GitHub - dweekly/ai-content-disclosure: Explainer: AI Content Disclosure for HTML — element-level markup for AI authorship transparency

AI Content Disclosure for HTML

Authors

• David E. Weekly

Participate

• GitHub Issues
• WICG Proposals Issue #261

Table of Contents

• Introduction
• Problem / Motivation
• Goals
• Non-Goals
• Proposed Solution
• Key Scenarios
• Detailed Design
• What Counts as AI?
• Alternatives Considered
• Privacy and Security Considerations
• Accessibility Considerations
• Internationalization Considerations
• FAQ / Common Objections
• Stakeholder Feedback
• References

Introduction

Web pages increasingly contain text produced with varying degrees of AI involvement — from light AI-assisted editing to fully autonomous generation. There is currently no standard HTML mechanism for authors to disclose AI involvement at element-level granularity within a page.

This explainer proposes an ai-disclosure HTML attribute and a companion <meta name="ai-disclosure"> tag, enabling authors to declare the degree of AI involvement in any section of a web page.

Problem / Motivation

A modern news article page might contain a human-written investigation alongside an AI-generated summary sidebar and AI-moderated user comments. Today, there is no standard way to label these sections differently.

Existing approaches operate at coarser granularity:

• WHATWG HTML #9479 proposes a page-level <meta> tag with four values. It does not support marking individual elements. Commenters on that issue (42+) identified element-level granularity as the critical missing capability.

• IETF draft-abaris-aicdh-00 defines an AI-Disclosure HTTP response header. It applies to entire HTTP responses and cannot distinguish mixed content within a page.

• C2PA 2.2 provides cryptographic provenance for media files (images, video, audio). It does not support HTML text content and is designed for file-level, not element-level, assertions.

Regulatory Context

The EU AI Act Article 50 (effective August 2026) requires that AI-generated text content be "marked in a machine-readable format and detectable as artificially generated or manipulated." Major platforms (YouTube, Meta, TikTok) already require AI disclosure in their policies. A standard mechanism would serve both regulatory compliance and voluntary transparency.

Goals

1. Enable authors to declare AI involvement at element-level granularity in HTML
2. Provide a machine-readable signal usable by browsers, search engines, accessibility tools, and research crawlers
3. Align with existing vocabularies — specifically the IETF AI-Disclosure header modes and IPTC Digital Source Type taxonomy — for cross-standard consistency
4. Complement (not replace) HTTP-level and cryptographic provenance mechanisms

Non-Goals

• Detecting AI content (this is voluntary declaration, not detection)
• Replacing C2PA cryptographic provenance
• Preventing AI content creation or prescribing editorial policy
• Defining how browsers should render the disclosure (that is UA-specific)
• Covering non-text media (images, video, audio — handled by C2PA and IPTC)
• Solving metadata integrity (spoofing is possible with any self-declared metadata; verified provenance requires cryptographic systems like C2PA)

Proposed Solution

Page-Level Declaration (Meta Tag)

For pages with uniform AI involvement:

<meta name="ai-disclosure" content="none">
<!-- or "ai-assisted", "ai-generated", "autonomous", "mixed" -->

The value mixed signals that different sections have different levels; inspect element-level attributes for detail.

Element-Level Declaration (HTML Attribute)

An ai-disclosure global attribute on any HTML element:

<article>
  <section ai-disclosure="none">
    <h2>Six-Month Investigation: City Budget Shortfall</h2>
    <p>Our reporters spent six months reviewing financial records...</p>
  </section>

  <aside ai-disclosure="ai-generated" ai-model="gpt-4o" ai-provider="OpenAI">
    <h3>AI Summary</h3>
    <p>The investigation found a $4.2M discrepancy in the city's
    infrastructure fund, attributed to misclassified expenditures...</p>
  </aside>
</article>

Disclosure Values

Four values, aligned with the IETF AI-Disclosure header and IPTC Digital Source Type vocabulary:

Optional Metadata Attributes

Schema.org Integration (JSON-LD)

For search engine discoverability, the same information can be expressed as structured data. The simplest form is just the level string:

{
  "@context": "https://schema.org",
  "@type": "NewsArticle",
  "headline": "Investigation: City Budget Shortfall",
  "author": { "@type": "Person", "name": "Jane Doe" },
  "aiDisclosure": "ai-assisted"
}

