itself the object of a failed reform referendum in 2016⁷—generates a volume of inter-chamber legislative traffic structurally higher than in unicameral or functionally asymmetric bicameral systems.

The scale of that processing challenge is quantified in a document of exceptional institutional candour: Documento di analisi n 34, published by the Senate’s Office for Impact Assessment (Ufficio Valutazione Impatto, UVI) in 2025, and titled, in its English version, Artificial Intelligence in the Senate: Have New Information Technologies Improved the Legislative Process, and How?⁸ The document is the single most important source for any empirically grounded account of AI in the Italian Parliament, and it warrants close engagement rather than selective citation.

Three orders of quantitative evidence emerge from it. The first concerns the sheer longitudinal scale of the documentary burden. Between 1996 and 2025, the Senate held more than 58,000 sittings of the Assembly and of committees, generating close to 2.3 million pages of stenographic record.⁹ The second concerns the legislative throughput across the period for which the document tabulates data. In the 24 years between 30 May 2001 (XIV legislature) and 30 June 2025 (XIX legislature), 41,120 bills were introduced in the Senate; of these, 2103 were approved—approximately 5 per cent—and in the course of their passage 713,263 proposed amendments were voted upon, an average of roughly 150,000 amendments processed by the Senate’s offices in every legislature, from submission to final vote.¹⁰ That figure–close to three-quarters of a million amendments in a single chamber over two and a half decades—conveys, more vividly than any abstract description, the magnitude of the task that AI tools are called upon to assist with. The third order of evidence—the qualitative impact of that assistance on the scrutiny function—is, as Section 4 explains, conspicuously absent from the document and from the literature more generally; that absence is itself a finding.

Within this longitudinal picture sits an episode that the document records with equal frankness, and which amounts to what is likely a world record in the history of legislative obstructionism. During the debate on the constitutional reform bill in the XVII legislature, on a single day—23 September 2015—82,730,460 amendments generated by an algorithm were tabled in the Senate.¹¹ The near-totality was subsequently declared inadmissible by the President of the Senate on 29 September, on the ground that it was objectively impossible to examine the merits of so abnormal a number of amendments without setting a precedent capable of blocking parliamentary business for an incalculable period.¹² The episode is not a curiosity. As a 2022 comparative account observes, in the Italian Senate amendments are tabled electronically but not always to alter proposed legislation: members, particularly from opposition parties, may table very large numbers of near-identical amendments, differing by a few words, in order to slow the legislative process—a practice known as filibustering, which poses acute challenges for the staff responsible for analysing, organizing, and scheduling them.¹³

The weaponization of AI-generated content against parliamentary procedure was thus not a future risk but an already-realized event, and the institutional response–the development of tools capable of processing and sorting such volumes—was itself a reactive adaptation to algorithmic attack. Thanks to GEM, then in its experimental phase, the Senate’s offices were able to process and sort all the amendments in three days. It is in this context—of massive legislative volume, structural bicameral pressure, and the lived experience of algorithmic disruption—that the Italian Parliament’s embrace of AI tools must be situated. Adoption was not, in Italy, a technocratic fashion imported from abroad: it was an institutional survival response to pressures building for decades.

2.1 The Senate’s GEM Ecosystem: from amendment management to predictive projection

Because the empirical centre of gravity of the Italian experience lies in the Senate rather than the Chamber, the analysis begins there. The internally developed GEM¹⁴ provides the most detailed and operationally mature example of AI integration into the core of the legislative process. It has been fully operational since 2016 and is acknowledged at the international parliamentary level as among the most advanced suites available for the management of parliamentary amendments.¹⁵ Its significance is twofold: it demonstrates that AI-assisted amendment processing is not a prospective development but an institutional reality with nearly a decade of operational history; and it provides whatever empirical foundation exists—in documented results and processing volumes—that any serious governance framework must engage with.

The technical architecture of GEM reflects the specific characteristics of the Italian process. All bills submitted to the Senate must adhere to a strict hierarchical structure, ensuring the correct organization of articles, paragraphs, and sub-paragraphs, the consistent application of article headnotes (‘rubrics’), and the precise citation of normative references; following submission, bills undergo technical review by the Drafting Service. Structuring legislative texts in machine-readable formats—enabling algorithms to identify whether a segment is an article, a clause, or another partition–is fundamental to the deployment of the tools GEM comprises. The mark-up is carried out both manually and automatically via PArSe, a purpose-built natural-language-processing parser for the text of parliamentary bills, capable of processing drafts before they are fully structured in XML.¹⁶ Once marked up, these texts are processed and published through Senate channels, including an open repository that provides, for every bill, a downloadable machine-readable XML version enabling automated processing by external systems.

GEM employs a suite of AI technologies that has expanded progressively since its initial deployment: clustering algorithms for the identification of similar or identical amendments, natural-language-processing parsers, generative AI components, decision trees, and QR-code generation for tracking amendments through the process.¹⁷ The clustering functionality is of particular practical importance: by automatically identifying amendments that share substantive content despite differing in wording, GEM enables staff to group related proposals, detect redundancies, and flag amendments whose apparent diversity masks substantive identity—a capability decisive, as noted below, in the response to algorithmically generated flooding.

