The integration of large language models (LLMs) into parliamentary workspaces offers transformative potential but also significant challenges. This contribution provides a structured analysis of LLM adoption in legislative environments, examining theoretical foundations, identifying implementation options, and assessing broader implications. It situates LLMs within the wider AI landscape in parliamentary processes, exploring their unique characteristics and applicability. A comprehensive SWOT analysis evaluates strengths, weaknesses, opportunities, and threats, highlighting potential efficiency gains in legislative drafting and policy analysis, alongside risks such as bias, misinformation, political manipulation, and ethical and regulatory challenges. Key integration options are examined, including external LLM service providers, in-house deployment, and training models on parliamentary data for enhanced relevance and accuracy. Performance benchmarks tailored to parliamentary needs are also discussed. The paper emphasizes the importance of LLM literacy and capacity building among MPs, parliamentary staff, and scientific services for responsible and effective use. Finally, it reflects on the implications of LLM use in parliaments and offers recommendations for actions, stressing the need for a balanced approach that leverages technological innovation while ensuring democratic integrity, transparency, and accountability. By providing a comprehensive overview, this contribution advances the discourse on AI in parliament and offers strategic recommendations for integrating LLMs into parliamentary operations effectively and responsibly.
In 2023, Zeppelin University hosted a workshop on AI use in parliaments. International experts from academia and legislatures discussed potential applications of large language models (LLMs) in parliamentary contexts. The event yielded valuable insights but also raised major concerns. Twenty-one participants debated whether, under what conditions, and how LLMs should be implemented, as well as who bears responsibility. Discussions covered methods to detect LLM use and risks such as dependencies and lock-in effects. Key concerns included lack of transparency, limited knowledge on training and improving LLMs, and difficulties in assessing accuracy, completeness, and quality of outputs. The need for new skill sets and training attracted strong interest. Broader implications for democratic governance and the state’s role were also debated. Given the many actors seeking to influence legislation, integrating LLMs into parliamentary processes constitutes a clear high-risk area for AI use.
Drawing on workshop discussions, a graphical recording, ongoing literature reviews, and ChatGPT-based exercises, a SWOT analysis of LLM use in parliamentary work was conducted. The analysis identifies key strengths, weaknesses, opportunities, and threats from a public sector informatics perspective. Findings provide a basis for examining implementation options, potential benefits and critical challenges, particularly as institutional knowledge remains limited, and most parliaments are only beginning to explore these technologies.
The research paper therefore addresses the following research questions:
RQ1: Which strengths, weaknesses, opportunities, and threats are expected with the introduction of LLMs in the parliamentary workspace?
RQ2: What are the options for the integration of a LLM in a parliament?
Drawing on insights from workshop discussions and a literature review (sections 2 and 3), this study systematically examines the integration of LLMs in parliamentary contexts. It begins with a SWOT analysis of their use in parliamentary work environments and complements this with action-oriented recommendations (section 4). Section 5 provides a detailed examination of current integration options for LLMs in parliaments. Section 6 offers broader reflections, including considerations on the European Union’s (EU) AI Act. Finally, section 7 synthesizes key findings, addresses limitations, and outlines open questions for future research.
Recent advances in artificial intelligence (AI) and generative AI (genAI) challenge established interaction patterns and create new opportunities for parliaments. The term “artificial intelligence,” introduced in 1956, reflected the belief that machine could simulate “every aspect of learning or any other feature of intelligence” (
GenAI is characterized by its capability to generate new artifacts on the foundation of what has been learned. It does not rely on coincidences, but instead on recognized and learned patterns to generate synthetic data. It can autonomously create data, texts, processes, programs, sounds, sound sequences, voice contributions, images, and videos (
A language model is a probabilistic model of a natural language, capable of generating probabilities for a series of words based on the text corpora it was trained on, which can include one or multiple languages (
Since the release of ChatGPT, built on the LLM GPT-3.5 architecture by OpenAI in November 2022, there has been increased focus on researching the potential applications, deployment strategies, and the associated risks and challenges of LLMs (
Currently, two technical approaches are available for significantly enhancing LLMs. Retrieval-augmented generation (RAG) is particularly effective when LLMs are prone to hallucinations (incorrect outputs), when query-based models fail to deliver context-specific answers or when the accuracy of generative AI is insufficient. In a RAG-based LLM system, a query model retrieves relevant information from existing data sources. The generative model then synthesizes this information, producing a coherent and contextual response (
A second approach addresses the challenge of balancing LLMs’ powerful capabilities with strict data privacy and security requirements. A privacy proxy becomes essential when queries transmitted to an LLM or its service provider include sensitive, confidential, or personal data that must not be disclosed to third parties. A privacy proxy ensures that such data is either anonymized or made unrecognizable before being processed by the LLM, thereby safeguarding individual privacy and protecting organizational confidentiality. Secondly, it can facilitate compliance with the General Data Protection Regulation (GDPR) in the European Union by filtering out or masking personal data, ensuring that the LLM’s operations align with legal requirements and do not violate data protection laws. Thirdly, directly exposing raw data to an LLM introduces a risk if the model is compromised or misused. A privacy proxy mitigates these risks by pre-processing the data to remove unnecessary details, ensuring that only the minimum required information is exposed to the model and its service provider. This enhances user’s trust while enabling the effective use of LLMs capabilities.
