Artificial intelligence is increasingly being explored as a tool to support the legislative process, from identifying gaps in existing laws and engaging citizens in the democratic process, to drafting, amending, and evaluating legislation. While AI should never replace democratic decision-making, it can help legislators and parliamentary staff make laws clearer, better aligned with policy goals, and easier to implement. The European Parliament has already invested in digital tools, including AI-powered applications, to support the legislative work of MEPs and staff.

The upcoming STOA workshop “AI for legislative work” on 4 November 2026 will explore the potential of AI across the legislative cycle, highlighting both its opportunities and the challenges it raises. The workshop will also feature a talk from Monica Palmirani, Professor of Computer Science and Law at the University of Bologna (Italy). In this interview, she discusses the issues around implementing AI in the democratic process and presents the award-winning GENAI4LEX project, designed to support legislative work in the Italian Chamber of Deputies.

---

What does it actually mean to use artificial intelligence in the process of drafting legislation? At which stages of the legislative process can AI intervene, and what are currently its most promising uses?

Monica Palmirani: AI can support the day-to-day work of parliamentary staff in practical ways. Members of parliament and civil servants can benefit from AI, particularly when preparing to draft a bill. For example, they can use it to retrieve relevant information and analyse how a proposal relates to existing legislation.

AI can help with retrieving legal definitions already established in previous legislation and checking for consistency. It can make it easier to apply amendments during the legislative process, saving time, or check references in order to avoid errors. Another application of AI in legislative drafting is creating an executive summary of a proposal and checking it against the full text, particularly when last-minute changes are made.

AI is also useful for translating legislation when institutions are required to publish it in multiple languages. This is relevant not only to the European Parliament: Switzerland, for example, must publish its legislation in three languages.

AI can also spot whether a law contains bias or is aligned with a particular policy. For example, if a government's or parliament's agenda includes gender balance, digital transformation, or climate change, AI can check whether a bill complies with these policies or whether, in practice, it does not.

These are some examples of how AI can support – but not replace – legal professionals, who must closely monitor its output and carry out quality checks.

---

How does GENAI4LEX work, and what concrete problem in parliamentary work was it designed to address?

Monica Palmirani: GENAI4LEX began thanks to Italy's Chamber of Deputies, which in 2024 launched a call for expressions of interest, inviting universities across Italy to submit projects. The GENAI4LEX project was carried out by a consortium of around 30 professors and researchers.

The main idea was to collect heterogeneous documents – such as Italian legislation, European legislation, Constitutional Court decisions, and bills and proposals from the Chamber of Deputies – and transform them into a common standard, called Akoma Ntoso, which is designed to be machine-readable and fully digital. On top of this harmonised standard, the project aimed to provide AI services, such as searching for definitions across multiple databases.

For example, if I want to draft a bill on end-of-life issues, I would want to understand which definitions exist in European-level documents, Constitutional Court decisions, or other sources. I can query the system and retrieve information from this heterogeneous dataset. We use AI because this is obviously not just a matter of matching words: we need a database of semantic vectors. If a definition does not exist in these repositories, generative AI can attempt to create one, using the available material to suggest a definition.

Another application was the calculation of what we called a probable unconstitutionality index. For each article of a bill, we check whether there are Constitutional Court decisions on similar issues and what the outcomes of those decisions were. We then return a list of potentially relevant Constitutional Court decisions that should be taken into consideration.

The GENAI4LEX project was evaluated positively by Italy's Chamber of Deputies, and some of its elements have since been integrated into the Chamber's platform.

---

How do you integrate transparency and accountability in this process, in an area as critical and delicate as legislation?

Monica Palmirani: As researchers, we try to ensure that evidence is available by recording critical information in metadata. This information can then be provided to decision-makers to support transparency.

For example, if the system tells you that a bill has a 70% probability of being unconstitutional, we try to explain step by step how that result was reached. We also provide the documents used to calculate the final score and present the information in a legible, clear way for decision-makers. It's important to explain the criteria on which the AI system operated.

---

European law involves navigating an enormous amount of material: EU legislation, national legislation from 27 Member States (all in different languages), case law, parliamentary documents, and different legal traditions. What are the benefits of adopting AI in the European legislative context?

Monica Palmirani: Comparative law can benefit greatly from comparing terminology and translations across languages with AI. Suppose we want to examine how an European directive has been implemented in Italy. We can use AI to check whether certain articles are in direct conflict with the directive or not. In many instances, European legislation needs to be adapted to local circumstances. When we want to look at these specific provisions across member States, AI makes this kind of comparison across languages much easier.

Using AI, we can also create a map of connections and identify, for example, a Spanish law that resembles an Italian law in a particular respect and shares a common basis in an EU directive. This kind of legislative mapping is a major advantage of using large language models, even when the relevant domestic laws are written in different languages. This function can also be useful for States aiming to join the European Union: for example, if we wanted to measure how closely Albania's national legislation aligns with the European community, we could use AI to make that comparison, even though Albanian is not one of the official languages in the EU.

---

How should politicians and parliamentary staff be trained to use AI correctly and transparently? What kind of AI literacy should become part of democratic institutions?

Monica Palmirani: Civil servants and politicians alike should understand both the risks and the benefits of AI, and training is fundamental for everyone involved in the legislative process.

The first step in AI literacy is to make everyone understand that AI is a tool that requires an additional level of critical thinking. Critical thinking is always important, but when using this kind of technology, we need to strengthen it through counterfactual thinking, fact-checking, and seeking different or opposing opinions. We should not ask AI to confirm our hypotheses, because chatbots can be very effective at reinforcing our existing opinions and biases.

---

What is the one thing in the democratic process you would never want to delegate to an AI system?

Monica Palmirani: I know that some parliaments around the world have wanted to do this, but one thing I would not want to see is AI being used to calculate or predict the likelihood that a bill will be approved.

In my opinion, using AI to predicting a bill's chances of success is risky for democracy because such predictions could be used to discard bills with a low probability of success. This would undermine one of parliament's most important roles: opening up discussion and enabling participation. Minority opinions must be brought to the table, even when those proposals are unlikely ever to become law: they may open a discussion that society is not yet ready to have, but that needs to begin somehow.

As of today, generative AI is based on probabilistic mathematical models. But legislators should imagine a future that disputes society's current trends. I don't want the trends of the past to determine the trends of the future without the vision of a better society.