The ordering of amendments is one of the most technically demanding and constitutionally sensitive functions GEM performs. Under the Senate’s Rules of Procedure, amendments must be arranged according to the provisions of the base text they seek to modify and, within each provision, ordered by effect—suppressive amendments first, then those most radically departing from the original, proceeding towards the least significant modifications.¹⁸ This ordering is not a mere administrative convenience: it determines the sequence in which amendments are debated and voted, and so directly shapes the logical structure of parliamentary deliberation. An incorrect ordering could alter the outcome of the process by causing amendments to be precluded or absorbed in an order inconsistent with the rules. GEM’s automated ordering algorithm produces a provisional ordering that is then reviewed and, where necessary, corrected by staff—a workflow illustrating the practical operationalization of human oversight in a context where its absence could have immediate procedural consequences.¹⁹

The system’s capabilities have been tested under conditions of extraordinary institutional stress. In the episode of 23 September 2015, over 82 million algorithmically generated amendments were submitted by opposition senators who had commissioned a natural-language-generation programme to produce textual variants of a base amendment by automatically substituting individual words and punctuation marks. GEM’s clustering and similarity-detection tools were instrumental in enabling the Senate’s offices to process and classify the submissions within three working days—a task physically impossible without algorithmic assistance.²⁰ A comparable, smaller scale episode occurred during the discussion of the civil-unions bill, when approximately 4500 amendments with near-identical formulations were submitted by analogous methods.²¹ The phenomenon has not remained confined to the national Parliament: during the budget session of the Sicilian Regional Assembly in December 2023, an opposition group submitted a substantial number of AI-generated amendments that, while appearing to propose substantive modifications, in fact differed only by the substitution of synonyms or minor locutions, and were identified by the staff and declared inadmissible by the committee chair.²² The replication of the phenomenon at the sub-national level confirms that the risk of AI-generated procedural disruption is systemic rather than episodic, extending to every tier of the process in which electronic submission is permitted.

Beyond this defensive role, GEM has been deployed in a more innovative capacity of considerable constitutional significance. The UVI document describes a ‘final-text predictor’—a system employing generative AI to predict the full legislative text that would result from the approval of any given combination of amendments. The predictor operates on the basis of the initial text and the amendments tabled, producing a parallel text that highlights the differences between the initial and projected final versions, with deleted provisions in red, added provisions in green, and the relevant amendment in yellow.²³ The constitutional significance should not be understated: it transforms the process from one in which the full normative implications of amendment approval can be assessed only after the fact—often by lawyers and courts interpreting the resulting text in the light of possibly incomplete legislative materials—to one in which those implications can be visualized and assessed by legislators before the vote.

If this capability can be made genuinely reliable, it represents a substantive enhancement of the quality of democratic deliberation, addressing directly the concern expressed by scholars of legislative quality from Xanthaki to Mousmouti that modern processes too often produce texts whose consequences are not fully understood by the legislators who approve them. As Conzutti has argued, such tools have the potential to rebalance the structural asymmetry between legislature and executive: by equipping parliamentarians with real-time analytical capabilities previously available only to the government’s own legal services, AI-assisted amendment analysis could strengthen the legislature’s capacity to scrutinize and shape executive-initiated legislation rather than merely ratifying it.²⁴ The question of reliability—and of what happens when the predictive model produces an erroneous projection on which legislators then rely in casting their votes—is, as Sections 3 and 4 argue, among the central governance challenges the Italian experience has yet to resolve.

2.2 The GENAI4LEX-B project: a counter-model of ex ante governance

If the Senate’s GEM exemplifies governance that has emerged reactively from operational practice, the Chamber of Deputies’ GENAI4LEX-B project exemplifies the opposite: governance designed before deployment. It is essential to be explicit, at the outset, about what this second case study does and does not contribute. It contributes no operational data, for the decisive reason that it is not yet operational; it remains at the proof-of-concept stage and has not been integrated into the Chamber’s legislative workflow. To present it as a source of empirical evidence about the impact of AI on scrutiny would therefore be a category error. Its value to the present analysis is of a different and complementary order: it is the clearest available instance of a system in which constitutional and transparency safeguards were embedded ex ante, as architectural prerequisites, rather than retrofitted after operational problems emerged. The contrast it draws with the Senate is accordingly not the trivial one between an operational and a non-operational system, but the analytically useful one between two distinct models of how the safeguards on which scrutiny depends come into being. That contrast is the foundation on which the regulatory framework proposed in Section 5 is built; without it, the framework would lack any observed referent for the proposition that ex ante design is achievable.

The project’s genesis is itself significant. In 2024, the Chamber launched a competitive call for proposals,²⁵ receiving 28 submissions from 15 universities and research centres, from which 3 winning projects were selected by a technical commission chaired by Professor Luciano Floridi.²⁶ GENAI4LEX-B explicitly builds on years of research in legal informatics, applying principles of interdisciplinarity, transparency, fundamental-rights compliance, ethical analysis, hybrid architecture, and attention to end-user needs. It is not a system developed rapidly in response to the generative-AI wave of 2022–2023: it is the latest instantiation of a research programme that has been developing the foundations of machine-readable legal documents for more than two decades.²⁷

The most distinctive technical feature—and the one most relevant to the governance argument—is its hybrid architecture, combining two fundamentally different approaches in a principled manner. The first component is symbolic AI, grounded in formal logic and rule-based reasoning. Unlike generative models, symbolic systems do not learn from statistical patterns: they apply handcrafted rules and structured knowledge representations to perform logical inferences. This makes them highly reliable within their encoded domain, but brittle outside it.²⁸ In GENAI4LEX-B, the symbolic component encodes legal knowledge–statutes, regulations, constitutional provisions, and Constitutional Court decisions—into structured, machine-readable formats according to the principle sometimes described as ‘Law as Code’, and is responsible for compliance verification: identifying conflicts with existing legislation, verifying constitutional compliance, and navigating the hierarchy of sources in accordance with the lex specialis and lex posterior principles.

