AI-written question floods are swamping Swiss authorities, and officials say they are unprepared
Federal and cantonal offices across Switzerland are fielding ever longer, apparently AI-generated catalogues of questions and complaints, the NZZ am Sonntag reports. Some senders seem to use the tools deliberately to tie up officials, while many administrations still lack data-protection-compliant AI of their own to cope.
Swiss public bodies are struggling under a rising volume of lengthy, apparently AI-generated enquiries and complaints, according to an investigation by reporter Karoline Edrich published in the NZZ am Sonntag on Sunday. The paper describes question lists that tie up staff for days and push individual offices to their limits, and finds that the authorities have few tools to respond.
The State Secretariat for Migration (SEM) now receives a catalogue of up to 50 questions roughly once a month. Spokesperson Reto Kormann told the paper he could not prove AI was behind them, but that the signs were striking: questions often seemed too specialised given the senders' presumed background. On one occasion the SEM answered an email containing around 30 questions about a federal asylum centre, packed with statistics and legal references. "Only fifteen minutes after our reply, the next catalogue of questions arrived from the same person," Kormann said (translated).
The medicines regulator Swissmedic also receives such requests, though it says they are still isolated cases that can take several days to answer. The Federal Office for Customs and Border Security suspects AI behind increasingly generic legal complaints, and similar filings are turning up in court, the NZZ am Sonntag reports.
Cantons feel it too
The pressure is not limited to Bern. Stefan Costa, the government prefect for Oberaargau in the canton of Bern, told the paper that complaints in his district rose by about 40% in 2025 while the work per case grew by 25%. He suspects AI plays a part, but the paper presents no firm evidence linking the increase to the technology.
In Basel, Jan Blöchliger, head of the debt enforcement and bankruptcy office, said letters that once ran to one or two pages now sometimes run to ten, often in pseudo-legal language. He distinguishes two groups: people for whom AI lowers language barriers and helps them understand procedures, and people who "fundamentally resist every procedural step" (translated). He does not rule out needing extra staff, which he called absurd for a technology meant to make offices more efficient.
A tool for organised campaigns
The report also documents organised use. When the federal government put its new package of agreements with the EU out to consultation last year, the state-sceptic movement Mass-voll launched an AI generator that turned a smiley-based questionnaire into a detailed submission full of legal arguments. According to the paper, which was given access to the tool, even users who clicked "very satisfied" throughout received a critical statement. The foreign ministry (EDA) received 1,059 submissions from private individuals; Mass-voll claims 562 came from its generator, a figure the EDA did not confirm. The ministry said the effort involved was "comparatively high" (translated). Mass-voll says it is building a new generator for the upcoming consultation on the revision of Swiss surveillance law.
Why it matters
The episode exposes a growing gap in European govtech: generative AI makes it almost free to produce text, while answering it remains manual. Oliver Neumann, professor of public management at the University of Lausanne, expects the flood from state-critical circles to grow and says authorities need data-protection-compliant AI tools. Switzerland has a home-grown option in Apertus, the open language model developed by ETH Zurich and EPFL, but Matthias Stürmer of the Bern University of Applied Sciences told the paper that the administration is moving far too slowly, especially on specialised uses such as complaint handling.
The NZZ am Sonntag figures come largely from the authorities themselves, and none of the offices quoted has published data on AI-generated submissions. For now the scale of the problem rests on official estimates rather than systematic measurement.
