Executive Summary

The Centre for Democracy Aarau conducted a pilot study on artificial intelligence and multilingualism in the federal administration on behalf of the Federal Delegate for Multilingualism and presented it in Bern on June 16, 2026. The study shows that AI tools such as machine translation and large language models can break down language barriers, but simultaneously carry the risk of losing cultural specifics and endangering cohesion between language communities. Translation quality varies significantly: German is supported considerably better than French and Italian, while Romansh barely functions at all. AI is already widespread in the federal administration, although the legal framework is still incomplete.

People

Topics

  • Artificial intelligence and language policy
  • Multilingualism in public administration
  • Machine translation and language quality
  • Cultural diversity and national cohesion

Clarus Lead

The study reveals a central dilemma for the Swiss administration: While AI tools simplify administrative processes and facilitate the recruitment of members of linguistic minorities, there is a danger of gradual erosion of multilingual practice. The asymmetric performance of AI systems in favor of German reinforces existing imbalances between language communities and could in the long term undermine the unifying function of languages in federal Switzerland. This requires rapid political decisions about the use of these technologies.

Detailed Summary

The pilot study documents that large language models and translation tools are already anchored in the daily work of the federal administration – without a complete legal framework in place. For French- and Italian-speaking employees, these tools offer a practical advantage, as they enable work in the preferred official language without requiring manual translation. This could increase the attractiveness of public service for linguistic minorities.

At the same time, the study identifies three critical risks: First, AI availability tends to promote monolingual communication rather than multilingual exchange, which erodes mutual language comprehension. Second, cultural specificity is lost in machine translation – expressions, regional references, and contextual nuances are smoothed over. Third, continuous AI use weakens active language competencies of users, which in the long term reduces the quality of human multilingualism.

Nicoletta Mariolini concludes from the findings that multilingualism will be shaped by AI in the future and that conscious political steering is necessary to use this technology in the interest of linguistic diversity. The Language Ordinance should serve as a guiding framework, but is not yet sufficiently specified for AI scenarios.

Key Findings

  • AI tools are already widespread in the federal administration, although the legal framework is incomplete.
  • Machine translation works significantly better in German than in French, Italian, or Romansh – a systemic imbalance that reinforces linguistic inequalities.
  • AI facilitates administrative work and recruitment of linguistic minorities, but simultaneously promotes monolingual communication and cultural loss.

Critical Questions

  1. Evidence and Data Quality: On what basis is the assessment of translation quality for Romansh made? Were concrete error rates or examples documented?

  2. Conflicts of Interest: Who funds the Centre for Democracy Aarau, and are there institutional dependencies on the Federal Chancellery that could influence its recommendations?

  3. Causality and Alternatives: Is it documented that AI actually leads to less multilingual communication, or does this reflect existing trends in administrative culture? Could targeted training calibrate AI systems toward cultural sensitivity?

  4. Feasibility of Recommendations: The study calls for measures but names these only implicitly. Which concrete regulatory or technical measures are considered priority, and who bears responsibility for implementation?

  5. Technical Asymmetries: Does the study explain why Romansh is particularly poorly supported by AI systems – is this a data volume, infrastructure, or design problem?

  6. Language Competency Erosion: Are there data or scenarios that show at what level of AI use measurable competency loss occurs?

Bibliography

Primary Source: Challenges of Artificial Intelligence for Multilingualism in the Swiss Federal Administration – Centre for Democracy Aarau / Federal Chancellery, June 16, 2026

Normative Foundations:

Verification Status: ✓ June 16, 2026

This text was created with the assistance of an AI model. Editorial responsibility: clarus.news | Fact-checking: June 16, 2026