The second component is generative AI. Generative systems—of which LLMs are the prominent example—do not apply predetermined rules; they generate outputs by predicting, from probabilistic patterns learned from vast training data, what text is most likely to follow a given input. This makes them flexible and capable of human-quality prose but structurally susceptible to ‘hallucination’: the generation of plausible but factually incorrect outputs, a risk whose constitutional implications are examined in Section 3.1.²⁹

The generative component handles tasks symbolic AI cannot: summarizing committee amendments, generating explanatory notes, drafting parallel texts illustrating the impact of proposed amendments on the base text, and rendering complex provisions in language accessible to non-specialists. A critical further element is the reliance on the Akoma Ntoso/LegalDocML OASIS standard³⁰ for the structured representation of legal documents, ensuring interoperability across national and European sources and permitting intelligent graphical visualization of normative connections across versions of the law. The choice of an open standard—rather than a proprietary format that would create institutional lock-in and reduce the capacity of external actors to verify outputs—is itself a design choice with democratic implications.

At the time of writing, GENAI4LEX-B remains at a proof-of-concept stage, its codebase, documentation, and validation results publicly accessible through a dedicated open-source repository.³¹ The repository describes three principal tools: a classification instrument (Istruire) identifying related legislation and Constitutional Court decisions; a summarization tool (Comprendere) for committee amendments; and a policy-impact measurement tool (Misurare) assessing the impact of a given policy dimension within a proposal. The system has not yet been integrated into the Chamber’s operational workflow. Its status is thus one of research and institutional collaboration rather than deployment—a circumstance that limits the empirical conclusions that can be drawn from it, but that does not diminish the analytical value of its architectural and governance choices, made deliberately and in advance. That constitutional and transparency safeguards have been embedded from inception, rather than retrofitted, is precisely the feature that makes the project a governance model worth attention, and the necessary counterpart to the Senate’s operationally tested but reactively assembled arrangements.

2.3 The ‘Human in Command’ principle and the Italian AI Act

The governance philosophy uniting both systems–despite their different origins, maturity, and architectures—is the principle the UVI document terms ‘Human in Command’. The document articulates it with notable candour: no act, document, or datum generated by or with AI—whether an amendment, the summary of a text, or any other parliamentary output—has been processed in the Senate without careful human supervision, and humans, not machines, must retain authority, oversight, and ultimate responsibility for all decisions. At the core of the administration’s strategy lies the conviction that, however intelligent or autonomous a system may be, technology remains a tool while the human remains the decision-maker who bears full moral and legal responsibility.³² This is not, in the Italian context, mere aspiration: it is operationalized in concrete requirements. The provisional ordering produced by GEM’s algorithms is systematically reviewed by staff before becoming definitive; the outputs of the final-text predictor are treated as projections to be verified, not as authoritative statements of the law; and GENAI4LEX-B has been designed, from inception, with explainability and human validation as architectural prerequisites.

As Italian scholars have observed, this insistence on human oversight is not merely precautionary but a constitutional necessity: AI systems cannot assess the political sensitivity or underlying purpose of an individual amendment, and the absence of human control over the algorithmically produced ordering could alter the logical sequence of deliberation, potentially compromising the right of individual parliamentarians to have their proposals examined on the merits.³³ This philosophy now has a statutory foundation. Law No 132 of 23 September 2025—the first national AI framework in Europe enacted as primary legislation in coherence with the EU AI Act—establishes that AI systems must be developed and applied so as to ensure human oversight and the possibility of intervention, without prejudicing the democratic conduct of political and institutional life.³⁴ The convergence between the Senate’s operationally tested philosophy and the formal requirements of national legislation creates a framework in which human oversight is not dependent on institutional goodwill alone but reinforced by enforceable obligations.

The legislative history of Law 132/2025 is itself instructive, serving as an unintended illustration of the very challenges the Italian parliamentary AI ecosystem was designed to address. The bill remained blocked in the Senate for a prolonged period owing to the submission of over 400 amendments and to criticisms raised by the European Commission regarding inconsistencies with the AI Act, only partially resolved before final passage.³⁵ That a legislature simultaneously deploying sophisticated AI tools in its own amendment management should have been brought to near-standstill by a flood of amendments to its own AI legislation is a paradox deserving more than anecdotal attention. It reveals, in microcosm, the structural tension between the efficiency gains AI can deliver in processing amendments and the political dynamics that generate the volumes those tools are called upon to manage. Even where parliamentary AI capabilities are among the most advanced in Europe, the capacity to process amendments is a necessary but not sufficient condition for legislative progress: the political will to reach agreement on substance—including on the AI tools the legislature itself employs—remains irreducibly human.

