Italy's Chamber of Deputies: testing generative AI to help draft legislation and check it against existing law

The problem

Writing a new law is slow and exacting work. Before a bill takes shape, parliamentarians and the staff who support them have to read across the statutes, regulations and Constitutional Court rulings already in force, to keep the new text in step with them and clear of contradictions. Legal language is technical, and reading it precisely takes time. The body of law it sits within keeps moving as well, as statutes are amended and revised, so a draft that was sound one month may not be the next.

Committee amendments add another layer. They tend to be many and detailed, and someone has to absorb what each one would change and summarise it for decision-makers. Checking that a draft respects the constitutional limits and the policy commitments already set is a further slow step. And the sources themselves, Italian and EU law, court rulings and parliamentary records, sit in different places and are hard to bring together.

The solution

In 2024, the Italian Chamber of Deputies, one of the two chambers of Italy's Parliament, ran a call for proposals for generative AI tools to support its work. Generative AI is AI that produces text or other content, and here the aim was tools that could help with research, drafting and checking legislation. GENAI4LEX-B was one of three projects picked from 28 put forward by fifteen universities and research bodies. It is led by Professor Monica Palmirani at the University of Bologna's Alma AI institute, and brings together several universities, the National Research Council's Institute of Legal Informatics and Judicial Systems, and three spin-off companies.

The project set out to take on the slower, more repetitive parts of preparing a bill. Its prototypes do three things: they surface the laws, regulations and rulings relevant to a draft; they condense committee amendments for the people who must act on them; and they test whether a draft is consistent with the rules already in force.

Underneath, GENAI4LEX-B combines two kinds of AI. Symbolic AI turns legal rules into a formal, computer-readable representation, an approach sometimes called "law as code". Generative AI sits alongside it, writing fluent text to draft and condense documents. Pairing the two lets the system produce text that is both faithful to the law and easy to read. To classify and find material, it tags documents with EuroVoc, the European Union's multilingual thesaurus of legal and policy terms, and it holds them in Akoma Ntoso, an international XML standard that puts legal texts in a shared machine-readable form so different systems can exchange them. Compliance checking runs on a separate reasoning engine called Houdini, which models legal rules in a companion standard, works through a draft to flag clashes with existing law, and surfaces consequences a drafter might not have foreseen.

The team was deliberate about where its data came from. Professor Palmirani stressed that the system used only "data coming from safe sources, parliamentary sources", and that it respected "the hierarchy of sources", so the material it worked from carried the authority that law requires.

The impact

GENAI4LEX-B is a set of prototypes still in testing, so its effect on the legislative process has not yet been measured.

1. Backed by Parliament through a competitive call

The Chamber of Deputies backed the project through its call for proposals and is the intended user of the tools, which puts a national parliament behind the approach.

2. The benefits the team anticipates

The team expects that automating research and amendment summaries could shorten the time it takes to draft and review a bill, and that keeping the tool current with legislative changes could reduce errors caused by working off out-of-date material. Professor Palmirani also argues that easier access to the relevant law and analysis could support democratic debate, giving those discussing a bill a clearer view of what it touches. These are expected outcomes from a prototype, not results from a system in service.

3. A possible route to a wider European role

The team intends to trial it with a group of users at the European Commission's Directorate-General for Informatics, which could point towards use beyond the Italian Parliament.

Lessons learned

• Training AI on legal language is unusually hard, and never quite finished. Legal terms carry precise, and sometimes contested, meanings, so a model has to be continually trained and corrected to read them correctly. The team treated this as continuous work.

• The computing costs were a real constraint, and the infrastructure proved fragile under testing. Running generative AI across large bodies of legal text is demanding, and the team felt it. Professor Palmirani described working in small pieces to keep costs down and watching the hardware closely: "there have been crashes. Once a week, while we were doing the test, there was always someone of us who was monitoring the performance of the machine because if it dropped in performance, we would pull it up." Sustaining a tool like this, she noted, requires ongoing funding and maintenance of that infrastructure over time.

• Users need to understand both what the tool does and what it has learned from. For a system used in lawmaking, the team saw explainability as essential. It made clear that the model drew only on authoritative parliamentary sources, and it produced a short guide to help members and parliamentary staff make sense of what the tool returns.

• Putting AI inside the legislature raises constitutional, not merely technical, questions. The team weighed how the tool might affect decision-makers' independence and the importance of keeping bias out. Professor Palmirani raised a sharper question about the separation of powers: "If the same tool is used by both the executive offices and the parliamentarians, how do we divide the roles and the tripartition of powers between who proposes, who governs, who legislates and who executes?" Her suggestion was that parliaments taking on such tools, especially under the EU AI Act, might establish task forces to monitor their use.

Launch year: 2024

This case study was written with assistance from artificial intelligence.