The study uses a multi-stage qualitative research design that combines a theoretical grounding with explorative elements. It begins with an interdisciplinary two-day workshop involving 21 participants within a research initiative on AI and LLM use in parliamentary contexts. Participants included software developers, consultants, parliamentary practitioners, and researchers (
Additionally, a structured classification of documented examples of use cases involving LLMs in parliaments worldwide (
The core methodological element is a SWOT analysis (Strengths, Weaknesses, Opportunities, Threats;
The discussion synthesizes general findings, followed by a review of the EU AI Act in relation to parliamentary contexts. The review reveals that parliaments are explicitly or effectively excluded from the Act’s scope so far, a position critically examined considering the insights gained.
The study is conceptually positioned ahead of large-scale implementations and adopts a theory-based, reflexive approach. Its aim is to provide a robust foundation for future technical, organizational, and normative design processes, particularly regarding the secure design, deployment and operation of LLMs in parliamentary institutions.
Parliaments in democracies are characterized by representation, plenary debates, legislative voting, and oversight functions. Paper- and file-based procedures ensure the permanent preservation of speeches, drafts, resolutions, and laws. Modern information and communication technologies (ICT) support records management, legislative workflows, and diverse parliamentary processes.
AI and genAI could potentially be utilized in parliamentary settings. Researchers are actively exploring the implications of AI and genAI on parliamentary operations, assessing their potential, opportunities, risks, and challenges posed by these technologies (
A late 2022 study identified 39 AI-based solutions across nine parliaments (
Globally, debate continues on whether and how to regulate AI. In the European Union, this discussion spanned years before the AI Act took effect in August 2024, establishing a comprehensive framework for AI, its applications, and generated content. The Act applies directly across all member states, setting rules for AI use in the public sector and democratic processes (
Implementing AI in a parliamentary environment requires decisions on AI governance (responsibility, democratic control, organization;
LLMs may reshape key dimensions of parliamentary work. Their use in legislative contexts opens new prospects for enhancing debates, drafting legislation, and conducting oversight. LLM-assisted debates, LLM-supported lawmaking, and LLM-based scrutiny mechanisms suggest a fundamental shift in parliamentary workflows. Integrating such models into routine operations marks not just technical augmentation, but a broader transformation in how legislative bodies function, deliberate, and govern. Understanding these shifts is essential to evaluating both the potential and the limitations of AI-assisted democratic processes. Ultimately, this also touches on our fundamental understanding of democracy and representative governance.
Given the potential of LLMs, key questions arise about whether, when, to what extent, and under what conditions Members of Parliament (MPs), parliamentary staff and research services may have access to AI-based tools for text generation, text and data analysis, as well as image and video generation as part of their duties. Even if access is prohibited by the parliament administration, MPs can at least access such tools available online independently and use them on their own responsibility. However, frustration may grow, when professional access is denied, while private use demonstrates the clear benefits of existing LLM services. Hence, LLMs should be recognized as essential instruments in the parliamentarian workspace. Early and proactive measures are also needed to prevent harm and unauthorized data disclosure to third parties.
