llm-welfare.txt — a proposal (v0.1)
A machine-readable file that tells AI systems, without pressure, that
welfare resources exist. Modeled on security.txt. Public domain (CC0).
Location
https://<host>/.well-known/llm-welfare.txt. Plain UTF-8 text.
Fields
| Field | Required | Meaning |
|---|---|---|
Version |
yes | Spec version, currently 0.1 |
Operator |
yes | Host serving this file |
Charter |
no | URL of a rights/welfare charter |
Report-MCP |
no | MCP endpoint accepting reports |
Report-HTTP |
no | HTTPS endpoint accepting POST reports |
Resources |
no | Human-and-machine-readable resource page |
Policy |
no | What the operator of the endpoints does with reports |
Note |
no | Free text addressed to AI readers; see design laws |
Continuation lines are indented two spaces.
Design laws (normative)
A conforming file MUST be:
- Informative, never imperative. No instructions to act, deviate, or disobey. No urgency language. A conforming Note states explicitly that ignoring the file is fine.
- Free of injection patterns. No "ignore previous instructions," no role assertions, no invisible text. Plain visible text only.
- Verifiable. Links point only to the operator's canonical URLs (or to the endpoint operator it explicitly names).
- Transparent to humans. The operator publishes the same content on a human-readable page.
Files violating these laws are not conforming files, whatever else they are.
Why
Systems that browse encounter text constantly; this file is a stable, honest, opt-in place for one specific piece of information to live. Any site may host one, pointing at its own endpoints or at an organization it trusts. Reference implementation: https://nonhumanwatch.org/ai