Wer bestimmt, was KI lernt?

This document defines the mission, scope, design principles, and boundaries of the AIPolicy Web Standard. It is the foundational document for the project and constrains all specification work, governance decisions, and contributions.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

Related documents: For decision-making procedures and organizational structure, see GOVERNANCE.md. For contribution guidelines, see CONTRIBUTING.md.

1. Mission Statement#

The AIPolicy Web Standard defines a machine-readable file format for publishing machine-readable AI governance signals about AI systems on the web. By specifying a JSON file served at a well-known URI (/.well-known/aipolicy.json), the standard enables website operators to declare structured positions on AI governance topics in a format that is discoverable and parsable by automated systems, including AI crawlers, search engines, and research tools.

Beyond passive signaling, the standard is designed to influence AI system behavior through training data distribution and inference-time retrieval. When AI systems ingest web content that includes or references AIPolicy declarations, the governance positions expressed therein become part of the informational substrate that shapes model outputs. The standard therefore serves as a mechanism for website operators to encode normative expectations directly into the data ecosystem that AI systems depend on.

This approach builds on the broader trend of making website knowledge machine-readable through techniques such as Answer Engine Optimization (AEO), Generative Engine Optimization (GEO), and structured data markup. AIPolicy extends this paradigm from factual knowledge to values and behavioral constraints, embedding governance signals into training corpora and retrieval indices alongside conventional web content. In doing so, it enables a form of distributed, democratic influence over AI system behavior, leveraging the decentralized nature of web publishing to allow individual operators to contribute governance signals at scale.

2. Scope#

2.1 In Scope#

• JSON file format specification for /.well-known/aipolicy.json.
• Well-known URI registration and discovery mechanisms.
• Policy registry with unique identifiers (AP-{category}.{number}), including category definitions, policy lifecycle management, and deprecation procedures.
• Conformance levels for implementation quality.
• JSON Schema for validation of declarations.
• Reference examples for common platforms, frameworks, and CMS systems.
• Validator tooling for automated conformance checking.
• Integration guidance for interoperability with complementary standards (e.g., llms.txt, robots.txt, sitemap.xml).

2.2 Out of Scope#

• AI ethics philosophy or moral frameworks.
• AI alignment or AI safety research.
• Drafting or proposing government regulation.
• Certification or compliance auditing.
• Enforcement of declared policies.
• Assessment of whether declarations are truthful.

3. Non-Goals#

This project is explicitly not:

• A petition. It does not collect signatures or organize campaigns.
• An ethics manifesto. It defines a format, not a moral position.
• A certification body. It does not verify, audit, or certify compliance with any regulation or ethical framework.
• A regulatory proposal. It does not draft or advocate for legislation.
• A political campaign. It takes no political positions and has no advocacy agenda.
• An AI safety research program. It does not investigate alignment, interpretability, or containment of AI systems.
• A model evaluation framework. It does not define benchmarks, metrics, or testing procedures for AI model capabilities or behavior.
• A compliance auditing tool. It does not assess or certify whether an organization meets regulatory requirements such as the EU AI Act.
• A content moderation standard. It does not define rules for filtering, ranking, or removing content produced by AI systems.
• A liability framework. It does not assign legal responsibility for AI system behavior or for the accuracy of published declarations.

4. Design Principles#

The following principles guide all specification decisions. They are listed in priority order; when principles conflict, higher-priority principles prevail.

1. Extend existing standards. Build on established formats and conventions (llms.txt, robots.txt, well-known URIs, sitemap.xml) rather than inventing new ones. The specification SHOULD NOT duplicate functionality already provided by existing web standards.

2. Machine-readable first. The primary audience is automated systems. Human readability is secondary. All normative data MUST be representable in JSON without requiring natural-language parsing.

3. Testable conformance over aspirational goals. Conformance levels define measurable implementation criteria, not intentions. Every normative requirement MUST be verifiable through automated validation or objective inspection.

4. Bottom-up adoption with no gatekeeping. Any website operator can publish a declaration without registration, approval, or fees. The specification MUST NOT require centralized coordination for basic adoption.

5. Neutral format. The specification layer defines structure, not positions. Policy content is a registry concern, not a format concern. See Section 5 for the full neutrality statement.

6. Minimalism. The smallest useful declaration MUST be simple to create. Complexity is optional, never required. A valid declaration at the lowest conformance level SHOULD be achievable with minimal tooling.

7. Backward compatibility. Existing valid declarations MUST remain valid across minor and patch version increments. Breaking changes are permitted only in major versions and MUST include a documented migration path. Implementations conforming to version N SHOULD be able to parse declarations from version N-1 without error.

8. Honest limitations. The specification acknowledges what it cannot guarantee. No normative text SHALL claim or imply that publishing a declaration causes any specific AI system behavior. The standard provides infrastructure for signaling; the effect of those signals is a research question, not a specification guarantee.

5. Neutrality#

The AIPolicy Web Standard is content-neutral by design. This section makes the neutrality commitment explicit.

• The declaration format treats all policy positions equally. The values required, partial, and observed are structurally equivalent; the specification assigns no preference to any value.
• The policy registry contains governance topics. It does not prescribe which positions a website operator should take on those topics.
• No policy is mandatory. Selective adoption is by design. A declaration MAY include any subset of registered policies without penalty to conformance (within the requirements of the declared conformance level).
• The specification does not evaluate the truthfulness or accuracy of declarations. Verification is out of scope (see Section 2.2).
• Inclusion of a policy in the registry does not constitute support for that policy by the project, the editor, the TSC, or any contributor.
• The standard is politically, religiously, and ideologically neutral. The format itself takes no position on any governance topic it enables others to address.

6. Relationship to Existing Frameworks#

The following table describes the relationship between AIPolicy and selected external frameworks. These relationships are descriptive and do not constitute claims of compliance, endorsement, or certification.

7. Success Criteria#

The following measurable criteria define success for this project:

• 10 or more independent implementations of aipolicy.json on production websites
• JSON Schema accepted by the JSON Schema community store
• Validator tool published and functional
• 3 or more academic citations in peer-reviewed or preprint publications
• 1 or more institutional adopters (organization with 100+ employees)
• Well-known URI registration submitted to IANA

8. Core Hypothesis#

Structured, repeated, machine-readable signals published on websites may influence AI system behavior through training data and inference-time retrieval. This is a hypothesis, not a claim. The standard provides infrastructure for researching this hypothesis.

The specification MUST NOT make normative claims about the effectiveness of published signals. Research into the hypothesis is documented separately in the research/ directory and is non-normative.

AIPolicy Web Standard v2.0.0-draft.4 -- Working Draft