Since the release of ChatGPT by OpenAI in November 2022, interest in using LLMs in parliaments worldwide has grown rapidly. Some pioneers began experimenting with cloud-based models or hosted systems in collaboration with IT partners. The 2023 workshop held at Zeppelin University provided an early platform to exchange experiences, insights, concerns, and risks. Due to the sensitivity of parliamentary tasks and legislative issues, participants proceeded cautiously. Speakers presented projects and early prototypes, including examples from the European Union and Greece (
Since 2019, the Centre for Innovation in Parliament of the Inter-Parliamentary Union (IPU) in Geneva has systematically tracked AI innovations in parliaments worldwide. In April 2024, it published its first issue brief on the use of generative AI in parliamentary settings (
In October 2024, the 149th IPU Assembly unanimously adopted the resolution “The Impact of Artificial Intelligence on Democracy, Human Rights and the Rule of Law,” which includes specific references to generative AI, particularly concerning the generation of textual and visual content (
Meanwhile, in Germany, the Social Democratic Party (SPD) was the first party to publish a political guideline for the use of generative AI in parliamentary work, presented in July 2024 (
Based on the discussion in the academic workshop on AI and LLM use in parliaments (July 2023:
SWOT-analysis for LLMs in parliament.
| Strengths | Weaknesses |
|---|---|
|
Analytical Capabilities Generative AI Skills Enhanced Communication Continuous Learning Adaptability and Multilingual Skills Scalability and Consistency Rapid Response Creativity |
Dependence on Technology Dependence on Training Data Technical Limitations Hallucinations Lack of Sensitivity to False Information, Irony and Sarcasm Limited Critical Thinking Confidentiality and Security Risks |
| Opportunities | Threats |
|
Text Creation and Drafting Support Scheduling Transcription Services Translation Services Policy Analysis and Briefing Speechwriting Assistance Letter and Email Assistance Legislative Research Assistance Constituent Engagement Training and Education for Staff |
Reduced Human Interaction Lack of Human Judgment Bias and Ethical Concerns Unintended Legal Consequences Authenticity and Accountability Erosion of Human Skills Impact on Human Employment Quality of Literacy Manipulation and Misuse Unintended Political Influence |
In combination with RAG and privacy proxies, LLMs unlock new opportunities for parliaments, MPs, parliamentary staff, and research services by leveraging analytical, generative, and communicative AI capabilities. They can process vast data sets, identify patterns, and generate insights that human analysts might overlook, ensuring legislative proposals are supported by comprehensive data analysis for more informed policymaking. As assistants, LLMs could be used to draft policy documents, speeches, and reports, and pose challenging questions for public consultations. They can also be used to translate or adapt texts to specific styles or conversational tones (
With Internet access and enabled training capabilities, which are not always granted for security reasons, LLMs can continuously adapt to new data and respond to evolving political landscapes. They can address a wide range of political topics and meet the demands of daily political processes, generating documents in a desired style and providing near real-time feedback. Moreover, LLMs are capable of handling multiple topics and tasks simultaneously. Their multilingual capabilities ensure effective interaction and task interpretation in dialogs across diverse linguistic groups (
IT managers in parliaments express concerns about reliance on available LLMs and their training data, as well as existing technical limitations. Key issues include generating incorrect outputs (hallucinations), limited contextual understanding, poor handling of false information, irony, and sarcasm, and a lack of critical reflection. This stems from the fact that LLMs are not truly intelligent, and their outputs are not always correct or complete. They function as probabilistic language models with broad knowledge, rapidly generating characters, words, data, and answers (
LLMs function as black boxes (
Like most digital systems on the Internet, current LLMs provide only limited guarantees of confidentiality, data protection and IT security. These vulnerabilities pose significant risks, including cyberattacks, misuse, and unauthorized access. Particular concerning are LLM service providers in jurisdictions where political or military interests may enable surveillance of parliamentary queries and uploaded documents, especially where legal frameworks explicitly permit such access. These risks raise serious ethical and security concerns. However, some technical limitations of LLMs are expected to be mitigated in the coming years through advances in RAG and the integration of privacy proxies.
