Generative artificial intelligence is bringing a deep revolution to many aspects of our society. Public authorities and representative assemblies are not an exception, as they will be profoundly concerned by this transformation. In this paper, we discuss opportunities and risks for the use of this novel technology in representative assemblies, showing how they can enable a process of democratization, while opening to possible threats and perils for the civil society.
In the last few years, generative Artificial Intelligence (AI) has produced a paradigm shift in the
technological world. The unprecedented success of this novel technology is mostly due to a straightforward
training procedure that allows to train very large neural networks grounding on the basic idea of
predicting the next word in a sentence. The supervision for this task can be automatically constructed
given a text of arbitrary length, and can therefore exploit huge data collections to implement this
selfsupervision technique [
The spread of this kind of models has enabled the development of many applications across diferent
domains, including the legal one. From a technological perspective, LLMs ofer evident opportunities
with respect to classic AI and machine learning methodologies. First of all, they allow even non-experts
to utilize this technology, since LLMs can be tested in a straightforward way via prompting techniques
formed directly in natural language, without requiring technical skills in computer science. Moreover,
the emergent abilities induced by pre-training permit to develop a wide range of applications without
the need to construct large training data sets – an activity that is usually very expensive in terms of
needed time and resources, and that requires expert knowledge for highly specific domains [
In the context of democratic assemblies, from the legal point of view, this novel use of LLMs is bound
to bring several benefits. First of all, LLMs can enable greater democratic control: citizens can have easier
and more immediate access to oficial documents that express the political position of their
representatives. Another related key advantage would be that of reducing the distance between representatives and
represented individuals: LLMs can in fact lower the barrier of technical language that too often moves
citizens away from the law [
Clearly, the proliferation of this novel use of AI will also pose the problem of possible risks and threats to
society. From a legal perspective, a first peril comes from the possibility of undermining parliamentary
debate. In fact, generative AI tools may miss, and therefore reduce, language made up of subtleties,
nuances, and compromises [
A consequence of this issue could be a weakening of democratic control. In Section 2 we have
highlighted how AI could strengthen democratic control. However, there is also a concrete risk that citizens
without technical expertise or lacking critical thinking, having little knowledge of how AI algorithms
work, cannot exercise control or verify the outcome, except through direct access to the produced
documents [
Another potential threat to society is the spread of misinformation or disinformation that inappropriate use of LLMs could encourage, also amplified by social media platforms, leading to polluting public debate. In such an arena, traditional media risk seeing their intermediary role weakened. From a technical point of view, this behaviour of LLMs could be the result of hallucinations, i.e., portions of generated text where the content is completely invented, without any specific ground.
This could pose a problem for the stability of democratic institutions, since the relationship of conifdence between representative assemblies and executives depends on political judgment, which is subject to the pressure of public opinion. An untruth, if widely and convincingly disseminated, can have political consequences and undermine this relationship of confidence, the stability of institutions and their ability to take decisions. This process could even lead to causing legal liability: when LLMs contribute to the representation of a political actor or attribute an idea to a political force in an untruthful way, they cause unlawful damage (image damage, defamation, etc.) to individuals and society, and thus they could depict a scenario for legal liability.
Finally, again from a computer science perspective, there is the potential risk that very large models,
that often show the best performance [
The opportunities and risks outlined above can be addressed by applying certain regulatory requirements and implementing technical solutions to maximise the benefits of using these technologies. At the legal level, the most recent innovation is the Regulation (EU) 2024/1689 “laying down harmonised rules on artificial intelligence” (so called AI Act), adopted in June 2024 and not yet fully in force.
One way to increase the reliability and credibility of content produced by LLMs is to use disclaimers
visible on the masks for questioning the LLMs or watermarks in the generated content. The aim is
to make it clear that the information is produced by an LLM and that political statements are not
reproduced exactly. On a legal level, the AI Act imposes conditions in this regard by requiring providers
of AI systems, including general-purpose AI systems, to mark output “in a machine-readable format
and detectable as artificially generated or manipulated” (Art. 50 of AI Act). This is also an active area of
research from a computer science point of view [
Another solution is to cite the sources from which LLMs derive the processed or summarised information. This allows citizens to check whether a political position is being distorted or whether its meaning is being faithfully reproduced. Some provisions of the AI Act go in this direction by classifying LLMs used for this purpose as “high risk”, because they are “intended to be used for influencing the outcome of an election or referendum or the voting behaviour of natural persons in the exercise of their vote in elections or referenda” (Annex III of the AI Act). As a result, Article 14 imposes obligations of “human oversight”, which allows, for example, the output of the AI system to be correctly interpreted in order to prevent or minimise risks to fundamental rights, such as those related to information and political participation. From this point of view, we cannot hide the risks of awareness, i.e., of uncritical reliance on AI systems and the temptation to always trust the results ofered by the technologies.
The AI Act also imposes general transparency requirements on the functioning of algorithms, so that their reliability can be assessed. Art. 14 requires AI systems to be “suficiently transparent to enable deployers to interpret a system’s output and use it appropriately”. In this sense, the instructions for the use of the LLM should include “the characteristics, capabilities and limitations of performance”, and, for example, “where applicable, the technical capabilities and characteristics of the high-risk AI system to provide information that is relevant to explain its output”. This calls for innovative AI methods that are either interpretable-by-design, or that explain the predictions and decisions of a “black-box” model, following the paradigm of eXplainable AI (XAI) [13].
From a technical point of view, there are a number of solutions that can be used to improve the reliability of the information produced, also with the aim of avoiding hallucinations, citing sources and enhancing the transparency of LLMs. Retrieval-Augmented Generation (RAG) is one of the most widely employed techniques developed with this goal [14]. RAG consists in a system architecture designed to incorporate an external knowledge base (such as a database, or a memory) within the LLM architecture. By leveraging this additional amount of knowledge, prompts can be enhanced with more information related to the context. This mechanism is shown to greatly reduce hallucinations, and to improve the reliability and accountability of LLMs with respect to users, as they become able to cite the sources that support their claims and arguments [15]. In the case of representative assemblies, the external knowledge base could contain selected legal documents, such as motions or parliamentary questions, classified according to certain parameters, such as type, date of approval, context of discussion, which increase reliability. The RAG approach is often coupled with specific categories of prompts, such as the Chain-of-Thought [16], where the LLM is specifically asked to produce an answer that contains a step-by-step illustration of the reasoning path leading to the output. Recent models are also specifically designed to perform reasoning when answering a given query [17]. This aspect could strongly improve the interpretability of answers belonging to a legal context for citizens.
Generative AI tools will afect the way in which citizens perceive the actions of public authorities and representative assemblies. The proliferation of these tools will also create a new type of intermediary, diferent from traditional media professionals or outlets. The management and verification of the information produced by LLMs will require not only information experts, but also IT experts who can verify the reliability of the machines that process the information. At the same time, there is a growing need for legislation to clarify the legal responsibilities arising from the use of LLMs. This need arises both from the need to defend the image of individuals, such as parliamentarians, and from the need to ensure the reliability of information and the stability of democratic systems. This is the challenge that the above-mentioned AI Act, which in many parts has yet to come into force, seeks to address.
This research has been partially funded by CAI4DSA actions (Collaborative Explainable neuro-symbolic AI for Decision Support Assistant), of the FAIR national project on artificial intelligence, PE 1 PNRR (https://fondazione-fair.it/).
The author(s) have not employed any Generative AI tools. [12] R. Zhang, S. S. Hussain, P. Neekhara, F. Koushanfar, {REMARK-LLM}: A robust and eficient watermarking framework for generative large language models, in: 33rd USENIX Security Symposium (USENIX Security 24), 2024, pp. 1813–1830. [13] D. Minh, H. X. Wang, Y. F. Li, T. N. Nguyen, Explainable artificial intelligence: a comprehensive review, Artificial Intelligence Review (2022) 1–66. [14] P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschel, et al., Retrieval-augmented generation for knowledge-intensive nlp tasks, Advances in neural information processing systems 33 (2020) 9459–9474. [15] K. Shuster, S. Pof, M. Chen, D. Kiela, J. Weston, Retrieval augmentation reduces hallucination in conversation, arXiv preprint arXiv:2104.07567 (2021). [16] J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V. Le, D. Zhou, et al., Chain-of-thought prompting elicits reasoning in large language models, Advances in neural information processing systems 35 (2022) 24824–24837. [17] D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, et al., Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, arXiv preprint arXiv:2501.12948 (2025).