Web Standards for AI:

Architecture of Human-Machine Symbiosis

​

​

​

Paola Di Maio, PhD

W3C AI KR CG @TPAC KOBE JAPAN BREAKOUT SESSION

Speculating on standards that preserve web integrity while enabling authentic socio technical co-evolution

What's Happening Now

AI is here. We might as well make the most of it.

It's delivering incredible new capabilities which were unthinkable until recently—capabilities we still don't fully understand.

The web is here, hopefully to stay. Open Web always been under threat, since the beginning, But we are still here

​

How will the web change in the age of AI?

What's Happening Now… Cont d

Corporations and Institutions RULE THE WORLD

Hire the best, pay high, make them work hard,

AI is soon going to be ruling the world

Influence Governments, Policies, Economies

AI companies EMBRACE TO SOME EXTENT THE vision for the open web *so called evangelists?

But they operate on ruthless market logic, open web is whitewashing

corporations - AI

W3C's Role in Shaping Human-AI Co-Evolution

W3C creates UNIQUE opportunities for humans to interact with technology

Gives a voice to stakeholders NOT CORPORATIONS, NOT INSTITUTIONS

Encourages diverse participation, gives voice to academics, philosophers, CITIZENS users at large in the technology agenda

THANK YOU W3C for the opportunity to be part of the process

an one-sided optimization

cPARTICIPATION IS IMPORTANTCo-Evolution

​

PARTICIPATION IS IMPORTANT BECAUSE

we learn what is going on

learn how to interact with diverse stakeholders

we contribute to shaping our world

we learn how to take fire and handle difficulties

we become resilient and develop staying power

SOMETHING BEAUTIFUL IS HAPPENING

an one-sided optimization

One of The Problems with AI (besides Misrepresenetation)

Asymmetry

We don't know what AI does with our information

Why?

• There are no standards
• There are no simple enforcement mechanisms (other than wars)

What We're Working On

W3C AI KR Knowledge Representation Community Group

• We keep track of AI developments using KR, The only way to make sense of AI complexity is use KR
• Modeling the we standard development process *incl AI STANDARDS
• Applying joint optimization *see separate paper
• Developing new web standards and tools for web/AI
• Developing AI agents to streamline the process of creating web standards

see PROOF OF CONCEPT DEMO

The Landscape of AI Standards is Messy SEE SEPARATE TALKs and papers

NOT FIT FOR PURPOSE

Does not cover or minimize the risks

What Are Web Standards?

A collection of technical specifications organized by the organizations that develop them, primarily the World Wide Web Consortium (W3C).

These standards can be broadly categorized by their function in building the web.

They provide the foundation for interoperability, accessibility, and innovation on the web.

Key Categories of Web Standards

Foundational Markup & Styling

HTML, CSS, SVG — structure and presentation of web content

Interactivity & Programming

DOM, APIs, WebRTC — dynamic behavior and complex applications

Accessibility & Internationalization

WCAG, i18n standards — making the web available to all people

Data, Metadata & Semantics

XML, JSON, SKOS — how data is structured, shared, and related

What We Need: Web AI Standards

Automating the process and service

While

Preserving interactivity, interaction, and writeability

We Need Process automation for Web AI Standards

But hav limited time and resources to figure them out because, they are complex, messy

​

Can AI help?

SEE DEMO

The Question

What would meaningful web standards for AI look like?

Unlike previous web standards focused on document interchange, AI standards must address the interface between human cognition and machine intelligence across billions of context-rich interactions.

We're transforming the web's fundamental nature.

How AI Integrates with the Web Today

Extraction Mode

Training by scraping the web, treating it as corpus rather than network. Breaks linkability and attribution.

Oracle Mode

AI systems answering queries without web navigation. Threatens decentralization.

Agent Mode

AI browsing and interacting on behalf of users. Could enhance or undermine universal access.

Embedded Mode

AI capabilities built into web applications and browsers. Risks opaque, proprietary intelligence layers.

Web Principles at Risk

• Universal Access — Anyone can publish, anyone can access
• Decentralization — No single point of control or failure
• Interoperability — Systems work together through open standards
• Separation of Concerns — Content, presentation, behavior are distinct
• User Agency — Users control their experience and data
• Openness and Transparency — View source, understandable technology

Foundational Layer

Interaction Protocols

Conversational State Management

How do AI systems maintain context across sessions, devices, and time? What's the standardized format for conversational memory?

Multi-modal Interaction Schemas

Voice, text, gesture, physiological signals. How do these modalities interoperate and preserve meaning across modal boundaries?