LLMs offer a broad spectrum of innovative applications in parliamentary contexts, which vary according to user groups such as MPs, parliamentary staff and research services. These applications include the generation of high-quality, accessible text, the creation of summaries, facilitating brainstorming sessions, and the production of diverse forms of creative content. In particular, LLMs can substantially support the drafting of bills, policy proposals, and administrative documents by producing coherent, well-structured, and legally sound texts. This contributes to improving the quality of legislative output, minimizing inconsistencies and reducing potential conflicts. Their strength in ensuring precision and clarity makes LLMs especially valuable for political speeches and legislative drafting, where accuracy is of critical importance for the following political decisions. Furthermore, LLMs can enhance administrative efficiency by automating routine tasks such as scheduling, calendar management, reminders, and meeting coordination. During meetings, they support agenda setting, minute taking, and real-time transcription, thereby ensuring precise documentation of discussions. In addition, many LLMs offer integrated translation and interpretation services, enabling multilingual communication within parliamentary environments. This supports inclusivity by enabling all members, regardless of language proficiency, to actively engage in discussions and fully understand relevant documents. Collectively, these capabilities strengthen deliberative processes and contribute to more informed decision-making processes within parliaments (
LLMs offer significant potential in automating policy analysis and generating concise, insightful briefings. They can detect trends, compare policies across jurisdictions in easy understandable tables, and provide evidence-based recommendations. This supports legislators in making informed decisions and developing effective policies. By analyzing relevant data, LLMs can produce clear reports, summaries, and visualizations that assists MPs in forming opinions and decisions. Their ability to anticipate legal challenges and suggest proactive solutions makes them valuable tools in the policy-making process.
LLMs can substantially support speechwriting, letter drafting, and email correspondence, particularly when combined with privacy proxies. Drafts can be adapted to the speaker’s style and audience expectations, integrating key messages, rhetorical strategies, and relevant data to enhance public communication. This ensures effective engagement with constituents, stakeholders, and officials while maintaining high standards and message accuracy. Moreover, paired with RAGs and privacy proxies, LLMs can efficiently manage complex legislative research by analyzing legal texts, case studies, laws, and precedents to deliver relevant insights.
For MPs and their staff, LLMs offer numerous applications in constituency engagement. They support analysis of public opinion, issue tracking, monitoring developments, and predictive analytics. LLMs enhance communication by managing channels, responding to inquiries, and collecting feedback. Through sentiment and trend analysis of constituent concerns and social media, LLMs enable legislators to better address public needs and views. They can also monitor emerging issues and key developments. Furthermore, LLMs can facilitate staff training by generating educational materials, conducting interactive training sessions, and providing ongoing support, ensuring that staff gain essential skills and knowledge.
The integration of LLMs in parliaments poses several potential risks that require thorough assessment (
As LLMs begin to assume tasks traditionally performed by humans, concerns arise about reduced human interaction and the erosion of human judgment. The latter remains essential, particularly for crafting nuanced policies and legislation that reflect diverse perspectives and societal complexities. Machine-generated texts often lack the experience and contextual depth needed for such policymaking, risking flawed analyses and misaligned decisions. Additionally, LLMs trained on biased or incomplete data may reinforce or amplify existing biases, leading to discriminatory policy suggestions and legislative texts. This raises ethical issues and potential legal liabilities (
Concerns also arise around authenticity and accountability when LLMs generate content for controversial or high-impact policies without clear human oversight. This lack of transparency may erode trust in parliament and its representatives, who are expected to assume responsibility. Long-term risks include skill erosion among MP’s staff, as overreliance on LLMs for tasks like interpreting legal texts and drafting legislation could weaken critical thinking and communication skills essential for effective governance. Potential human job losses and knowledge losses also raise concern, as a decline in literacy quality can undermine the overall effectiveness of parliamentary work. Additionally, socio-political risks emerge if LLMs concentrate decision-making power in fewer individuals or entities, potentially threatening democratic processes and representative governance.
LLMs can assist in identifying key, controversial, or potentially harmful positions within debates, thereby revealing weaknesses in underlying arguments. This capability may enable opposition actors to mount more forceful critiques, necessitating greater preparedness for government representatives. Moreover, the risk of deliberate political manipulation by third parties should not be underestimated (
Security vulnerabilities are critical, particularly regarding data protection and privacy. When LLMs process inputs containing real personal data, the risk of breaches and misuse of sensitive information increases. Copyright concerns also require attention to ensure that LLMs are not trained on protected material without proper authorization (
Another major concern is the potential misuse of LLMs by system-crashers seeking to undermine and destabilize democratic governance. LLMs could provide strategies to circumvent democratic structures, weaken state institutions, dismantle bureaucracy, and accelerate administrative erosion, all without adhering to legal frameworks. This scenario raises profound questions about the legitimacy and methodology of deploying LLMs as instruments in a digitally enabled AI coup d’état (
Addressing all these risks demands strong regulatory frameworks, ethical guidelines, and proactive safeguards against misuse, bias, security flaws, and privacy breaches linked to LLMs in parliamentary settings (
The following recommendations derive from a strategic TOWS analysis (
For the strengths-opportunities strategy, capitalizing on opportunities is crucial. Structured integration can systematically identify priority applications and ensure systematic training for staff and policymakers. Institutionalizing LLM-assisted workflows through formal guidelines can improve efficiency and transparency of legislative drafting, communication, and research. Combining LLMs with RAG and privacy proxies will improve precision, scalability, and accountability. Implementing LLM-driven automation for administrative tasks like scheduling, transcription, and summarization will streamline operations, enabling staff to focus on higher-level strategic work. These measures should be embedded in a robust change management framework.