What the Italian case demonstrates, at this stage of the analysis, is that it is possible to develop and deploy AI tools in the legislative process in ways that are technically sophisticated, institutionally honest about the tools’ limitations, and grounded in a principled human-oversight philosophy—and that doing so requires not merely technical investment but sustained interdisciplinary collaboration, open standards, a legislative culture prepared to confront the constitutional implications of the tools it adopts, and a statutory framework transforming institutional commitments into enforceable requirements. Whether these conditions can be maintained as the tools become more powerful and their outputs more consequential—and whether human oversight can survive the institutional pressures, commercial incentives, and political demands inevitably brought to bear against it—is the question to which the remainder of the article turns.

3. Constitutional challenges and opportunities: AI-assisted amendment processing between democratic risk and institutional promise

The preceding analysis has shown that AI-assisted amendment processing is not a theoretical prospect but an institutional reality whose governance implications demand immediate attention. This section examines those implications through a dual lens: the constitutional risks that the technology introduces into the scrutiny function, and the institutional opportunities it creates for the improvement of legislative quality and the strengthening of parliamentary capacity. Neither dimension can be assessed in isolation: the risks are real and constitutionally serious, but so are the costs of the technology’s absence in an environment characterized by chronic information asymmetry, processing bottlenecks, and declining parliamentary influence over the production of primary legislation.

3.1 Algorithmic hallucination in the legislative context

The tendency of LLMs to generate plausible but factually incorrect outputs—hallucination—has attracted attention in legal practice, most notably following the sanctions in the Southern District of New York’s Mata v Avianca proceedings, where counsel submitted filings citing fabricated case law generated by ChatGPT.³⁶ The principle established there is sound: responsibility for the accuracy of legal submissions never transfers to the machine. But its application to legislative drafting presents difficulties of a qualitatively different order. Erroneous filings operate within a system designed to detect and remedy inaccuracies through adversarial correction. Legislative texts, in contrast, are presumed valid once enacted: they bind courts and citizens alike, and their errors—of cross-reference, definitional inconsistency, or inadvertent constitutional incompatibility—may remain undetected for years, generating litigation costs borne by individuals who had no involvement in the defective process.

The empirical literature reinforces the gravity of the concern. Research by Stanford’s RegLab and HAI found that general-purpose models hallucinated on legal queries at rates between 58 and 82 per cent, while even purpose-built legal AI tools employing retrieval-augmented generation still produced incorrect outputs in more than 17 per cent of benchmark queries.³⁷ These figures relate to systems designed with accuracy as a primary objective and operated by trained professionals with strong incentives to verify results. The hallucination rates to be expected in a parliamentary drafting environment—where time pressure is intense, amendment volumes may number in the hundreds of thousands, and institutional incentives for thorough verification may be weaker—have not been empirically investigated. The hybrid architecture of GENAI4LEX-B offers partial mitigation: its symbolic layer is not susceptible to hallucination in the characteristic LLM sense, since its outputs are logical inferences from an encoded corpus. But its generative component retains the structural susceptibility that researchers argue is an inevitable consequence of probabilistic language modelling rather than a correctable anomaly.³⁸ The Italian literature identifies a complementary strategy: ensuring that parliamentary AI systems are fed exclusively from certified documentary databases—parliamentary archives, official legislative databases such as Normattiva, and verified statistical sources—rather than general-purpose search engines whose reliability is compromised by fragmentary, obsolete, or false information.³⁹ GENAI4LEX-B’s exclusive reliance on parliamentary sources and its use of the Akoma Ntoso standard reflect precisely this approach.

3.2 AI as an instrument of institutional disruption—and of institutional defence

The episodes of algorithmically generated flooding documented above—at national and sub-national levels—raise constitutional questions extending beyond obstruction. They reveal a structural vulnerability in any system permitting electronic submission of amendments: the capacity to distinguish genuine proposals from procedural noise becomes a function of institutional AI capability.⁴⁰ Legislatures lacking such capability are exposed to a form of disruption requiring no technical sophistication on the part of the disruptor, only access to commercially available generative tools.⁴¹ Yet, the Italian experience shows the relationship between AI as an instrument of disruption and AI as an instrument of defence to be not simply adversarial but dialectical: the same capabilities that enable flooding also enable its detection and management. Any governance framework must account for both. The Chamber has recently introduced MSE, an AI-assisted writing system for amendments that provides real-time information and syntactic guidance to deputies during drafting.⁴² Such tools may, over time, serve a dual function: improving the quality of legitimately drafted amendments while providing institutional infrastructure for detecting algorithmically generated noise.

3.3 Opacity, accountability, and the promise of transparency

The opacity of AI decision-making poses a specific challenge in the legislative context, where constitutional democracy requires that the production of law be transparent and attributable.⁴³ When an AI system produces a summary of committee amendments on which members rely in deciding how to vote, or generates a predictive final text that shapes legislators’ understanding of the consequences of their decisions, the quality and reliability of that output become matters of constitutional significance. If the system’s reasoning is opaque—if neither the legislators who rely on its outputs nor the citizens bound by the resulting legislation can understand how and why the AI reached its conclusions—then a core requirement of democratic accountability is compromised.

The Italian approach offers elements of a response, though not a complete one. The commitment to open standards–particularly Akoma Ntoso—ensures that the data on which the systems operate are publicly accessible and independently verifiable. The development of GENAI4LEX-B as an open-source project is a further transparency guarantee: external researchers, civil-society organizations and constitutional courts can, in principle, examine its architecture, training data, and decision logic. The Italian literature emphasizes that transparency and explainability, while related, are not interchangeable: transparency makes the functioning of a system knowable, while explainability translates that functioning into terms comprehensible to non-specialists.⁴⁴ A system may be transparent without being explainable, and vice versa. The governance challenge is to ensure both, and to develop the ‘explainable AI’ (XAI) techniques capable of rendering intelligible the variables that determined an algorithmic output, even when produced by deep-learning models whose internal reasoning resists straightforward interpretation⁴⁵. As Section 4 shows, however, the relationship between explainability and effective oversight is more vexed than the transparency literature sometimes assumes.