Turn-taking and Interruption Protocols

Natural conversation requires sophisticated management of who speaks when and how interruptions are handled.

Semantic Layer

Knowledge Representation

Knowledge Graphs for AI Context

Not just RDF triples, but richer structures representing uncertainty, temporal dynamics, source attribution, and confidence levels.

Grounding and Reference Resolution

Standardized mechanisms for resolving references across systems. Persistent, globally unique identifiers for conversational artifacts and shared understanding.

Ontological Interoperability

Translation layers, shared upper ontologies, and mechanisms for semantic alignment without requiring identical representations.

Evaluation and Trust Layer

Most Critical, Most Neglected

Capability Declaration and Discovery

AI systems should declare what they can and cannot do in machine-readable formats. Verifiable capability assertions, not marketing claims.

Explainability Interfaces

Standardized ways for AI systems to expose reasoning processes, confidence levels, and decision boundaries.

Safety and Alignment Attestation

Mechanisms for systems to declare training objectives, alignment methods, and safety constraints that others can verify.

Agency and Delegation Layer

As AI Systems Become Agentic

Authorization and Capability Delegation

How do users grant AI systems permission to act on their behalf? What's the OAuth equivalent for AI agency? Fine-grained, revocable permissions across platforms.

Inter-AI Protocols

When your AI assistant interacts with mine, what protocols govern that interaction? How do we prevent adversarial behavior while enabling cooperation?

Action Schema and Workflow Representation

Standardized ways to represent complex, multi-step tasks spanning multiple AI systems, human interactions, and external services.

Ethical and Rights Layer

Philosophically Deep Territory

Data Provenance and Lineage

Every piece of information an AI system uses should be traceable to its source through standardized metadata. Enables copyright compliance, bias detection, and accountability.

User Rights and Data Sovereignty

Standards that encode the right to be forgotten, right to explanation, right to human review, and right to port AI interaction history and preferences between platforms.

Consent and Preference Management

Sophisticated frameworks for expressing and enforcing user preferences about data use, interaction styles, privacy boundaries, and acceptable AI behaviors.

Where Web Integrity is Breaking Down

Attribution Collapse

The web's citation mechanism (hyperlinks) replaced by AI synthesis without clear source attribution. Breaks the epistemic fabric of the web.

Death of Serendipity

AI that "efficiently" answers queries without requiring web navigation eliminates aimless browsing and unexpected discovery.

Context Collapse

The web preserved context through site structure and page relationships. AI extraction flattens this into decontextualized synthesis.

Enclosure of the Commons

AI training represents massive enclosure—taking freely shared web knowledge and locking it inside proprietary models.

Preserving Web Integrity Through AI

Mandatory Attribution Standards

AI systems cite sources with hyperlinks at the claim level, maintaining the web's connective tissue while adding synthesis capabilities.

Federated AI Infrastructure

Distributed models running locally or on community infrastructure. Preserves decentralization through open protocols and fine-tunable models.

Transparent Reasoning Trails

AI systems expose reasoning through standardized interfaces—sources, confidence levels, alternative interpretations considered.

User-Owned Context

Conversational history and preferences stored in standardized formats that users own and can port between providers.

Concrete Standards We Need

AI Attribution Markup Language

Conversational Context Protocol

AI Capability Declaration Format

Provenance Chain Standard

Standard way to embed source citations, confidence levels, and reasoning chains in AI-generated content

Open standard for representing conversational state, memory, and preferences (the RSS for AI conversations)

Machine-readable declarations of what AI systems can do, how they were trained, and their limitations

Every piece of AI-generated content carries verifiable chains showing sources, transformations, and editorial control

The Path Forward

We're not adding AI features to the web—we're transforming its fundamental nature.

Without intentional action, we'll have an AI-mediated internet that has lost what made the web revolutionary: openness, decentralization, and preserved attribution.

Standards work must happen now, while patterns are still forming.

The question is whether AI-web integration will preserve the web's liberatory principles or create new centralized control.

NOW

1. Modelling the web standard development process
2. The Challenge: Making the Demo Run

https://colab.research.google.com/drive/1m1Ea-I3Fdx4h1p3uVD04sSn2Ut4pXZie?usp=sharing

NEEDS TO BE DEBUGGED

BUT WAIT

​

ME TO GEMINI, MINUTES BEFORE THE BREAKOUT SESSION

GEMINI APPPLIED THE PATCH

THANK YOU FOR LISTENING

​

Talk to me

via AI KR CG

or W3C SLACK or email me *