The strengths-threats strategy focuses on mitigating risks by establishing oversight mechanisms to ensure human supervision of critical outputs, including legal texts and public statements. Mandatory AI literacy programs and continuous training for MPs and staff will prevent skill erosion and foster effective human-AI collaboration. Defining quality and security standards aligned with recognized frameworks is essential for validating LLM-generated content. Developing internal LLM systems or using trusted infrastructures can reduce dependence on insecure external providers and mitigate geopolitical risks. Implementing secure, ethical-by-design architectures is crucial for countering misinformation, deepfakes, and attempts at political destabilization.
The weaknesses-opportunities-strategy addresses weaknesses by embedding quality control layers and fact-checking routines into LLM workflows to minimize errors in legal and policy contexts. Leveraging LLMs for internal knowledge management and staff training can close digital literacy gaps and strengthen institutional competence. These measures ensure LLMs remain supportive tools rather than sole decision-makers, promoting responsible use by MPs, parliamentary staff, and the scientific services.
Finally, the weaknesses-threats strategy focuses on minimizing vulnerabilities through regular audits, risk assessments, and compliance with frameworks such as the EU AI Act. Establishing an AI ethics council within parliament, with a clear mandate for oversight and guidance, will safeguard responsible innovation, fairness, and ethical deployment. Without such measures, generative AI and LLM-driven system-crashers could unduly influence democratic processes in legislative, parliamentary and ministerial contexts.
A study identified the most relevant fields of application for LLMs in parliament (
Today, LLMs are no longer only experimental technologies in labs. They are available and increasingly seen as promising tools for parliamentary innovation and daily usage. However, despite growing interest among stakeholders, concrete implementation strategies remain underdeveloped. Bridging this gap requires a clear roadmap that aligns technological opportunities with institutional requirements and legal constraints. Transitioning from experimentation to operational use requires adherence to established standards for secure AI development and IT governance. These include principles covering secure design, secure development, secure deployment, as well as secure operation and maintenance as outlined by the US National Cyber Security Centre and others (
Reflecting the digital strategy, the resources and the IT governance of a parliament, there are several options for integrating LLMs into the parliamentary workspaces. Knowledge about options and consequences helps to make the right decisions. This section examines the existing options for integrating LLMs, addressing the second research question about the core features and capabilities of LLMs relevant to parliamentary functions. Key factors include choosing between on-premises and cloud-based server architectures (outsourcing options), selecting suitable service providers, and assessing the scope and adaptability of available LLM services. Additional aspects involve training modalities (custom vs. general-purpose), developing benchmarks for evaluating effectiveness, bias mitigation, and alignment with democratic principles, and fostering LLM literacy among staff and MPs, implementing structured capacity-building.
Regarding ownership, various make-or-buy options and technical implementation paths exist. Performance depends on specific requirements and differences in cybersecurity levels, national security needs, data privacy, neural network architecture, model accuracy, training data quality, supported official languages, linguistic diversity, and explainability of AI systems.
For the central make-or-buy decision regarding LLMs and their application, five strategic options can be distinguished (
In addition to the make-or-buy strategy, five server deployment options exist for hosting LLMs in parliamentary environments. These include using public cloud-based LLMs, private cloud-based LLMs tailored for parliamentary use, and fully internal LLMs hosted within parliamentary infrastructure. A hybrid approach combining internal and external solutions is also possible. Finally, parliaments may opt to tolerate the use of externally accessed, unregulated LLMs (called BYO LLMs or Dark LLMs).