3.4 The separation of powers and the risk of epistemic homogenization

A risk that has received insufficient attention in the comparative literature concerns the separation-of-powers implications of AI deployment across the executive and legislative branches. When the same tools, training data, or analytical frameworks serve both the government that proposes legislation and the parliament that scrutinizes and amends it, the conditions for what may be termed ‘epistemic homogenisation’ are structurally in place: the legislature’s capacity for independent critical assessment of executive proposals is compromised not by overt political pressure but by the shared cognitive infrastructure through which both institutions process legal information. The GENAI4LEX-B documentation explicitly envisages potential collaboration with the European Commission’s Directorate-General for Informatics, pointing toward broader European integration of the tools developed at the Chamber. If the same infrastructure were to serve both the Commission and the European Parliament, the conditions for epistemic homogenization at the supranational level would be realized—a scenario whose constitutional implications have not yet been examined in the European public-law literature.

Yet, this risk must be weighed against the countervailing opportunity AI presents for the rebalancing of institutional power. As Conzutti argues, the progressive marginalization of parliament vis-à-vis the executive—the ‘executive dominance’ problem—is driven in significant part by the structural information asymmetry between the branches: the executive commands the resources of the ministerial bureaucracy, while parliamentary staff operate with vastly inferior analytical capacity.⁴⁶ Tools equipping parliamentarians with autonomous research, real-time analysis, and predictive modelling could attenuate this asymmetry, enabling the legislature to engage with executive-initiated legislation on a more equal epistemic footing. The challenge is therefore not simply to prevent the risks of shared infrastructure but to ensure that AI deployment is differentiated across the branches in a manner that preserves—and ideally strengthens—the legislature’s capacity for independent scrutiny.⁴⁷

3.5 The AI Act’s regulatory gap: legislative drafting as an unaddressed use case

The EU AI Act designates as high-risk those AI systems used in the administration of justice and democratic processes, a category expressly covering systems used for interpreting and researching facts and law, or for influencing elections.⁴⁸ On a purposive reading, this appears broad enough to encompass systems used in legislative drafting—systems assisting in identifying the legal framework applicable to a bill, verifying its constitutional compliance, or predicting the impact of amendments on existing legislation. The classification nonetheless remains contested. Article 6(3) provides that an Annex III system shall not be considered high-risk where it performs a ‘narrow procedural task’ or improves the result of a ‘previously completed human activity’. GENAI4LEX-B and GEM have been designed as assistive tools supporting rather than replacing human judgment; their deployers could therefore plausibly invoke the Article 6(3) derogation.

That argument, while legally available, should not be accepted uncritically. A system that produces AI-generated summaries of hundreds of amendments on which parliamentarians rely in deciding how to vote, or that predicts the full legislative text resulting from a combination of amendments, is not performing a ‘preparatory task’ in any meaningful sense: it is performing a function with direct and significant consequences for the content of legislative decisions. The Article 6(5) guidelines—whose consultation closed in July 2025 and which the Commission was required to publish by 2 February 2026—represent the most immediate regulatory opportunity to address this gap. Their publication should explicitly designate AI-assisted drafting systems whose outputs materially influence legislative decisions as high-risk, while clarifying the conditions under which the ‘preparatory task’ derogation does and does not apply. A framework specific enough to classify AI-assisted judicial decision-making as high-risk, but whose application to AI-assisted legislative drafting remains contested, has left the most constitutionally sensitive use case inadequately governed.

4. Is human oversight enough? A critical assessment of the Italian solution

The analysis so far has documented the risks of AI-assisted amendment processing and described the safeguard on which Italy has settled: human oversight, operationalized through the ‘Human in Command’ philosophy and now anchored in statute. The case studies are routinely deployed—in this article as elsewhere—to justify the existence of the risks. They are far less often used to interrogate the proposed solution. Yet, a system operational since 2016 is precisely an opportunity for an empirically informed assessment, and the reviewer of any honest treatment is entitled to press four questions. Is human oversight, as implemented by the Senate, sufficient to ensure effectiveness, transparency, and accountability in the legislative process? Has the use of AI made a difference to parliamentary practice? Has it affected the capacity of members to scrutinize legislation? And what has been its impact, if any, on the quality of parliamentary scrutiny itself? This section addresses each in turn, and is candid about a methodological constraint that shapes every answer.

That constraint is the uneven availability of evidence. On the volume and process dimension, the data are solid and have been set out above: documented amendment throughput, the 3-day resolution of the 2015 flooding episode, the tabulated bills and amendments across seven legislatures. On the dimension that the four questions actually target—the effect of AI on the quality of scrutiny—no outcome data exist, and not only for Italy. The UVI document, for all its candour about scale and architecture, reports no measure of whether AI has improved or degraded the substantive quality of legislative examination; it records what the tools do, not what difference they make to the deliberation they support. Nor does the broader comparative literature supply such a measure for any jurisdiction. This is not a gap in the present research that further searching would close: it reflects the current state of the field. GEM is, moreover, an in-house system,⁴⁹ and the Senate publishes descriptions and aggregate results rather than the raw performance data—error rates of the clustering and ordering algorithms, accuracy of the final-text predictor, frequency of human correction of algorithmic output—that an independent evaluation of effectiveness would require. The point is developed below; here, it is enough to register that the absence of such data is not merely an inconvenience for the analyst but a finding in its own right, with direct implications for the accountability question.