An expanded form supports optional metadata. All fields except level are strictly optional — publishers may have legitimate reasons not to disclose specific tools or providers:

{
  "@context": "https://schema.org",
  "@type": "NewsArticle",
  "headline": "Investigation: City Budget Shortfall",
  "author": { "@type": "Person", "name": "Jane Doe" },
  "aiDisclosure": {
    "level": "ai-assisted",
    "tool": "Claude 3.5 Sonnet",
    "provider": "Anthropic",
    "description": "AI used for copy editing and fact-checking assistance",
    "methodologyUrl": "https://example.com/ai-methodology"
  }
}

(Proposed as a comment on schemaorg/schemaorg#3391.)

Key Scenarios

Scenario 1: News Article with AI Summary Sidebar

A newsroom publishes an investigative piece with a human-written article and an AI-generated summary:

<meta name="ai-disclosure" content="mixed">
<!-- ... -->
<article ai-disclosure="none">
  <h1>Exclusive: City Budget Shortfall</h1>
  <p>After six months of records review...</p>
</article>

<aside ai-disclosure="ai-generated" ai-model="gpt-4o" ai-provider="OpenAI">
  <h3>Key Takeaways (AI-Generated)</h3>
  <ul><li>$4.2M discrepancy found...</li></ul>
</aside>

Scenario 2: Blog Post Written with AI Editing

A blogger writes a post and uses an LLM for grammar, style, and clarity improvements:

<meta name="ai-disclosure" content="ai-assisted">
<!-- ... -->
<article ai-disclosure="ai-assisted" ai-model="claude-3.5-sonnet"
         ai-provider="Anthropic">
  <h1>My Trip to Kyoto</h1>
  <p>The bamboo grove felt otherworldly at dawn...</p>
</article>

Scenario 3: Fully Automated Content Feed

An automated system generates weather reports without per-instance human oversight:

<meta name="ai-disclosure" content="autonomous">
<!-- ... -->
<div ai-disclosure="autonomous" ai-model="weather-llm-v2"
     ai-provider="WeatherCorp">
  <h2>San Francisco Bay Area Forecast</h2>
  <p>Expect fog clearing by noon with highs near 62°F...</p>
</div>

Scenario 4: Human-Only Publication Asserting Provenance

A literary journal positively asserts that no AI was used:

<meta name="ai-disclosure" content="none">
<!-- ... -->
<article ai-disclosure="none">
  <h1>The Weight of Feathers</h1>
  <p>She found the letter tucked inside a volume of Neruda...</p>
</article>

Note: ai-disclosure="none" is a positive assertion. The absence of the attribute means "unknown," not "none."

Detailed Design

Inheritance

• Element-level ai-disclosure attributes override the page-level <meta name="ai-disclosure"> tag
• Children inherit their nearest ancestor's ai-disclosure value unless they specify their own
• Absence of the attribute means unknown (not "none") — this is an important distinction that avoids false negative assertions

Relationship to HTTP AI-Disclosure Header

These are complementary. A CDN or reverse proxy can set the HTTP header; a CMS can set the meta tag; an author or AI tool can set element-level attributes. None supersedes the others.

Vocabulary Alignment with IPTC

What Counts as AI?

To address a frequently raised concern ("where do you draw the line with grammar checkers?"), here is boundary guidance:

Not AI (no disclosure needed):

• Deterministic spell-check (dictionary lookup)
• Thesaurus / synonym suggestions
• Search engines used for research
• Calculators and formula solvers
• Template-based mail merge

ai-assisted:

• LLM-based grammar and style tools (Grammarly AI, etc.)
• AI summarization of human-written text
• AI-powered translation (DeepL, Google Translate neural mode)
• AI rewriting of human-provided drafts

ai-generated:

• LLM-generated drafts from human prompts
• AI expanding bullet points into prose
• Chatbot-generated responses with human review before publication

autonomous:

• Agent-generated content without per-instance human prompting
• Automated reporting systems (weather, sports scores, financial summaries)
• Content produced by persistent AI agents with standing instructions

The boundary is generative/inferential AI — systems trained on data that produce novel outputs. Deterministic tools that apply fixed rules are not covered.

Alternatives Considered

1. Page-Level Meta Tag Only (WHATWG #9479)

Does not handle mixed-content pages — the single most requested feature in that issue's 42+ comments.

