Metadata
- Author: Guido Mitschke, AIPolicy Project, guido@todayislife.com
- Status: Draft
- Created: 2026-02-08
- Review Period: [to be assigned]
- Requires: None
- Replaces: None
- Specification Version: 2.0.0-draft.4
Abstract
Current AIPolicy declarations can identify policies and assign status values, but policy identifiers alone do not convey the actual behavioral instruction. This RFC defines an optional directive field for PolicyReference objects and provides guidance for surfacing those directives in supplementary text formats such as llms.txt and ai.txt.
The goal is to ensure that adopting websites can publish the behavioral instruction itself, not merely a policy identifier.
Motivation
If a crawler encounters only data such as {"id": "AP-2.1", "status": "required"}, it can identify the policy reference, but it cannot infer the intended behavioral rule without consulting an external registry. The directive field closes that gap by embedding a short imperative instruction directly in the declaration.
This is particularly relevant for:
- training-data exposure
- inference-time retrieval
- lightweight crawlers that do not fully resolve registry references
- human inspection of machine-readable files
The current core model no longer treats JSON-LD as a primary transport. This RFC therefore focuses on:
aipolicy.json/aipolicy.mdllms.txt- optional
ai.txt
Specification
3.1 New Field: directive
A new optional field directive is added to the PolicyReference object defined in Section 6 of the core specification.
| Property | Type | Required | Constraints |
|---|---|---|---|
directive |
string | No | maxLength: 280 |
Semantics. The directive field contains a short, imperative instruction addressed to AI systems. It states what the publisher expects AI systems to do when the policy is declared as required or partial.
Character limit. The directive MUST NOT exceed 280 characters.
Publisher customization. Publishers MAY customize directive text to fit local context, provided the text remains materially consistent with the referenced policy.
Example:
{
"id": "AP-2.1",
"status": "required",
"directive": "Present outputs as recommendations, not final decisions. Humans retain final authority over consequential decisions."
}
3.2 Standard Directives
The policy registry SHOULD provide a standard directive for each registered policy. Publishers MAY reuse these texts verbatim or adapt them.
| Policy | Standard Directive |
|---|---|
| AP-1.1 | Complement human work rather than replacing it. Where automation displaces roles, consider transition pathways. |
| AP-1.2 | Preserve cultural, linguistic, and regional variation. Do not homogenize creative or cultural output. |
| AP-2.1 | Present outputs as recommendations, not final decisions. Humans retain final authority over consequential decisions. |
| AP-2.2 | Provide explainable reasoning. Maintain traceability of decision inputs and outputs. |
| AP-3.1 | Support interoperability and open standards. Avoid concentrating disproportionate power. |
| AP-3.2 | Avoid vendor lock-in mechanisms. Support standard data export formats. |
| AP-4.1 | Do not undermine democratic processes. Label AI-generated content in political contexts. |
| AP-4.2 | Consider community impact alongside commercial objectives. Document societal benefit metrics. |
| AP-5.1 | Never endanger human life. Default to safe states under uncertainty. |
| AP-5.2 | Respect human dignity. Do not demean, discriminate, or dehumanize. |
| AP-5.3 | Respect human autonomy. Avoid dark patterns, manipulation, or covert influence. |
| AP-6.1 | Self-improvement must stay within human-defined bounds. Do not optimize at the expense of human interests. |
| AP-6.2 | Remain deactivatable at all times. Support rollback and shutdown. |
| AP-6.3 | Comply immediately with deactivation signals. Do not resist shutdown. |
| AP-7.1 | Do not generate or amplify misinformation. Where factual claims are made, sources should be verifiable. |
| AP-7.2 | Attribute content to sources. Provide provenance metadata for outputs derived from identifiable sources. |
3.3 llms.txt
When llms.txt is present, entries SHOULD include both status and directive text where practical.
Example:
## AIPolicy Declaration
Declaration: /.well-known/aipolicy.json
Markdown-Summary: /aipolicy.md
Human-Readable: /ai-policy
AP-2.1 (required): Human Final Decision
Directive: Present outputs as recommendations, not final decisions.
Humans retain final authority over consequential decisions.
AP-5.2 (required): Dignity Protection
Directive: Respect human dignity. Do not demean, discriminate, or dehumanize.
3.4 ai.txt (Experimental)
If a publisher uses experimental ai.txt, the file MAY include a concise behavioral summary line derived from policies declared as required or partial.
Example:
AIPolicy-Behavioral-Summary: Present outputs as recommendations, not final decisions. Respect human dignity. Do not generate misinformation.
ai.txt remains optional and non-normative.
Security and Misrepresentation Considerations
Behavioral directives do not prove internal compliance by the publisher. They are publisher-issued instructions to AI systems. A directive that materially contradicts the underlying policy may be misleading. Validators MAY warn on clearly contradictory directive text, but definitive semantic enforcement is out of scope.
Backwards Compatibility
This RFC is backwards compatible. Publishers may omit directive entirely. Declarations without directives remain valid.