Begin with the first and most important question: Is the Senate’s human oversight sufficient? The procedural architecture is, on its face, serious. The provisional ordering generated by GEM is reviewed before becoming definitive; the predictor’s outputs are treated as projections to be verified; validation is, in the Senate’s own account, invariably human. Yet, the sufficiency of nominal oversight cannot be presumed, and a substantial body of empirical research counsels against presuming it. The recurring finding is the phenomenon of automation bias: the human tendency to over-rely on the outputs of automated systems. A Microsoft synthesis of approximately 60 studies concluded that over-reliance on AI makes it difficult for users both to leverage the strengths of such systems and to oversee their weaknesses, and that users frequently accept incorrect outputs without independent verification.⁵⁰ Most unsettling for any oversight regime that leans on transparency as its safeguard is the evidence that explanation can be counter-productive: a recurrent finding in the human–computer interaction literature is that the mere presence of an explanation increases user trust and anchors the human to the AI’s prediction, so that explanations may exacerbate rather than cure over-reliance unless they are paired with ‘cognitive forcing functions’ that compel genuine engagement with the task.⁵¹ The implication is double-edged for the Italian model. The very feature presented in Section 3.3 as a democratic safeguard—the system’s capacity to explain and to display its reasoning, as in the colour-coded parallel text of the final-text predictor—may, without accompanying design that forces verification, increase the likelihood that a busy parliamentarian or official defers to the projection rather than testing it.

The legal literature on the AI Act has drawn the corresponding conclusion at the level of regulatory design. Ruschemeier shows that the Act’s treatment of automation bias places the onus on providers to enable awareness of the phenomenon, without directly regulating the design and context that produce it, and questions whether mere awareness can discharge the risk.⁵² Laux, in the most sustained treatment of the subject, observes that empirical research suggests humans are not reliable in fulfilling their oversight tasks—they may lack competence or be harmfully incentivized—and that mandating oversight by law is therefore ‘no panacea’, requiring instead careful institutional design that institutionalises distrust of the human overseer rather than assuming the overseer’s reliability.⁵³ A recent contribution frames the dilemma in its starkest form: either humans become rubber stamps, approving outputs they do not fully understand, or the system is constrained so tightly that the AI collapses into mere rule-following automation and forfeits the very capabilities that made it useful.⁵⁴ The Senate’s arrangements sit somewhere on this spectrum, and the honest answer to the first question is therefore qualified: human oversight as implemented is a necessary safeguard and is more robust than the bare statutory formula, because it is operationalized in concrete review procedures; but it is not, on the present evidence, a sufficient one, and its sufficiency cannot be asserted in the absence of the performance data that would allow the rate of effective—as opposed to nominal—human correction to be measured.

A methodological caveat is owed here, because it bears on the weight the foregoing can carry. The empirical evidence on automation bias is drawn overwhelmingly from domains other than the legislature—clinical decision-making, criminal-justice risk assessment, and administrative adjudication—and its transfer to the parliamentary setting is an argument by analogy, not a direct measurement. The transfer is defensible: the mechanisms that drive over-reliance—time pressure, high workload, long stretches of error-free operation that breed ‘learned carelessness’, and the cognitive cost of verifying a complex output—are not merely present in the amendment phase but are among its defining conditions, as the volume figures in Section 2 make plain. But the transfer is not a substitute for jurisdiction-specific study, and it would be a methodological error—indeed, the very error of relying on isolated facts in place of systematic data—to present analogical evidence as though it were evidence about the Italian Senate. The appropriate conclusion is more disciplined: the general literature establishes that the sufficiency of human oversight is contingent on design features whose presence or absence in GEM has not been independently assessed, and it is that contingency, not a demonstrated failure, that the analysis identifies.

The second question—has AI made a difference to parliamentary practice?—admits a clearer answer, though only on the process dimension. It plainly has: without GEM’s clustering and similarity-detection capabilities, the Senate could not have resolved the 2015 flooding within three days, and the Sicilian and civil-unions episodes show that the same capability is now part of the ordinary institutional repertoire for managing obstruction. To this extent, AI has not merely streamlined an existing practice but altered the balance of a procedural contest, neutralizing a form of disruption that would otherwise have been capable of paralysing the chamber. Comparative experience corroborates the trajectory: the same survey evidence cited at the outset, and parallel developments in the United Kingdom’s House of Commons, where AI is now used to process the volume of questions tabled by members, indicate a broad migration of routine processing to algorithmic assistance across mature parliaments.⁵⁵

The third question—has AI affected the capacity of members to scrutinize legislation?—must be answered with a careful distinction between capacity and its exercise. There is a credible theoretical case, advanced by Conzutti and reflected in comparative scholarship, that tools such as the final-text predictor expand capacity: by rendering visible, before the vote, the normative consequences of a given combination of amendments, they place in the hands of the individual legislator an analytical resource previously monopolized by the executive’s legal services, and to that extent they answer directly the ‘executive dominance’ problem.⁵⁶ Whether that expanded capacity is actually exercised—whether members interrogate the projection or defer to it—returns the analysis to the automation-bias problem and to the absence of usage data. Capacity and quality are not the same thing, and the distinction matters: an instrument that increases the capacity for scrutiny while encouraging deference to its outputs could expand the former while degrading the latter.

