AI Content Disclosure for HTML

Motivation

Web pages increasingly contain text produced with varying degrees of AI involvement — from light AI-assisted editing to fully autonomous generation. A modern news article page might contain a human-written investigation alongside an AI-generated summary sidebar and AI-moderated user comments. There is currently no standard HTML mechanism for authors to disclose AI involvement at element-level granularity within a page.

Related approaches address different parts of this problem:

• WHATWG HTML #9479 proposes a page-level <meta> tag with four values. It does not support marking individual elements. Commenters on that issue identified element-level granularity as the critical missing capability.
• C2PA provides signed provenance claims bound to digital assets, including support for text and regions within assets. This report instead defines lightweight author declarations and inheritance through the HTML element tree.

Regulatory context

The EU AI Act Article 50 distinguishes providers' obligations to mark AI-generated outputs in a machine-readable format and enable their detection (paragraph 2) from deployers' obligations to disclose AI-generated or manipulated text published to inform the public on matters of public interest (paragraph 4). These obligations have different exceptions.

Article 50(4) provides an exception to the text-disclosure obligation where content undergoes human review or editorial control and a natural or legal person holds editorial responsibility for its publication. The article itself does not specify reviewer competence. The Commission's interpretation, explained in its Article 50 guidance FAQ, describes substantive examination by people with relevant subject-matter knowledge and professional judgment. It excludes superficial checks such as spelling or grammatical correction. These are conditions for an exception to disclosure, not a general requirement that all AI-generated text receive expert review.

The ai-assisted value asserts that human review occurred before publication. It does not establish the reviewer's competence, the substance of the review, or editorial responsibility, and does not by itself establish eligibility for the Article 50(4) exception. This specification describes AI involvement and human review; its use alone does not establish regulatory compliance.

Where reader-facing disclosure is required, Article 50(5) calls for clear, distinguishable and accessible information by the first interaction or exposure. HTML disclosure metadata can support that presentation, but does not itself provide a reader-facing notice. This specification does not prescribe a rendering or labeling design.

Article 50 applies from 2 August 2026. As explained in the Commission guidance linked above, the transition to 2 December 2026 applies only to the paragraph 2 marking and detection obligation for AI systems placed on the market before 2 August 2026; it does not postpone deployers' paragraph 4 disclosure obligation.

Further context is provided by the Commission's Guidelines on transparency obligations for providers and deployers of certain AI systems and the Code of Practice on Transparency of AI-generated Content. The voluntary Code addresses provider marking and detection as well as deployer labelling. This report proposes an HTML mechanism for author-declared text metadata that can contribute to disclosure workflows; it does not claim that this format is prescribed by the Code or that unverified HTML declarations alone satisfy its measures.

Design 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. Make classification depend on AI involvement and human review, using the two questions in the classification procedure.
4. Relate to existing vocabularies, including the IPTC Digital Source Type taxonomy, while identifying differences in scope and meaning.
5. Complement (not replace) HTTP-level and cryptographic provenance mechanisms.

Scope and non-goals

This report defines author-declared metadata for textual content in HTML, including text alternatives, and rules for interpreting and inheriting that metadata. Its purpose is to communicate AI involvement and whether human review occurred. The following limits apply to this work:

• Detection and verification: the report does not detect AI use, verify the truth of declarations, or certify provenance. It does not define or verify cryptographic proofs of authorship or review, including proofs that either was performed by a human or a non-human system. Its metadata is self-declared.
• Content policy and quality: the report does not judge content quality, prescribe whether AI should be used, or determine how platforms should act on disclosure signals.
• Legal and professional requirements: the report does not determine legal compliance, establish editorial responsibility, or define "substantive review." Its definition of human review does not prescribe review depth or reviewer qualifications. Any additional requirements remain with the relevant authorities and frameworks; see Regulatory context.
• Reviewer identity and sign-off: the report does not identify or authenticate the person who reviewed a text, record approval to publish, or identify the person or organization accountable for publication. An ai-assisted declaration asserts that human review occurred; it does not attest to a particular person's sign-off or acceptance of editorial responsibility. Such records can be maintained separately.
• Additional provenance detail: media provenance, quantitative percentages of AI authorship, and registries of models or providers are outside this report's scope. Optional metadata does not create a registry or require a particular methodology-document format.
• Other communities' specifications: the report does not define HTTP disclosure headers, a JSON-LD disclosure vocabulary, Schema.org properties, or extensions to IPTC or C2PA. Examples and correspondences support coordination with those communities.
• Standardization and implementation: publication as a Community Group Report does not establish a W3C Standard, adoption into HTML, or browser support. See Implementation status and adoption for the proposed next steps.

