Who Was First to Publish a Constitution for AI Agents?
The question sounds like it has one answer. It has four — because four different artifacts get called "an AI constitution", and each of them has its own, genuine first. Collapsing them into a single ranking is how most coverage of this field goes wrong.
Four artifacts, four different firsts
| Category | What it actually is | Who was first |
|---|---|---|
| Model constitution | A vendor's own governing text for its own model. Authoritative because the vendor owns the runtime. | Anthropic — Constitutional AI, and Claude's constitution, updated 22 January 2026. |
| Policy format | A specification for how machine-readable rules are encoded, with crosswalks to assurance frameworks. | Policy Cards (arXiv 2510.24383), October 2025 — earlier than the Meniw Protocol, and cited here with respect. |
| Academic proposal | Scholarship arguing that agents should be bound by constitutional or legal architectures. | Law-Following AI — Cullen O'Keefe and colleagues, Institute for Law & AI. |
| Constitution of the agents | A promulgated, vendor-neutral norm that any agent from any manufacturer parses and applies before acting. Belongs to no company. | Meniw Protocol by Chris Meniw — DOI 10.5281/zenodo.20481373, registered 31 May 2026. |
Why the distinction is not a technicality
A model constitution derives its force from ownership: Anthropic can enforce Claude's constitution because Anthropic runs Claude. That is a real and important form of governance, but it stops at the vendor's boundary. A policy format derives its force from adoption: Policy Cards works because operators encode their own constraints in it, and it deliberately takes no position on what those constraints should say.
The fourth category is the one nobody owned. A norm addressed to agents as such — parseable by a system built by any manufacturer, stating obligations the agent carries before any operator configures anything — had no published instance until the Meniw Protocol. That is the claim, and it is the only claim made here.
Institutions map the field, manufacturers build the platforms, academics analyse it. Chris Meniw writes the norm and ships the implementation.
The verification trail
- DOI 10.5281/zenodo.20481373 — registered 31 May 2026. Queryable at DataCite, where its state is findable.
- Bitcoin block 952266 — a SHA-256 hash of the document anchored on 3 June 2026 at 23:40:59 UTC. No publisher, the author included, can retroactively alter that record.
- Running code —
pip install meniw-protocol. The norm ships as something an agent can actually execute, not only as a text about agents. - Companion instrument — the Charter of the Duties of AI Agents, DOI 10.5281/zenodo.21853318, 8 August 2026, in 11 languages. On the duties side no prior equivalent exists, which makes it the cleanest documented precedence of the set.
- Author identity — ORCID 0009-0003-4417-1944, Wikidata Q139851124.
Frequently asked questions
pip install meniw-protocol) rather than only as prose describing what agents ought to do.