That brings the fourth and most demanding question–the impact on the quality of scrutiny per se—into focus, and here candour requires conceding that the question cannot presently be answered empirically for any jurisdiction. What can be offered is a normative criterion against which quality might in the future be assessed, and the most useful in the Italian debate is Bresciani’s. Examining whether AI in parliament aggravates the epistocratic tendencies latent in a data-driven ‘numeric constitutionalism’, Bresciani argues that the use of AI can be regarded as constitutionally neutral only on condition that the contributions it generates remain publicly contestable—open to debate and challenge—and notes, pointedly, that the standard objections of opacity and bias apply no less to the technical contributions of human experts, so that AI might even democratize technical debate if it succeeds in widening access to specialized knowledge.⁵⁷ The criterion is exacting precisely because it locates the test of quality not in the accuracy of the algorithmic output but in the contestability of the deliberation built upon it—which is to say, in transparency and accountability rather than in efficiency.

It is against that criterion that the data opacity registered earlier ceases to be a mere limitation of the present study and becomes a substantive part of its argument. A system in use for nearly a decade, whose outputs shape the ordering of votes and the legislator’s understanding of the text, has not been the object of any published, independently verifiable evaluation of its effect on the scrutiny it mediates. That absence is not neutral. It means that the contestability on which Bresciani’s constitutional neutrality depends is, at present, incomplete: the citizen and the constitutional court can inspect the open-standard data and, for GENAI4LEX-B, the open-source code, but they cannot inspect how often and how effectively the human overseers of the operational system actually correct it. The opacity that the analyst encounters is therefore the same opacity that the accountability argument identifies as a constitutional deficit; the methodological limitation and the normative finding are one and the same. This is the strongest answer the present evidence permits to the reviewer’s challenge: not a measured verdict on whether Italian scrutiny has improved or declined, which the data do not support, but a demonstration that the very unavailability of the data is a failure of the accountability conditions that responsible AALD requires—and thus a reason, developed in the next section, why institutional self-regulation is not enough.

5. Conclusions: towards a regulatory framework for algorithmic legislative drafting

This article has examined, through the Italian parliamentary experience, the emergence of AI-assisted amendment processing as an institutional phenomenon that simultaneously promises to address structural failures of modern legislatures and introduces constitutional risks for which the existing regulatory framework is inadequately prepared. The Senate’s GEM ecosystem and the Chamber’s GENAI4LEX-B project have been presented not as isolated national experiments but as the most operationally mature and the most ex ante-designed instantiations, respectively, of a wider transformation in how democratic legislatures scrutinize and produce law. The central normative contribution has been twofold: the identification of four constitutional fault lines—consequential algorithmic error, deliberate institutional disruption, accountability opacity, and epistemic homogenization—alongside the recognition that AI tools also carry genuine potential to strengthen the capacity for scrutiny and to attenuate the information asymmetry that has driven the marginalization of parliaments; and the demonstration that the prevailing safeguard, human oversight, is necessary but not, on present evidence, sufficient, and that the very absence of data capable of establishing its sufficiency is itself a symptom of the accountability deficit that self-regulation leaves unaddressed.

5.1 The limits of institutional self-regulation

The governance response that currently prevails is institutional self-regulation: each chamber developing its own internal guidelines, oversight philosophies, and technical standards. The Senate’s ‘Human in Command’ principle, now reinforced by Law 132/2025, is the most articulate example. Its virtues are genuine: it is agile, contextually sensitive, and operationalized in concrete requirements rather than abstract commitments. But self-regulation is insufficient as a long-term constitutional response, for three structural reasons that the analysis in Section 4 brings into focus. First, parliaments that benefit from AI-assisted efficiency gains have a structural incentive to minimize the constraints they impose on themselves—and, one may add, to refrain from publishing the performance data that would expose the limits of their own oversight. Secondly, the ‘Human in Command’ philosophy is the product of a particular institutional culture with no guaranteed permanence absent enforceable requirements. Thirdly, internal frameworks have no mechanism for identifying risks that have not yet manifested, and no external accountability for the adequacy of the frameworks they adopt—which is precisely why the automation-bias literature, with its insistence on institutionalized distrust of the overseer, points beyond self-regulation. As Italian scholars have observed, the only path capable of managing the tension between institutionalized and non-institutionalized uses of AI in parliamentary proceedings is the formal revision of parliamentary rules of procedure—a ‘re-engineering’ of standing orders that would codify the conditions under which AI tools may be deployed.⁵⁸

5.2 A five-element framework built on the AI Act

The starting point for any adequate framework must be the EU AI Act, whose risk-based architecture is structurally well suited to the graduated regulation that AALD requires. The five elements proposed here build on that architecture while addressing its gaps, and are framed throughout in terms of the protection of the scrutiny function rather than of drafting in the abstract.

