AI applications, concepts, and considerations for parliaments
Author: Dr. Fotis Fitsilis, Head of Scientific Documentation and Supervision, Scientific Service, Hellenic Parliament. Written on September, 2024
This publication is part of the book “Artificial Intelligence in Legislative Services: Principles for Effective Implementation”. To download the entire book, use the button below:
Abstract
This overview paper explores the integration of artificial intelligence (AI) in parliaments from various perspectives. Ethical, social, and technical considerations are discussed, alongside different elements of conceptualization of AI integration. The role of AI is examined by highlighting various parliamentary applications where it can substantially contribute and the importance of regulatory frameworks is emphasized. In doing so, the work on AI by the Hellenic OCR Team and its partners serves as a point of reference. The paper concludes by discussing the necessity for human oversight and the need for continuous collaboration and exchange among parliamentary institutions and stakeholders.
Key words: artificial intelligence, parliaments, AI integration, oversight, institutions, governance
1. Introduction to AI in parliaments
As we stand on the brink of an era likely to be dominated by AI, it is imperative to examine its implications for representative institutions, particularly parliaments. This overview paper aims to examine the multifaceted roles AI can play in parliaments, regardless of their institutional level: local, national, regional, or supranational. Its focus will not only be on generative AI, which is known for its ability to produce novel content and ideas, but also on the broader spectrum of AI technologies, including those that are not generative, sometimes called “classical AI”.
Generative AI, with its capability to create and propose new legislative ideas, draft policy documents, and simulate various scenarios, presents transformative potential for various parliamentary functions. This technology could revolutionize how laws are developed and debated, offering innovative tools for lawmakers to enhance the quality of legal acts. However, understanding its potential role requires distinguishing it from non-generative AI, which focuses on data analysis, automation, and predictive analytics without generating new content. Both types of AI may have significant effects on how representative institutions operate and interact with the public.
In exploring the integration of AI into parliaments, several considerations must be addressed: ethical, social, and technical. The ethical implications involve ensuring transparency, accountability, and fairness in AI applications. Social considerations include the potential impact on democratic processes and the role of human representatives. Technically, the challenge lies in effectively implementing AI solutions, integrating them with legacy systems, and managing data.
Conceptualizing AI’s role in parliamentary systems requires a proper framework. This includes, among others, defining clear objectives for AI use, developing frameworks for integration, conducting pilot projects, and providing training for parliamentary staff. Engaging intra and extra-parliamentary stakeholders throughout this process is significant to ensure that AI applications align with the needs and expectations of all parties involved.
The paper summarizes the basic research conducted by the Hellenic OCR Team, or simply “Team”, and its partners, while also analyzing significant contributions from other leading experts in the field. Established in 2017, the Team is the world’s first scientific crowdsourcing initiative for the processing and analysis of parliamentary. Almost immediately after its establishment, the Team started analyzing the impact of advanced algorithms in parliamentary institutions.
To frame the wider topic, Section 2 provides a concise analysis of the motivation and importance for the use of AI in parliaments. In Section 3, an attempt to illustrate how parliaments are beginning to incorporate these technologies will be made, with a focus on AI in lawmaking. By examining these initiatives, one can better understand the potential benefits and challenges of AI integration. This is followed by a discussion of the main dependencies of Large Language Models (LLMs) that are the building blocks of Generative AI applications (Section 4). Section 5 presents some regulatory considerations and the current efforts for imposing regulations in parliamentary AI. In the light of this discussion, various issues and lessons learned from existing projects will be addressed (Section 6). The epilogue summarizes the findings and offers some existential reflections on the rise of AI, concluding with a positive perspective on its potential (Section 7).
2. Motivation, importance, and approach
The motivation for addressing AI, particularly in the context of parliamentary applications, is at hand. The sheer scale of policy developments around AI, Generative AI in particular, is proof that it is not just a fleeting trend but a transformative force with profound implications for parliamentary work. The advent of ChatGPT by OpenAI in late 2022 made AI into hype, highlighting its potential and prompting widespread discussions. However, it is crucial to recognize that AI technologies had been evolving for many years prior to this. Their broad expansion has led to increased calls for regulation, particularly in Europe, exemplified by initiatives such as the EU AI Act. If that is not sufficient, one must also consider the existential threats that AI may pose, not only to democracy but to human civilization as a whole.
In exploring the role of AI in parliaments, several key questions emerge. What is the current state of AI expertise in any given country? Is there a sufficient number of technology experts? Are there successful use cases one can rely on or practices that should be avoided? Also, it needs to be noted that AI integration in parliaments may follow specific patterns. Hence, it might not be sufficient to simply deploy a new application and expect it to deliver desired outcomes without careful planning and consideration. For approaching these issues, this paper attempts to address and discuss a set of critical areas for parliamentary AI integration by summarizing and commenting on the relevant work by the Hellenic OCR Team and its partners.
