First

Some disclaimers…

Artificial Intelligence

• Misinformation have always existed
• Code algorithms have hardcoded biaised since the Internet was created
• AI is learning from all of that
• You can't trust AI
• … but you can't trust the web

Ethical Principles for ML

Ethical Principles for Web Machine Learning
W3C Group Draft Note, 8 January 2024

• Documents ethical issues associated with using Machine Learning on the Web
• general consideration of harms, risks and mitigations relevant to Web ML
• Accuracy: deviation from a true value can affect life, including credit scoring, loan approval
• Bias: systematic deviation can disproportionately affect individuals or groups
• Privacy: operating without a user’s knowledge/consent, scraping personal information to train models

For more on AI, see AI and Web impact (last updated: August 2024)

Fast-pacing systems

• Agent protocols are evolving fast and competing
• We're still climbing the hype curve
• Regulators are trying not to lag
• We're all learning

So, why AI?

As users and developers, our modern world is more complicated.

• Convenience: we're lazy
• Ease of use: we want to get stuff done
• Cost: we're cheap

Why AI and the web?

• The web contains a LOT of information and services, including new information and services every day
• AI needs to leverage the web to be useful
• Keep in mind that a majority portion of the Web is not accessible through search engines

AI and client application

AIs need to:

• Help the application: image, audio, text, sensors, data
• Help the user: Get me a pizza, Go to a music concert and tell my friends about it, …

What do we need?

• Connect the applications to AIs
• Connect the AIs with ALL applications (native, web, miniapps, etc.): calendars, social platforms, wallets, online shops, etc.
• Connect the user to the AIs

Connecting applications with AIs

Applications need to leverage LLMS, and run new LLMs

• APIs to leverage Client side, preinstalled, LLMs
• High-performance API to load and run an LLM

Neural Network API

Web Neural Network API

• Hardware-agnostic abstraction layer for NN inference
• Use cases: Person Detection, Style Transfer, Image Captioning, Detecting fake video
• Sync/Async build and execution, device selection (cpu/gpu), power preference
• Operands: sigmoid, softmax, slice, gru, hardSwish, squeeze, etc.
• Issue #350 got resolved: Apple's CoreML can distribute a workload within a single ML graph across multiple devices including the Apple Neural Engine Apple

built-in AI APIs

• Experimentation in the Web Machine Learning CG
• Ship an LLM into the browser for edge computing
• Access LLMs capabilities primarily with API: translator, summarizer, writer proofreader, prompt (for web extensions)
• See also Built-In AI APIs from AC Meeting 2025

Connecting AIs with applications

AIs needs to understand the application intent, and interact with the application:

• using AI:

  markdown, screenshots, and code analysis

• using developer hints: AI agents struggle to navigate existing human-first interfaces

  declarative (metadata) and programmatic (APIs)

Using developer hints

• Native applications: see agent to agent protocols (MPC, etc.), accessibility APIs, etc. For Web browsers, see WebDriver.
• Web applications:
• declarative:
  • Accessible Rich Internet Applications (WAI-ARIA)
  • see LLM & Linked Data (AC Meeting 2025)
• API: see WebMCP (app-controlled UI)
• Server applications: see agent to agent protocols (MPC, etc.)

W3C Efforts

• Working Groups:
  • Web Machine Learning
• Community Groups:
  • AI Agent Protocol
  • Web Machine Learning
  • AI KR (Artificial Intelligence Knowledge Representation)
  • Web Incubator
  • Cognitive AI
  • Human Centric AI
  • Web AI for Time Series
  • Autonomous Agents on the Web

Work in progress

• Interest Groups:
  • Web and AI
• Workshops:
  • “AI Agents & the Web”

Thank you