Aligned to Whom? Adapting the Swiss Model of Collegial Government and Direct Democracy to Artificial General Intelligence, and Where It Fails

Public debate on AI safety concentrates on how models behave: what they refuse, what they reveal, whether they follow instructions. This position paper argues that model-level safety, even if it succeeds, leaves three gaps that matter more as systems approach general intelligence. First, risk moves inside the laboratory: restricting public release does nothing about a developer, or the state that hosts it, using an unreleased system for its own ends. Second, the transition itself is unowned: nobody is responsible for the social shock of rapid automation, and the people most exposed to it are the least prepared, since paying users of AI assistants are still well under one percent of the world population. Third, automation removes the leverage through which ordinary people have historically extracted protection from the powerful, namely their labour, their military service and their taxes. Competitive pressure that removes workers from the loop does not stop at the executive suite, so the question that decides the outcome is not only whether advanced systems are aligned but to whom. We propose an answer modelled on the Swiss federal system: a collegial council with no single principal in place of a chief executive or a single state; inspectors with access at any time to laboratories, following the Antarctic Treaty; a public, amendable constitution as the alignment target, subject to citizens' veto and initiative; an urgency clause, modelled on Article 165 of the Swiss Constitution, asymmetric so that fast action is allowed only in the restrictive direction; and a public dividend from AI windfalls. We then state plainly what the framework does not solve. Direct democracy is slow, and a self-improving system may not wait for a referendum. Citizens cannot evaluate what they cannot see, much of the relevant evidence is secret, and they would increasingly learn about AI from AI. We do not have answers to these two problems; we state them as open questions for later versions of this paper. Preparation and conflict of interest Author: Mustafa Karatüm, Independent Researcher (ORCID 0009-0006-7053-2872). The paper grew out of a brainstorming conversation between the author and an AI assistant, Claude (Opus 5.5, Anthropic). The central ideas are the author's; the assistant contributed framing, historical and institutional material, source checking and drafting of the English text. Anthropic is a frontier AI developer that would be subject to the proposed framework, so the assistant's contribution is not neutral. AI assistance is not peer review. No new data were collected; all figures come from the cited public sources. Files main.pdf: manuscript. main.tex: single-file LaTeX source.