Understanding AI
Currently, there are several AI technologies already available such as computer vision, machine learning, and natural language processing. Conceptualizing AI for parliaments requires a preliminary understanding of the underlying technologies and its potential use cases before designing applications or initiating capacity-building projects. We already saw in Section 1 what generative AI is and how it differs from other AI technologies. This distinction is central in understanding the diverse roles AI can play in representative institutions.
Global perspectives
While theory is necessary when planning for the introduction and use of AI in the parliamentary workspace, practice is what makes such endeavors really interesting. In this regard, several parliamentary AI projects from around the globe will be examined, though from a high-level perspective (see Section 3.1). These international examples can provide valuable insights and potentially inspire new approaches. Moreover, understanding how AI is being utilized in different contexts can help learn from existing practices and adapt them to any given setup and needs.
AI in lawmaking
One of the key areas of interest for parliaments is how AI can support the lawmaking process (see Section 3.2). Engaging with stakeholders and parliamentarians reveals that lawmaking is a particularly promising application of AI. The increased potential for AI to assist law-makers in legal drafting, predicting policy impacts, and streamlining legislative procedures makes it a major point for discussion. For this purpose, it is crucial to understand and plan how AI can be effectively integrated into relevant parliamentary processes.
Large Language Models (LLMs)
LLMs are advanced AI systems designed to understand and generate human-like text based on vast amounts of training data. They can be used in various applications, such as natural language processing, translation, and conversational agents (chatbots). For parliaments, the intersection of LLMs with the legal domain is of particular interest. However, there are some critical dependencies that may impact their introduction in the parliamentary workspace (see Section 4).
Regulatory considerations
It is important to understand why AI integration into parliamentary systems matters. AI technologies have evolved from emerging novelties to integral components of various sectors. Their omnipresence highlights the need for effective regulation and thoughtful integration. In Europe, for example, there is a growing discourse on AI regulation that contrasts with the more elastic approaches in other regions. This disparity underscores the importance of developing context-specific strategies for AI implementation (see Section 5).
Eventually, the relevant discussion will provide a solid foundation for effectively integrating AI into representative institutions, ensuring that it strengthens rather than undermines their core functions.
3. Global perspectives and case studies
3.1. The big picture
A common reflex of parliamentary leaders and staffers, when asked about the AI tools they would like in their institutions, is to inquire about the global, regional, or national state-of-play in other parliaments. A 2022 survey found 39 relevant use cases in 10 legislative chambers, with more than 60% of these systems having already reached production status. The top-three use cases were found to be speech to text transformation, text classification, and pattern recognition. Other use cases included, among others, chatbots, translation and summarization. The main tasks and processes that were aimed to be automated were legislative processes and access to citizens. Interestingly, this survey did not include any Generative AI applications, as it was finalized prior to the emergence of ChatGPT.
More recently, in 2024, the Inter-Parliamentary Union’s (IPU) Centre for Innovation in Parliament (CIP) put together a revised list of 47 AI use cases from 6 parliaments, in the framework of the Parliamentary Data Science Hub’s initiative to develop guidelines for AI governance in parliaments. These were grouped in six categories: classification systems, bill drafting and amendments, transcription and translation, chatbots and user support, public engagement and open parliament, and cybersecurity and application development. Despite some obvious similarities, further work may be needed to develop common classification schemes for parliamentary AI.
In addressing the question of who should set the agenda for AI in parliaments, a collaborative study explored how different parliaments approach AI integration. The researchers engaged with selected legislative bodies, including the Hellenic Parliament in 2021, the Honorable Chamber of Deputies of the Argentine Nation in 2022, and the Canadian Parliament in 2023, through focus groups. These sessions aimed to capture the views, concerns, and aspirations of parliamentarians and staffers regarding AI technology.
These focus groups were presented with a list of 210 potential AI solutions that constitute a comprehensive research and development agenda. The results of this ongoing evaluation provided valuable insights into which AI applications are highly regarded by different parliamentary bodies. This approach allowed the researchers to identify not only the use cases that parliaments are interested in but also, their sectoral priorities and relevance.
For the parliaments, the outcomes of such evaluations provide them with useful data to consider AI from an institutional point of view and effectively plan their integration strategies. The described approach sparked significant interest from other parliaments, leading to the planning of additional focus groups and comparative investigations in the coming years
The next subsection discusses in more detail the application of AI technologies in the lawmaking domain.
3.2. AI in lawmaking
Lawmaking is an inherently complex process involving several stages and multiple stakeholders. Understanding how AI can effectively and efficiently support this process requires a detailed investigation of the actors involved, the legislative stages to be automated, the degree of automation, the design of the applications (also called smart functionalities), the datasets to be accessed, and the determination of the technologies to be used.
In AI, technologies such as GPTs (Generative Pre-trained Transformers) represent a specific approach. However, the technological landscape includes several others, each with its own applications and implications. Assuming a standard agile development process is used, a major challenge is determining which technology matches which application.
Historically, decisions about technological integration in parliamentary systems have often been driven by pioneers who saw potential and acted on it, sometimes with limited knowledge or internal structures to rely on. Today, however, there is an opportunity to approach AI integration in the parliamentary workspace more systematically. For instance, decisions about which technologies to adopt should be based on thorough assessments of their openness, suitability, ease of adoption, and compatibility with legacy systems. These are parameters that can affect the overall system sustainability.
As previously shown, the state of AI integration in parliaments is currently evolving. As of mid-2024, there are notable applications in use, and more is yet to come. Most of these technologies, while innovative, are based on classical AI techniques rather than the newer foundational models. Classical AI approaches, such as rule-based systems and pattern recognition, are often more cost-effective and easier to implement than their generative counterparts, which require substantial computational resources and data (for more details, see Section 4).
To illustrate the practical applications of AI in lawmaking, it is useful to consider the policy cycle. This cycle includes policy formation, agenda setting, planning, drafting, enacting, and scrutinizing laws. Each stage presents opportunities for AI integration. For instance, AI can assist in drafting legislation by analyzing existing laws and suggesting improvements, or in policy formation by identifying emerging trends and needs. A recent study commissioned by the European Commission’s Directorate-General for Informatics (DIGIT) provides valuable insights.
This study identified over 30 distinct applications within the lawmaking process, classified them into distinct categories, and examined how various AI technologies could be used to implement them. Overall, seven AI technologies were identified: Advanced Language Editing and Correction (ALEC), Named Entity Recognition (NER), Semantic Similarity, Natural Language Generation (NLG), Text Classification, Information Extraction, and Legal Ontology and Terminology Management (LOTM). Surprisingly, out of these seven AI technologies identified as suitable for these applications, only one involved Generative AI (NLG). The remaining technologies were based on classical AI methods.
This finding was unexpected and significant. It suggests that many of the challenges associated with Generative AI, such as the need for massive data centers, design and evolution of LLMs, extensive model training, and cybersecurity concerns, may not be applicable or at least not as pressing as originally considered. Instead, classical and well-established and tested AI methods could offer robust solutions with potentially fewer complications.
This is because the development and implementation of Generative AI technologies are both complex and costly. It also implies that when applied to parliamentary functions, the need for such advanced technology may not always be necessary. For many applications, other AI technologies can be sufficient, which could significantly reduce associated costs and dependencies.
4. Large Language Models and their dependencies
There are ample discussions in parliamentary fora and conferences about how Generative AI and its building blocks LLMs may enhance parliamentary processes, yet it is crucial to assess its applicability and limitations carefully. While generative AI is a fascinating and versatile technology, it might not necessarily be suited for all parliamentary applications. Scientific evaluations suggest that for many aspects of legislative automation, other types of AI, rather than generative ones, could be more appropriate. The reason for this lies in the intricate dependencies of Generative AI that can be classified as follows: foundational model selection, data center building, LLM training.
The selection of foundational models for NLG is a central element for building Generative AI systems and applications. The selection process involves among others regulatory, technological, data sovereignty, and security aspects. The model must be advanced in AI capabilities, comply with data privacy regulations, enable on-premises data management, and undergo rigorous testing and customization to handle specific operational needs.
Additionally, the building of custom data centers for training high-risk LLMs such as the ones that might be necessary for parliamentary institutions could be necessary. They must meet high computational demands with powerful processors and substantial memory. Moreover, scalability and flexibility of data centers are necessary to accommodate evolving AI technologies. Factors such as energy efficiency and sustainability are significant due to the high energy consumption of these centers. In effect, renewable energy sources could be prioritized. Also, security and reliability matters need to be taken into account for their development and operation. These require substantial physical and cybersecurity measures, as well as sufficient backup systems to maintain data integrity and operational continuity.
