Federal Assembly of Switzerland embeds trust at the heart of technology
Parliaments occupy an unusual position in the AI landscape, carrying a responsibility to reflect democratic values in their own practices. The Federal Assembly of Switzerland, which has handled over 500 parliamentary initiatives on AI, is putting that principle into practice. Its Parliamentary Library faces a substantial operational challenge: indexing several thousand items of parliamentary business annually across three official languages using over 200 predefined topic labels – all of which creates bottlenecks, inconsistencies and staff resource pressures. A keyword-based search system compounds the problem, returning entire documents when users need specific passages.
To address this, the Parliamentary Library has partnered with the Bern University of Applied Sciences on an AI pilot that involves automated topic modelling using machine learning trained on over 30,000 historical items of parliamentary business (1995–2024). This is coupled with a retrieval-augmented generation (RAG) system combining semantic search with generative AI to deliver contextually enriched answers with visible sources.
The results transform library staff into quality controllers, improving speed and consistency. But what makes this initiative instructive is not simply the efficiency gains but how they were achieved.
Why parliaments are different: Democracy requires distinct AI principles
The Federal Assembly is not adopting AI merely to improve efficiency. That logic alone could apply to any organization. What sets this initiative apart is its explicit recognition that parliaments must safeguard their democratic legitimacy and institutional sovereignty, while placing trust at the very core of any AI deployment in the public sphere. In an era when technologies can reshape decision-making and information flows, trust is not optional: it is the foundation upon which democracy itself stands. Operational gains, however significant, cannot substitute for this fundamental responsibility.
The pilot project has been designed to ensure that AI enhances efficiency and quality through semi-automation in parliamentary processes without compromising trust or democratic values. At every stage – from conception and multidisciplinary development through to testing and deployment – AI must be transparent, explainable and accountable, as well as compliant with the Federal Constitution, the Federal Act on Data Protection and the Council of Europe Framework Convention on Artificial Intelligence. This reflects the fundamental principle that trust must be embedded at the heart of technology in parliament, ensuring that AI serves democracy responsibly, strengthens public confidence and supports parliamentary work effectively.
These considerations shaped the design decision:
- User sovereignty is non-negotiable: Unlike business AI tools that might optimize for adoption rates or engagement metrics, the Swiss system gives users the power to completely switch off AI generation and view only source documents. This is an important democratic principle embedded in the architecture. Users remain in control – always.
- Human judgement stays supreme: The system positions AI explicitly as assisting intelligence, not replacing it. Library staff retain final authority on topic assignments and RAG users verify sources themselves. This approach preserves human accountability, which is essential in institutions where decisions affect democratic processes.
- Trust is not a feature; it is the foundation: For businesses, user trust might be desirable for adoption. For parliaments, trust in institutional processes is fundamental. If MPs, staff or citizens doubt the integrity of parliamentary information systems, the institution itself is weakened. That is why transparency, explainability and source attribution are treated as core requirements, particularly in AI-supported systems that generate or mediate access to parliamentary and legal information. Mitigating risks such as loss of traceability and hallucinated content is therefore a prerequisite for operational deployment in the parliamentary context.
- Building trust is a process: The project team built trust through concrete and verifiable design choices implemented during the pilot phase. These include control features that allow users to adjust date ranges, languages and visible sources, where every generated result is presented together with the underlying parliamentary documents. Feedback mechanisms also enable users to rate outputs and provide qualitative comments. In addition, the interface makes explicit what the AI suggests and which sources are used, and enables results to be checked directly, reinforcing user control and accountability in practice.
- Example-setting is important: Like other parliaments, the Federal Assembly faces the challenge of integrating AI into its work in a manner consistent with democratic values. As institutions that regulate AI and debate its societal implications, parliaments carry a particular responsibility to reflect these principles in their own practices. By striving for ethical, transparent and human-centred AI use in internal processes, they strengthen the credibility of their oversight and legislative role. To date, the Federal Assembly has dealt with more than 500 parliamentary initiatives addressing the challenges and implications of AI. This pilot project was conceived as a modest, practical contribution, illustrating how AI can be implemented responsibly, in compliance with legal frameworks and with trust placed at the centre of parliamentary operations.
This distinction is relevant for parliaments exploring AI adoption. The question is not merely “Will this improve efficiency?”. Other questions come into play: “Will this strengthen or weaken our democratic institutions?”, “Does it safeguard human accountability?”, “Does it enhance transparency or introduce new black boxes?”, and “Will citizens and stakeholders trust it?”. The Swiss approach offers a modest, practical example: efficiency gains are pursued, but only when delivered through systems that uphold democratic values, respect sovereign decision-making and rely on clearly defined data and verified sources. Technology that functions well but undermines institutional trust ultimately risks eroding confidence in parliamentary processes.
Key implementation insights
A key insight emerged during user testing: completeness takes precedence over technical neatness. In parliamentary work, minority debates and niche issues illustrate how relying solely on the most obviously relevant sources can omit important context – a risk amplified when AI systems are used for retrieval and knowledge augmentation. That is precisely why, when prototypes were evaluated by Parliamentary Library staff, users consistently preferred systems that retrieved all potentially relevant material – even if this included some irrelevant results – rather than risk missing critical documents.