Mixed-content news page

A newsroom publishes an investigative piece with a human-written article and an AI-written summary reviewed by a human before publication. Each region declares its own disclosure value, so no page-level default is needed. Elements outside these regions have an "unknown" disclosure status:

<article ai-disclosure="human-only">
  <h1>Exclusive: City Budget Shortfall</h1>
  <p>After six months of records review...</p>
</article>

<aside ai-disclosure="ai-assisted"
       ai-model="example-model-v2" ai-provider="Example Provider">
  <h3>Key Takeaways (AI-Generated)</h3>
  <ul><li>$4.2M discrepancy found...</li></ul>
</aside>
        

AI-edited blog post

A blogger writes a post and uses an LLM for grammar, style, and clarity improvements, then reviews the edited text before publication:

<meta name="ai-disclosure" content="ai-assisted">

<article ai-disclosure="ai-assisted"
         ai-model="example-model-v1" ai-provider="Example Provider">
  <h1>My Trip to Kyoto</h1>
  <p>The bamboo grove felt otherworldly at dawn...</p>
</article>
        

Automated weather feed

An automated system generates and publishes weather reports without human review:

<meta name="ai-disclosure" content="ai-autonomous">

<div ai-disclosure="ai-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>
        

Human-only assertion

A literary journal positively asserts that no AI was used:

<meta name="ai-disclosure" content="human-only">

<article ai-disclosure="human-only">
  <h1>The Weight of Feathers</h1>
  <p>She found the letter tucked inside a volume of Neruda...</p>
</article>
        

Note: ai-disclosure="human-only" is a positive assertion. Without an applicable element, ancestor, or page-level declaration, the value is "unknown," not "human-only."

Applying the classification procedure

Human authorship of notes, a prompt, or an earlier draft does not establish review of the resulting AI-generated or AI-edited text. Review does not require making edits. Translation follows the same review rule as other AI-involved text. See Boundary guidance for exclusions, including deterministic tools and AI-generated page structure that does not generate or modify the text.

If the production history is unknown, do not guess a disclosure value. Omitting a declaration makes no new assertion; the applicable value is then determined by Inheritance, and is "unknown" if no ancestor or page-level default supplies one.

This guide covers a single piece of content. When one page combines content with different disclosure values (for example, a human-written article with an AI-generated summary), disclose each part individually. A page-level default is optional; if one is supplied, use a valid disclosure value and override it on elements where a different value applies.

Scenarios

These scenarios illustrate the disclosure definitions; they do not establish legal compliance or eligibility for an exception to a disclosure obligation. See Regulatory context for how the human-review assertion relates to EU AI Act Article 50.

IPTC Digital Source Type

The IPTC Digital Source Type vocabulary describes how media was produced. The following examples show related terms for text; they are not a lossless mapping or an IPTC endorsement. Source type alone does not establish whether human review occurred, so consumers cannot derive this specification's disclosure values from an IPTC term alone.

IPTC's digitalCapture describes capture from real life by a digital camera or recording device, rather than human authorship of text.

Schema.org / JSON-LD

JSON-LD is a useful complementary approach to AI disclosure, particularly for publishers already using structured data. HTML attributes associate disclosure directly with elements in the document; JSON-LD can describe the corresponding work and richer metadata such as contributors, methodology, and evidence links. Authors can use both together, keeping declarations about the same content consistent.

The group's charter covers coordination with related standards, but defining a JSON-LD disclosure vocabulary or extending Schema.org is outside the current scope of this specification. Those designs belong with the respective communities. Related discussion is tracked in schemaorg/schemaorg#3391 and issue #17.

The following examples share possible designs for discussion. The aiDisclosure property and its expanded fields are illustrative; these examples do not define standardized Schema.org terms or imply support by search engines or other consumers. JSON-LD is not required for conformance to this specification.

A simple form could associate a disclosure value with a work:

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

An expanded form could carry additional metadata:

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

PROV-O and Dublin Core

W3C PROV-O describes provenance through entities, activities, agents, and their relationships. The Dublin Core DCMI Metadata Terms include properties for creators, contributors, and provenance. These vocabularies offer richer attribution and provenance expression that can complement the disclosure values defined here.

Integrating these vocabularies is outside the current scope of this specification. No mappings to PROV-O or Dublin Core, or additional HTML attributes derived from them, are defined at this time. They are linked as related work for authors who need more expressive metadata alongside AI disclosure.

C2PA

C2PA binds signed provenance claims to digital assets. Its technical specification includes unstructured text and regions of interest within assets; it is not limited to whole image, audio, or video files. Cryptographic validation establishes the integrity and signer of a claim, not by itself the truth of a claim about human authorship or review.

This report defines advisory HTML declarations and DOM inheritance, without cryptographic binding. The approaches can complement each other, but no C2PA integration or automatic mapping to this report's disclosure values is defined.

Text watermarking

Text watermarking embeds a detectable statistical signal in generated text. Research such as TextSeal explores detecting and locating watermarked passages, including within documents containing both human-written and AI-generated text.

As of September 2026, Anthropic is rolling out text watermarking in supported Claude models. Its published explanation describes a method based on SynthID-Text, with an Anthropic-specific key and a detection API in private preview. The signal is embedded through word selection, rather than HTML attributes or hidden characters.

Watermarking can provide evidence of AI involvement that complements voluntary, author-declared disclosure. A detected watermark does not establish whether a human reviewed the content, so it cannot by itself distinguish ai-assisted from ai-autonomous. Failure to detect a watermark does not establish human-only provenance. This specification does not define watermark generation, detection, or automatic conversion of detection results into disclosure values.

IBM AI Attribution Toolkit

The IBM AI Attribution Toolkit is a research prototype for structured AI attribution statements via human self-reporting. It uses a multi-axis model (proportion, contribution type, initiative, human review) compared to this specification's single-axis model. The approaches are complementary: the toolkit addresses the "author guidance" layer, while this specification addresses the "machine-readable metadata" layer.