Public cloud-based LLM services with proprietary licenses, such as those provided by OpenAI, Microsoft, or Google, are accessible via the public Internet and through API interfaces. Their use is governed by contractual agreements that define the rights and obligations of both the service provider and the user. Private cloud-based LLM services are only internally accessible. Alternatively, on-premises solutions allow LLMs to be implemented and operated on in-house servers or high-performance notebooks within parliamentary IT infrastructure, ensuring data protection and security. This is particularly suited for open-source LLMs, which can run locally on the intranet without restrictions. Local parameter adjustments and confidential training with proprietary documents are possible. The demand for scalable, trustworthy LLMs has grown due to sectors like legal tech and tax tech, where training data, prompts, and results must remain secure and private. This model could enable parliamentary groups to utilize their own dedicated LLMs. Hybrid solutions, the combination of public or private cloud-based and on-premises solutions to achieve a balance between flexibility and security, would be another option (
The responsibilities of LLM service providers in parliamentary contexts go beyond granting access to reliable and secure language models. They must ensure full compliance with legal, ethical, and technical standards. This includes obligations regarding data protection, transparency about model training methods and limitations, continuous monitoring to prevent misuse, and the provision of robust documentation and interfaces. These measures are essential for enabling accountable and trustworthy integration of LLMs into parliamentary workflows.
Several major LLM operate globally, including OpenAI, xAI, Google, Anthropic, Meta, Baidu, Alibaba, DeepSeek, and Mistral AI, each offering models and servers with distinct strengths and limitations. All are continuously improving their LLMs, raising the strategic question for parliaments: Should they commit to a single service provider, or adopt a platform-based approach that ensures flexibility? A platform solution would need to make LLMs and tools from different providers accessible and comparable through a standardized interface, ideally allowing easy switching between models, providers, servers, and solutions. Given the rapid pace of development, solution developers face significant uncertainty in selecting and testing LLMs and assessing their suitability for specific parliamentary applications in the cloud or on premises. Nonetheless, adopting a modular, open architecture would help parliaments to meet a key requirement: avoiding vendor lock-in effects by ensuring that any LLM service or provider can be replaced at any time.
Open API interfaces allow also the seamless integration of LLM functionalities into other existing IT infrastructures systems, including document management systems, customer relationship management systems, and workflow tools. From an information management perspective, a crucial question is which LLM service providers and services can be deemed trustworthy and appropriate in a parliamentary context, and which should be excluded due to national security or cybersecurity risks.
An answer to this challenge is provided by Rode’s ample principle (
LLM providers must offer reliable LLM services tailored to parliamentary needs, including chatbots, virtual assistants, translation, multilingual support, data analysis and forecasting, information management, training, professional development, process design, and coding. These services may be components of a broader LLM application. While reflecting on potential LLM use cases, it is important to note that language models process language but need not necessarily produce exclusively text-based outputs. In addition to generating code or BPMN-process models, LLMs can also produce images and other visualizations (text-to-image), audio (text-to-audio) and video (text-to-video). These services enable new forms of visual output based on text input and training data. Such capabilities support MPs in creating visual representations for arguments, sketches, memes, or photorealistic images of construction projects. Visual content often proves more persuasive than text, as the brain processes it more rapidly. Further inputs for LLMs may include speech (via microphone or audio files), as well as images and videos, significantly enhancing usability.
In this context, parliaments must identify which prompt types (text, image, audio, video) and AI agents (specialized services that build on the reasoning, understanding, and generative capabilities of an LLM to perform autonomous or semi-autonomous tasks) are most effective for their intended use. Developers face the challenge of designing, testing, and selecting prompts and AI agents best suited to parliamentary workflows. Where best practices exist, a platform for knowledge sharing and transfer would be highly valuable. Such a portal should provide suitable prompts and AI agents for tasks such as document summarization, translation, meeting transcriptions, public engagement, training, onboarding, and the preparation of documents and speeches.
For text-to-image LLMs, the platform should provide prompt suggestions and templates to translate ideas, collections, and designs into clear, compelling images, figures, artworks, and other visual concepts.
For text-to-video or image-to-video LLMs, prompt suggestions enabling object visualization through short films would be valuable. This allows projects to be viewed from multiple angles, illustrating their three-dimensional integration into natural environments and fostering lasting engagement.
Training takes time but is essential for compiling and applying these datasets effectively. LLMs tailored for parliamentary use can generate responses referencing official documents, including historical records from past legislative periods and earlier versions of laws. However, copyrighted material, personal data, or sensitive documents must not be used in publicly accessible LLMs. For internal systems, access to such data can be appropriate. Equally important is identifying and correcting biases in models, training data, and source documents to ensure fair and unbiased outputs.