In this context, false negatives were far more problematic than false positives. This user preference directly informed system configuration, prompting a deliberate shift from the technically optimal confidence threshold of 0.1 to a lower threshold of 0.05. That adjustment highlights a central lesson from real-world deployment: technical performance metrics alone are insufficient, and professional judgement and user experience must guide system behaviour.
The pilots also surfaced several practical and transferable lessons:
- Multilingual support is not a simple extension task but requires language-specific models and careful handling of imbalanced data sets, particularly for less-represented languages.
- The choice to rely exclusively on open-source tools proved decisive: it avoided vendor lock-in, kept costs manageable and enabled deployment on existing infrastructure without specialized hardware.
- From a retrieval perspective, the study confirmed that hybrid approaches outperform single-method solutions. Combining semantic (embedding-based) retrieval with traditional lexical search consistently delivered better relevance and coverage than either approach used in isolation.
Most importantly, the projects showed that a human-in-the-loop design is not optional in parliamentary settings – it is essential. In both automated indexing and RAG, AI systems were explicitly designed to support, not replace, expert judgement. Library staff guide the systems through parameter selection, verify outputs and provide structured feedback that feeds directly into iterative improvement. This approach delivers measurable efficiency gains while preserving the contextual understanding, accountability and reliability that parliamentary information work requires. Rather than diminishing human expertise, the systems amplify it, ensuring that speed and scale are achieved without compromising institutional trust or democratic responsibility.
Why this matters now
Parliamentary institutions worldwide face similar challenges: growing document volumes, multilingual requirements, limited resources and demands for transparency. The Swiss pilot demonstrates that AI can support parliamentary work without replacing human expertise, enhancing efficiency while ensuring results remain reliable and trustworthy. Staff who are experts in supporting MPs assess AI outputs, check information and ensure that only accurate, verified results reach legislators.
More broadly, the pilot illustrates how parliaments can govern AI by example. By combining careful piloting, transparency, human oversight and continuous user feedback, the Swiss approach provides a practical template for implementing AI responsibly, balancing operational efficiency with democratic values and institutional trust. Rather than pursuing full-scale deployment immediately, the Federal Assembly opted for incremental pilot projects, allowing assumptions to be tested, feedback to be gathered, and systems to be iteratively improved without jeopardizing critical services. This “start small, learn fast” approach reduces risk while building organizational knowledge and trust.
The IPU’s recently published Guidelines for AI in Parliaments emphasize exactly this: transitioning from theoretical road maps to tangible implementations, guided by rigorous evaluation and interdisciplinary collaboration. Switzerland is putting these principles into practice, demonstrating that efficiency and democratic integrity are not competing goals – they are complementary when AI is designed thoughtfully.
Getting started: Questions for your parliament
If your parliament is considering similar initiatives, ask yourself the following questions:
Tasks and users
- What repetitive, time-consuming tasks could AI assist with in your library or information services?
- Who are your actual users and what do they need to ensure the system supports their work effectively?
Data and verification
- Do you have labelled historical data that could train models and support reliable outputs?
- How will you ensure that AI outputs are checked, validated and verified by human experts before being used for decision-making?
Governance and principles
- What democratic principles must your AI systems embody and how will you ensure user sovereignty and human accountability?
- What legal and constitutional frameworks govern AI use in your institution and how will you ensure compliance?
Trust and transparency
- How will you build trust through transparency, explainability and user control?
- How will you measure success not only in terms of efficiency, but also in terms of trustworthiness, transparency and alignment with parliamentary values?
Implementation and iteration
- What is your pilot–test–learn cycle and who needs to be involved to ensure meaningful feedback and learning?
- How will you capture user feedback, adapt the system iteratively and evaluate whether AI supports rather than replaces human judgement?
Data quality and risk management
- How will you address requirements relating to multilingualism and document diversity to ensure equitable and accurate retrieval across languages and sources?
- Are there risks of bias, errors or black-box outputs and, if so, how will you monitor and mitigate these risks and communicate them to users?
The bottom line
The experience of the Parliamentary Library of the Federal Assembly of Switzerland offers an encouraging message: the key to using AI effectively is knowing what users need and developing systems in such a way that the results are transparent, accountable and trustworthy.
The technology is ready. The tools are increasingly accessible. Parliaments must now ask themselves whether they are ready experiment, learn and adapt while staying true to democratic principles.
This is not just about adopting technology that works. It is about adopting technology that strengthens parliament as an institution, preserves accountability and maintains the trust essential to democratic governance.
For more information: See the full technical paper “Transforming Digital Access in Parliament: Topic Modeling Meets Retrieval Augmented Generation” (citation in “Sources” below).
Contact person: Ms. Jacqueline Kucera, Head of the Parliament Library, Federal Assembly of Switzerland
Sources