2. HTTP Header Only (IETF AI-Disclosure)

No element-level granularity. Not accessible to client-side tools processing the DOM. Cannot distinguish mixed content within a page.

3. C2PA Manifest for HTML

C2PA is file-based cryptographic provenance. HTML pages are dynamically assembled from templates, databases, and user input — they are not single files with stable hashes. C2PA and this proposal are complementary, not competing.

4. data-* Attributes Only

Using data-ai-disclosure instead of a dedicated attribute was considered. Trade-off: data-* attributes have no semantic meaning to browsers or assistive technology. A dedicated attribute signals intent for future browser integration (e.g., address bar indicators, accessibility announcements).

5. RDFa / Microdata Only

Too verbose for common cases. RDFa requires namespace declarations and multi-attribute markup for simple assertions. The ai-disclosure attribute provides a lightweight default; RDFa or JSON-LD can supplement it for richer structured data needs.

Privacy and Security Considerations

Responses to the W3C Security and Privacy Self-Review Questionnaire:

• No new information about the user (reader) is exposed. The attribute describes the content, not the person viewing it.
• No new fingerprinting surface. The attribute is author-declared metadata, not a browser API.
• Disclosure is voluntary and advisory. It carries no integrity protection. For verified provenance, pair with C2PA.
• Model/provider metadata is optional to avoid requiring disclosure of trade secrets or proprietary tooling.
• Prompt text is not embedded in HTML. The optional ai-prompt-url attribute links to an external resource, giving authors control over what they disclose and when they revoke access.
• The voluntary nature means it cannot be relied upon for security decisions, the same as any self-declared metadata (robots.txt, rel=nofollow, Schema.org markup).

Accessibility Considerations

• Browsers and screen readers could optionally announce AI disclosure to users (e.g., "AI-generated content follows")
• The ai-disclosure attribute should be exposed via the Accessibility Object Model
• No change to existing content rendering — the attribute is purely informational
• Useful for accessibility research: understanding how AI-generated content affects comprehension for users with cognitive disabilities

Internationalization Considerations

• Attribute values (none, ai-assisted, ai-generated, autonomous) are English-language tokens intended for machine consumption, not display
• Human-readable presentation of disclosure status is a User Agent responsibility and can be localized
• The ai-prompt-url attribute can link to localized methodology pages

FAQ / Common Objections

"This is the evil bit — bad actors won't comply."

This standard serves responsible publishers, regulated industries, and AI tool vendors who want to be transparent. It is not a detection mechanism. The EU AI Act makes compliance mandatory for covered entities, and major platforms already require disclosure in their terms of service.

The analogy is rel=nofollow: voluntary, widely adopted because it aligns incentives, and useful despite being ignorable by bad actors.

"Where do you draw the line with grammar checkers?"

See What Counts as AI? above. The boundary is generative/inferential AI — systems trained on data that produce novel outputs. Deterministic spell-check and thesaurus tools are excluded.

"Metadata will be gamed like SEO dates."

True for any self-declared metadata. The standard enables honest disclosure; verification requires pairing with C2PA or regulatory auditing. The value is in the signal for those who choose to use it honestly — same as Schema.org structured data, which search engines use despite its spoofability.

"This will stigmatize AI-assisted content."

The granular levels (four values, not binary) allow publishers to distinguish "AI helped me edit" from "AI wrote everything." The ai-assisted level should carry no more stigma than acknowledging the use of a human copy editor.

"Everything will be AI-touched soon, making this meaningless."

That is exactly why granularity matters. Binary "AI/not-AI" is already inadequate. The spectrum from none to autonomous reflects the reality of modern content workflows and remains meaningful as AI tools become ubiquitous.

Stakeholder Feedback

References

1. WHATWG HTML #9479: Proposal: Meta Tag for AI Generated Content
2. IETF draft-abaris-aicdh-00: AI Content Disclosure Header
3. WICG #104: Content Provenance and Authenticity
4. WICG #141: Advancing Web Metadata
5. W3C: AI & the Web — Understanding and managing the impact of Machine Learning models on the Web
6. W3C Explainer Explainer
7. C2PA Specification 2.2
8. IPTC Digital Source Type Vocabulary
9. EU AI Act — Regulation (EU) 2024/1689, Article 50
10. EU Code of Practice on Transparency of AI-Generated Content (draft)
11. Bloomberg Container Timing Explainer (structural model for this proposal)