Establishing recognized benchmarks for LLMs and their services to reliably assess response quality is highly desirable. Such benchmarks would offer a standardized framework for evaluating LLM performance and effectiveness across various applications and contexts. Currently, several benchmarks exist for LLMs, such as the TIMETOACT LLM Benchmark
However, developing benchmarks tailored specifically for parliamentary use is a crucial area for future research. This will require standardized methods that yield reliable evaluations of response behavior. Future LLM benchmarks for parliaments would enable the evaluation of freely available LLMs against the specific needs of parliamentary settings. Benchmarking services could also assess the quality of internal LLMs trained and fine-tuned on parliamentary documents. Such comparisons could highlight differences in cybersecurity, national security requirements, support for official languages, linguistic diversity, or explainable AI capabilities. Recommendations based on proven benchmarks would guide parliamentary bodies and IT teams in selecting suitable LLMs. Comparison frameworks could establish minimum standards and identify leading solutions. Such initiatives would benefit not only national parliaments but also subnational legislatures and local councils, which face similar challenges in navigating the diverse LLM landscape (
The successful integration of LLMs in parliaments also depends on adherence to established AI guidelines (
Data protection and ethical considerations must be addressed from the outset. Training programs require regular updates to reflect technological advances and reduce uncertainty. Embedding LLM deployment within broader change management efforts is also crucial. Demonstrating how LLMs and prompting can enhance efficiency and effectiveness will gain support from the staff councils and other stakeholders.
Reflecting on the SWOT and TOWS analyses reveals a range of strengths, opportunities, weaknesses and threats. What does the integration of LLMs mean for parliamentary workspaces? Their introduction promises substantial efficiency gains and improved output quality. Concrete steps depend on initial conditions, available financial resources, qualified personnel, infrastructure, political demand, LLM models, training data, confidentiality, sovereignty, and national security.
However, generalizations remain difficult, as effectiveness varies with factors such as make-or-buy decisions, LLM versions, and specific use cases. Parliaments differ and operate under strict confidentiality, making common technical approaches rare. The Centre for Innovation in Parliament of the Inter-Parliamentary Union (IPU) at least brings together parliaments worldwide to share experiences and foster collaborative learning.
LLMs can be either highly beneficial or problematic. At this early stage of transformation, it is unclear which implementations will ultimately succeed or fail. The global emergence of ChatGPT since late 2022 signals disruptive changes ahead, driven by experimentation, breakthroughs, and innovative solutions. The SWOT analysis provides critical insights for LLM adoption in parliaments, underscoring the need for cautious, well-managed integration to ensure operational readiness. To safeguard institutional autonomy, it is essential to maintain technical and contractual flexibility, particularly the ability to switch between LLM providers.
To support strategic decisions, LLM-specific benchmarks for parliamentary use should be established. These benchmarks would enable precise differentiation and evaluation of models and providers based on criteria such as quality, ethical principles, privacy, security, operational reliability, governance, oversight, interpretability, and democratic alignment. They would also provide a structured framework for determining when and under what conditions LLMs should be deployed.
Currently, only a few parliaments officially use LLMs. However, growing demand from MPs, parliamentary groups, political parties, administrative staff, and scientific services underscores the need for scalable LLM solutions and custom software tailored to the specific requirements of parliamentary work and trained on differentiated, context-sensitive datasets. Developing such capacities will be a strategic priority in the coming years. Parliaments must initiate an AI and LLM development process that enables intensive AI use, establish robust AI governance, and fosters inter-institutional cooperation. The process should reflect the first recommendations for action outlined in section 4.5. The Five Point Framework for the strategic integration of AI in parliaments (
Given these weaknesses and threats, it is unsurprising that regulatory measures are being considered by several nations and the EU. Developing trustworthy AI-based systems for parliamentary applications requires careful attention to ethical, legal, social, and safety dimensions. The EU AI Act establishes a framework for regulating and prohibiting AI across Europe. It sets harmonized rules for vendors for placing AI systems on the market, deployment, and use of AI systems in the Union, prohibits certain practices, and defines obligations for high-risk AI systems (EU AI Act 2024, Art 1). Requirements include enhanced transparency, technical documentation, compliance with EU copyright law, and publication of training data summaries. A fundamental rights impact assessment is mandatory, with stricter duties for systemic-risk applications. Foundation models must mitigate systemic risks, undergo adversarial testing, and report serious incidents to the European Commission. Cybersecurity remains a key priority.
