AI in Swiss Administration: Between Strategy and Reality

Blog (EN)

A critical assessment — with facts instead of assumptions As a counterpoint and supplement to AI in Federal Administration

The Political Offensive: Efficiency as the Magic Word

In the Swiss Parliament, there is cross-party consensus: AI should make the administration more "efficient." FDP National Councillor Andri Silberschmidt demanded with a postulate the "efficiency improvement of administrative processes through process automation and artificial intelligence." The Federal Council recommended acceptance, Parliament followed.

Source: Republik: Rather AI than Civil Servants? (July 2025)

But those who look closely notice: Many parliamentarians talk about AI as if the technology were invented yesterday — while the Federal Administration has been operating productive AI systems for years.

What the Confederation Actually Does: More than "Nebulous"

The claim that the Confederation acts "nebulously" does not withstand fact-checking. Clear governance exists:

The Competence Network CNAI

The Competence Network for Artificial Intelligence (CNAI) at the Federal Statistical Office maintains a publicly accessible project database with over 60 AI projects from the Federal Administration — including subject area, responsible office, and contact persons.

Source: CNAI Project Database

Concrete AI Projects in Operation

Contrary to popular belief, productive systems are already in use:

  • ADELE (Federal Statistical Office): AI system for automated categorization of land use using satellite imagery
  • Downy Mildew Forecasting Model (Agroscope): Machine learning model for viticulture
  • Several chatbot projects in various departments (EAK, BFS, SECO, VBS)

Source: Netzwoche: These Five Chatbots Are to Relieve the Federal Administration (October 2024)

Strategy and Implementation Plan

On September 13, 2024, the Federal Council tasked the Federal Chancellery with developing an AI strategy. This was adopted with three principles:

  1. Responsible use while maintaining data protection and information security
  2. Competence building in the administration
  3. Efficiency improvement through relief from routine work

Concrete implementation measures are to be defined by the end of 2025.

Source: Federal Chancellery: Artificial Intelligence

The Swiss LLM: Apertus — Open, Not Closed Off

The rumors about a "secret Federal LLM only for the administration" are false. Reality is the exact opposite:

Apertus is the first large Swiss language model — developed by ETH Zurich, EPFL, and the national supercomputing center CSCS. It was published in September 2025 and is completely open source:

  • Source code public
  • Model weights downloadable
  • Training data transparent
  • Over 1,000 languages (incl. Swiss German and Romansh)
  • 15 trillion training tokens
  • Two variants: 8 billion and 70 billion parameters

The model is the first LLM to fully meet the transparency requirements of the EU AI Act and is available to everyone via Hugging Face and Swisscom.

Sources:

The Justified Criticism: Structural Challenges

Despite these advances, there are legitimate concerns:

1. Fragmentation and Resource Shortage

While the Confederation moves forward, many municipalities lack the foundations for basic digitization. The federal patchwork remains a challenge.

Source: Inside-IT: Swiss Municipalities Lack Resources for Digitization (June 2024)

2. Security Concerns

The Xplain scandal showed that data protection was neglected by IT service providers. The question of how AI systems should access sensitive administrative data without creating new attack surfaces remains virulent.

Source: Inside-IT: Data Protection Was Neglected at Xplain (May 2024)

3. Regulatory Gap

Switzerland does not yet have comprehensive AI legislation. Only on February 12, 2025, did the Federal Council discuss a situational analysis. A consultation draft is to follow by the end of 2026 — thus with a multi-year delay compared to the EU.

Source: Federal Chancellery: AI Regulation

4. The "Efficiency" Discourse

The fixation on "efficiency" often conceals what is really meant: staff reduction. The SVP explicitly expresses positive views on the idea that AI systems could reduce staffing levels. SVP National Councillor Franz Grüter: "Our Federal Administration has grown strongly in recent years — both in staffing and costs."

Source: Republik: Rather AI than Civil Servants? (July 2025)

The Uncomfortable Questions That Must Be Asked

  • Who checks AI systems for discrimination? The Austrian AMAS case shows: algorithms reproduce historical disadvantages.
  • Which decisions may be automated? The new Data Protection Act allows "automated individual decisions" — without clear boundaries.
  • Who is liable for errors? Responsibility remains with humans — but how does a case worker control a model with billions of parameters?
  • Where is the public debate? While Apertus exemplifies transparency as an open-source project, a broad societal discussion about AI in administration is lacking.

Conclusion: Differentiation Instead of Polemics

Swiss AI policy is neither a "glass case without knowledge" nor a model example. Reality lies in between:

What works well:

  • Structured governance through CNAI
  • Transparent project database
  • Apertus as a globally unique open-source LLM
  • Clear strategy with implementation plan

What is missing:

  • Legal framework (delayed until 2026+)
  • Resources at the municipal level
  • Societal debate about automated decisions
  • Clarification of liability issues

A critical perspective remains necessary — but it should be based on facts, not rumors about "secret" Federal LLMs.

Sources

Official Sources

Netzwoche

Apertus / Swiss AI Initiative

Other Media

First published on clarus.news

Zurück zur Übersicht