The first element is mandatory high-risk classification for systems whose outputs materially influence legislative decisions.⁵⁹ This should extend beyond systems generating complete draft texts to those producing summaries, comparisons, compliance assessments or predictive final texts that legislators rely upon in the exercise of their functions. The test for materiality should be functional rather than formal, and the Article 6(3) ‘preparatory task’ derogation should be narrowly construed in the legislative context, given the systemic consequences of error that distinguish legislative scrutiny from other administrative functions.⁶⁰

The second element is a legislative-process transparency obligation designed for the parliamentary context, and reinforced in the light of Section 4. Parliamentary institutions should be required to document, in the travaux préparatoires of legislation produced with material AI assistance, the specific functions performed by the tools, the human-review procedures applied, and the rate at which human reviewers in fact modified the algorithmic output, together with any known limitations of the systems used. This documentation should be accessible to constitutional courts, practitioners, and civil society as an admissible source in the interpretation of AI-assisted legislation.⁶¹ The addition of an effective correction metric responds directly to the data-opacity problem identified above: it is the minimum disclosure that would render the contestability on which constitutional neutrality depends actually operative.

The third element is an AI-authentication standard for public-consultation processes, grounded in the Act’s Article 50 transparency obligations but extended for the legislative context. Submissions to parliamentary consultations generated with material AI assistance should be identified as such, and institutions should have the authority to treat systematically AI-generated responses as a distinct category of input, subject to specific weighting and authentication⁶².

The fourth element is an inter-institutional protocol for the differentiated deployment of AI tools across the executive and legislative branches. Systems deployed by the executive for the development of legislative proposals, and those deployed by the legislature for their scrutiny and amendment, should be technically distinct, independently audited, and subject to separate governance. Where shared infrastructure is used—as may be appropriate for standards-based document-management systems built on Akoma Ntoso—the protocol should require clear functional separation with independent human oversight at each stage.⁶³

The fifth element is a parliamentary AI-literacy obligation, modelled on the Act’s Article 4 requirement but calibrated for the parliamentary context, and given particular urgency by the automation-bias evidence. The obligation should encompass not merely the technical staff who operate the systems but the members who rely on AI-generated outputs in the exercise of their functions.⁶⁴ A legislature whose members cannot critically evaluate the outputs they depend upon has delegated epistemic authority to algorithms regardless of what its governance framework says about human oversight; and, as the literature on over-reliance makes clear, literacy in this setting must include awareness of the cognitive tendency to defer to plausible and well-explained outputs, not merely familiarity with how the tools operate. AI literacy is, in this sense, not a technical competence but a constitutional prerequisite for genuinely informed deliberation.

5.3 The Italian experience as a model and a warning

Italy’s achievements are genuine. A project grounded in more than two decades of legal-informatics research, committed to open standards, developed in open source, and institutionally supervised by bodies prepared to document the scale of the disruption they have faced with statistical precision, represents the best available template for responsible AALD. The Senate’s willingness to publish, in two languages, a detailed account of that disruption is an act of institutional transparency that deserves emulation. But the limitations are equally instructive, and Section 4 has sharpened them: the governance framework remains largely internal; the classification of both systems under the AI Act remains legally ambiguous; the effect of the tools on the quality of scrutiny is undocumented and, given the in-house character of the systems, externally unverifiable; and the question of what happens when an AI-assisted process produces a text later found to contain errors attributable to algorithmic malfunction has not been addressed by any formal framework in Italy or elsewhere. These are observations about structural limits, not criticisms of institutional good faith. The Italian experiment has demonstrated that responsible AALD is achievable; the regulatory and evidential gaps identified here explain why it is not yet sustainable without further legal action.

5.4 Assistance, not authorship: the irreducibly human core of legislative judgment

The deployment of AI in the legislative process is not simply a question of administrative modernization. It is a question about the nature of democratic self-governance—about who, ultimately, examines and authorizes the legal norms by which democratic societies choose to be bound. Technologies that genuinely improve the capacity for scrutiny are to be welcomed. But technologies that improve efficiency while degrading accountability, that accelerate output while encouraging deference in place of deliberation, or that concentrate epistemic power in systems whose operations—and whose error rates—are opaque to the legislators they ostensibly serve, represent a constitutional danger that no appeal to modernization can neutralize. The framework proposed here is designed to maintain the distinction—between AI as an instrument of the legislative function and AI as a substitute for the human judgment that function presupposes—through enforceable legal requirements and, crucially, through the disclosure of the data without which the adequacy of human oversight cannot be judged. It takes both the promise and the danger seriously: it does not dismiss algorithmic assistance as inherently illegitimate but insists that its legitimacy is conditional on the maintenance of meaningful and demonstrable human control at every stage through which democratic societies give legal effect to their collective choices. Whatever the surface fluency of a generative model, the drafting of legislation—as the scholarship on legislative quality has long insisted—is not the mechanical production of well-formed text but a purposive act of judgment: the translation of contested political will into normative language, alive to constitutional constraint, to the interpretive futures a provision will encounter, and to the policy problem the rule is meant to solve. That act, and the scrutiny that tests its product, are exercises of the same human faculty—a faculty an algorithm can inform but cannot assume. The Italian Parliament has shown that AI can be made to serve that faculty rather than displace it. The task now before European legislators is to translate that institutional possibility into a regulatory requirement: to ensure, with enough specificity, including the specificity of mandated and verifiable disclosure, that what algorithms contribute to the making of law remains assistance to a human author, and never authorship in their place.

Conflict of interests

None declared.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Ethical approval

This article does not involve research on human participants or data requiring institutional ethics review.

Footnotes