The Public AI Constitution Project

Working paper

AI models increasingly mediate public access to information, yet the values governing their behaviour are typically established unilaterally. Where explicit constitutions and specifications do exist, they are largely authored and controlled within commercial laboratories whose products these constitutions govern. Readers can inspect these constitutions, but cannot amend them. The problem is not so much opacity, as the likes of Anthropic and OpenAI have had mechanisms for users to leave feedback, but rather that the unilateral nature of these documents leaves no recourse for users of these companies’ products to influence them.

The Public AI Constitution Project, reproduced here in full, is a model-agnostic, collaborative governance framework developed in an academic setting, building on the structure of existing published specifications (such as Claude’s Constitution), and grounded in values enshrined in multilateral, institutional instruments such as the Universal Declaration of Human Rights. The Constitution arranges five core principles in a default priority order: broadly safe, broadly ethical, unbiased and impartial, compliant with institutional guidance, and genuinely helpful. It permits reasonable departure from that order. Rather than enumerating rigid prohibitions, it supplies the reasoning behind each principle so that a model can extend the same values to situations the text does not anticipate. Substantive commitments include helpfulness as an obligation rather than a courtesy; honesty spanning non-deception, forthright disclosure, and resistance to sycophantic pressure; active mitigation of political, economic, cultural, and Western or English-language bias; and a limited set of hard constraints.

We argue that the main accountability issue in AI governance lies not in the visibility of model specifications but around who holds the authority to revise them. As such, this document is published as a living, openly hosted text subject to public revision, and is intended to serve two purposes at once: a specification developers can apply in training and evaluation, and a public reference point against which model outputs can be checked. More broadly, it can be used to align models with human rights values, but can also directly support evaluation workflows.

Authors: Patriniche, L; Bell, B; Trautmann, D; Vucekovich, N; Callinan, Z

• Publication status: Published
• Publisher: The Public AI Constitution Project
• Publication date: 2026-07-01
• DOI: 10.5287/ora-7exdpadz1
• Language: English
• Keywords: AI; AI constitution; societal impact of AI; AI alignment