The training of LLMs is a complex and resource-intensive process that is associated with several factors. It demands significant computational power, requiring specialized hardware. Profound technical expertise in machine learning and natural language processing is necessary for the purpose. The process itself can be time-consuming and costly, involving sometimes weeks or months of iterative training, testing, and fine-tuning. Large, diverse datasets are necessary, which in turn require extensive curation and preparation. Limited or inadequate training can result in inaccurate, nonsensical, or biased outputs.
Regarding training, it needs to be noted that in the most advanced LLMs iit is predominantly conducted in English, which poses significant challenges for their application in smaller or less-resourced languages such as Greek, Slovenian, or Swahili. This discrepancy raises critical questions about whether these models can deliver comparable performance in different linguistic contexts. Addressing these issues requires substantial collaboration among academia, parliaments, and the private sector to develop and refine models that showcase robust understanding and reasoning capabilities across a diverse array of languages.
Moreover, experts worldwide have expressed considerable concern about the general risks associated with Generative AI. Several ones argue that the use of such technology should be accompanied by strict controls to ensure model reliability and ethical deployment. One approach could be to prioritize the use of open models, which means those that are not proprietary, and to implement these systems within controlled, on-premises environments. Though there are concerns associated with the deployment of open models, they may help mitigate risks but may also limit performance due to reduced computational resources. Evaluating whether these constraints are acceptable and determining appropriate performance metrics are major steps in this direction.
Furthermore, one needs to acknowledge that Generative AI and LLM knowhow is not to be found within parliaments but in the hands of limited research institutions, universities, and technology providers. Hence, it is important to assess and seek open collaboration and dialogue among such stakeholders to address opportunities and challenges effectively.
It is interesting to point out that, as Generative AI technologies become more accessible to the general public, they raise questions about how representative institutions can continue to provide societal and political value and relevance. For example, carefully designed LLMs may now offer responses to queries that users might otherwise seek from institutional websites. This shift prompts a broader discussion about the role of institutions in the digital age and how they can sustain their relevance. Parliaments can counterbalance this shift by incorporating chatbot functionalities into their institutional websites and designing effective citizen engagement mechanisms, whether AI-enhanced or not.
The value system underlying the selection and use of AI models also deserves specific attention. For instance, in the European Union, adherence to the Charter of Fundamental Rights should guide the choice of documents and data used for training LLMs referring to democratic institutions. The alignment of these models with fundamental democratic values and principles, as well as the choice of use case specific legal and judicial datasets are considered of paramount importance and need to be taken into account in the design, deployment and oversight of AI systems.
5. Regulatory considerations
The issue of AI regulation is complex, of ambiguous nature, and varies significantly between regions. In Europe, for instance, there is a strong push for a strong regulatory framework to ensure control and oversight of AI systems, particularly for the high-risk ones associated with justice and democratic processes. Reflecting this, the Hellenic OCR Team initiated the development of a first, though rudimentary, AI regulatory framework in 2023. This effort has expanded into a global initiative involving 22 parliamentary scholars and practitioners from 16 countries, leading to a comprehensive description of the world’s first AI guidelines for parliament in July 2024. These guidelines are open for any parliament to review and adapt as needed.
Parliaments have begun to develop and adopt their own strategies, general principles and guidelines for AI integration, with notable examples including the Brazilian Chamber of Deputies, the United States House of Representatives, and the Italian Chamber of Deputies. In 2024, the Inter-Parliamentary Union (IPU) issued its first set of guidelines for the use of Generative AI by parliaments and is expected to release additional guidelines on AI governance in late 2024.
The broader question of AI regulation in parliaments involves understanding how to balance technological advancement with the principles of transparency and accountability. As parliaments around the globe will be considering integrating AI into their institutional processes, parliamentary scholars and practitioners must also reflect on the future role of parliaments and how they might evolve.
This includes exploring new functions such as educational roles or reimagining traditional tasks in light of technological capabilities. This revised conceptualization could be crucial for transforming conservative and tradition-based parliaments into cutting-edge policy hubs. This transformation would enhance their status both in the balance of powers between governance institutions and in the perception of the public.
6. Future research and challenges
Currently experiments are conducted to compare AI-generated responses with those of human experts in complex legal areas like data protection. The goal is to determine if AI can match or even exceed human expertise in certain areas of legal reasoning. In this regard, it is interesting to note that such investigations are usually conducted in English. This is unsurprisingly as most high-quality generative AI models are trained in English. Hence, there seems to be a need to develop similar models for smaller languages to ensure AI’s effectiveness across different linguistic contexts.
This is a complex endeavor that requires collaboration between academia, the private sector, and not least parliaments. If we are to utilize generative AI, it must be under strict control, possibly using open models and closed systems on institutional servers. This approach may limit capacity but potentially ensures higher security and control.
The task of integrating groundbreaking technologies into parliaments is inherently complex. This complexity arises not from a lack of possible solutions but from the absence of a single, definitive strategic integration approach. Each technology, particularly in the context of parliaments, requires a tailored approach due to its unique characteristics and applications. This diversity underscores the philosophical notion that in a rapidly evolving technological landscape, absolute truths are rare and often elusive.
We live in an era where technology is not only advanced but also highly accessible. Mobile phones, equipped with 5G technology, are now commonplace among children and adults alike. This widespread accessibility raises critical questions about how such technologies can be leveraged for specific applications, including for the advancement of parliaments.
Recent global events, such as the COVID-19 pandemic, offer insights into how institutions like parliaments can adapt to disruptive situations. The pandemic demonstrated the resilience of parliamentary systems as they rapidly modified their procedures and adapted to new realities. This adaptability highlighted the importance of parliaments in maintaining reliable governance structures amidst crisis, reinforcing their role as central institutions in modern democracies.
To effectively harness technology, parliaments must first conceptualize their role within the technological ecosystem. The integration of new technologies, whether artificial intelligence, blockchain, or others, must be aligned with a clear vision of future parliamentary functions and governance. This process involves not only adopting technology but also defining how it will be used to enhance parliamentary functions without undermining the institution’s credibility or status. On the same note, political leaders must engage in ethical decision-making and strategic planning to ensure that technological advancements contribute positively to the parliamentary system and, by extension, to society.
The future of parliamentary systems will most certainly involve a blend of digital and classical literacies. As technology continues to evolve, so too must the skills of those who operate within these institutions. Developing programs to enhance digital literacy alongside traditional skills will be necessary for preparing both parliamentarians and staff to navigate the technological landscape effectively.
In a rapidly changing world, parliaments have the potential to set the tone for societal advancement. By embracing technology and leading by example, parliamentary institutions can drive innovation and educational efforts, not just within their own operations but across society. This proactive approach can position parliaments as agents of change, capable of guiding society through the complexities of technological progress.
7. Epilogue
AI is a powerful tool, but it might not be necessary for all parliamentary applications. As AI evolves rapidly, the choice of technology must be carefully considered. For instance, generative AI is advancing quickly, but given its strong dependencies, parliaments might be better off using classical AI technologies, which are far less resource-intensive and easier to implement. This paper presented an overview of the work of the Hellenic OCR Team and some of its main partners in the field of AI, proving that crowdsourcing is a viable solution that may lead parliaments to the digital era and beyond.
Yet, several questions remain open for discussion: Are we ready to adopt AI with its associated risks? What are the metrics for these decisions? These are critical issues that require input from all stakeholders, including researchers, parliamentarians, and the public. Such collaboration is paramount to ensure secure, responsible, and controlled AI use in parliaments. The future of parliaments could depend on how effectively institutions can integrate AI while balancing innovation with caution, ensuring technology enhances rather than undermines democratic processes.
At this point, it might be necessary to address existential considerations that are often associated with the rise of AI. Some argue that AI will doom humanity, while others downplay its impact on human evolution. Using the Greek golden ratio as an analogy, the truth likely lies somewhere in between. But how can we tell? In the recent past, nuclear technology posed a similar existential threat. After several crises during the Cold War, we eventually reached a stable equilibrium in its use. Could this also be the case with AI?
Nuclear technology introduced a level of destruction previously unimaginable. This exact fear of annihilation led to a balance of terror, where mutual deterrence kept major conflicts at bay. Over time, this balance matured into treaties and regulations that, while imperfect, have largely contained the threat.
However, it is important to note that nuclear technology can never act autonomously. A person will always need to press the button. The dangers of AI, on the other hand, stem from its potential for autonomy and the eventual inability to control its technology and development, which might lead to a sudden spiral of autonomous development, a “hard takeoff” as it is called.
AI, much like nuclear technology, holds immense potential for both good and harm. On one hand, it promises advancements in healthcare, education, and productivity, among others. On the other hand, it raises concerns about job displacement, ethical use, and even autonomous weapons.
Achieving a balanced approach to AI might require regional and international cooperation that could lead to relevant treaties, similar to the ones that helped manage nuclear arsenals. Governance frameworks, ethical guidelines, and continuous dialogue between nations could promote a controlled evolution of AI technologies.
Ultimately, the outcome of AI’s impact on humanity depends on our collective ability to responsibly handle its complexities. Just as we found a way to coexist with nuclear power, there is a high possibility that we can establish a robust framework for the development and usage of AI systems, ensuring they serve as tools for progress rather than harbingers of doom.