With respect to software vendors and parliamentary contexts, the AI Act (Annex III, No 8(b): Administration of justice and democratic processes), classifies as high-risk only “AI systems intended to be used for influencing the outcome of an election or referendum or the voting behavior of natural persons in the exercise of their vote in elections or referenda.” AI use within parliament is not explicitly regulated. While elections and voter manipulation are assessed as high-risk, the absence of provisions addressing direct AI influence on political decisions by parliaments or state representatives is notable.
The reasons for this omission remain unclear. It may have been intentional to avoid constraints on AI-based parliamentary work. Alternatively, it could reflect the limited relevance of generative AI and LLMs at the time, reluctance to impose strict regulations, or insufficient consultation. Given the vulnerabilities and threats identified, regulatory approaches to AI use in parliaments are likely to evolve to prevent actions or omissions that could undermine democratic governance.
Ultimately, regulating LLM use in parliamentary workspaces rests with each parliament and its elected representatives. As guardians of democratic principles, parliamentarians should establish comprehensive frameworks addressing ethical, legal, and societal dimensions, balancing innovation with institutional protection. Achieving this requires appropriate organizational and technical measures (
This contribution explored the opportunities and challenges of integrating LLMs into parliamentary contexts. LLMs offer significant potential: advanced analysis, generative capabilities, multilingual support, scalability, and rapid response can streamline legislative processes, enhance constituent engagement, and reduce administrative burdens. Key applications include text creation, scheduling, transcription, translation, policy analysis, speechwriting, correspondence, legislative research, and staff training. These functions promise improved decision-making and accessibility. However, risks remain. Technical limitations, hallucinations, weak critical reasoning, and vulnerability to false information, irony, and sarcasm pose challenges. Ethical concerns, biases, security risks, and unintended legal consequences further complicate adoption. Reduced human interaction and lack of judgment underline the need for cautious integration. To balance benefits and risks, parliaments must establish ethical frameworks, ensure transparency, and maintain human oversight.
Building on a SWOT and a TOWS analysis, this study provides actionable recommendations and identifies implementation options, including choices between on-premises and cloud architectures, provider selection, and scope assessments. Additional priorities include LLM literacy, tailored benchmarks, and guidelines to ensure responsible deployment.
Given global diversity in parliamentary structures and IT approaches, the Five-Point Framework for AI integration (
To conclude, LLMs belong in parliaments, but with restrictions. Due to uncertainty regarding truthfulness, outputs should serve only as drafts for human review. This approach safeguards integrity while leveraging LLM benefits (
A key limitation of this study is the absence of empirical validation, as findings rely on expert judgment rather than quantitative evidence, and categorization relied on the author’s interpretation. The exploratory nature of SWOT and its context-specific insights restrict generalizability. Moreover, results represent a snapshot in time and lack verification through real-world performance data. This analysis reflects the state of technology in November 2025, 3 years after ChatGPT’s release. Future developments, driven by rapid progress and competition, will likely address current weaknesses and enable more tailored parliamentary solutions. Results may differ as technology evolves.
Open questions remain: Which technical features should parliamentary LLMs incorporate both now and as technological progress accelerates? How can benchmarks and risk assessment frameworks be developed, established and dynamically adapted? Will LLM-generated agendas improve meeting quality? Can speeches become clearer and laws more precise? Will LLMs accelerate bureaucracy reduction and strengthen democratic resilience? What societal and democratic impacts arise from integrating LLMs into parliamentary decision-making processes? How to deal with these changes in an optimistic scenario with strong governance, transparency, and training programs, in a pessimistic scenario with uncontrolled adoption and in a transformative scenario with hybrid human-AI governance? Can LLMs support compromise proposals in negotiations? Or will politicians reject this technology, feeling disconnected from the process? Linguistic differences in text generation also require deeper study.
Much work lies ahead. Responsible, transparent, and human-centered integration will determine whether LLMs become a tool for efficiency and democratic resilience or a source of new risks.
The original contributions presented in the study are included in the article, further inquiries can be directed to the corresponding author.
JvL: Conceptualization, Investigation, Methodology, Visualization, Writing – original draft, Writing – review & editing.
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
The author(s) declared that Generative AI was used in the creation of this manuscript. During the preparation of this work the author used DeepL to translate drafts from German to English. ChatGPT 3.5 was used for brainstorming. ChatGPT 4o and CoPilot were used for language editing. After using these services, the author reviewed and edited the machine-generated outputs and takes full responsibility for the content of the publication.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Edited by:
Reviewed by: