AI, web, and standards

A presentation at Web Standards for Policy Makers course by Hidde de Vries

AI, web, and standards Hidde de Vries, Web Standards for Policy Makers, 24 February 2026

(I’m here in personal capacity)

AI AI and web AI and web standards

AI

Students of AI often learn… linguistics, computer science, mathemathics, psychology, neuroscience, philosophy of language, philosophy of science

AI doesn’t mean the same to everyone, it’s an effective phrase to sell ideas

Natural Language Process ng i i i i i i branch of AI concerned with read ng, wr t ng and commun cat ng in human languages (eg speech recognition, machine translation)

Mach ne Learn ng i i i the process in which machines learn to class fy by themselves after being shown a large amount of labelled examples

Generat ve AI i large language models that were trained to generate text, media or code in response to prompts

“word calculators” “near human intelligence”

Agents task-oriented systems that can act on a human’s behalf

human toll data-taggers, people near data centres copyr ght ignored and violated major ethical concerns b as i i i i i accidental or purposefully susta nab l ty magnitudes more water, electricity used

AI and web

i i T he web s open for everyone to read and wr te

LLMs train on the open web

LLMs train on the open web

LLMs put a strain on the web matthiasott.com/articles/webspace-invaders / drewdevault.com

LLMs co-‘author’ the web

LLMs co-‘author’ the web

from: Alexandra Souly ea, Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples

i T he web s where many access LLMs

i T he web s where consequences of LLM usage are felt

Impl cat ons of AI deployment for web users (the good) Content may be more clear and relevant Easier to find things via more fuzzy search and structured result Very educated guesses at text alternatives i i More useful recommendations

Impl cat ons of AI deployment for web users (the bad) Possible increase of filter bubbles Decreased factuality of web content Security risks of AI agents acting on the web More phishing risk as it is less work i i Cost increase due to lower click through rates

AI and web standards

CEN/CENELEC Web and AI IG J C 21 + 5 working groups (many beyond web) Web Machine Learning WG + CG C Securing AI (SAI) IETF C2PA AI Preferences WG Agent to Agent BOF Content Credentials i T i T C 56 T ETSI CGs: AI Agent Protocol, Autonomous Agents on the Web, AI Knowledge Representation, Web AI for ime Series ECMA T T T T W3C C DA A Agent c AI Foundat on MCP AGEN S.MD

Web Machine Learning WG (W3C) Web Neural Network API github.com/webmachinelearning/webnn/blob/main/explainer.md

Web Machine Learning WG (W3C) Web Neural Network API dedicated low-level API for neural network inference hardware acceleration github.com/webmachinelearning/webnn/blob/main/explainer.md

Web Machine Learning WG (W3C) Web Neural Network API object detection, face detection, machine translation, noise suppression github.com/webmachinelearning/webnn/blob/main/explainer.md

Google’s Chrome Built-in AI eam + Members of the Web Machine Learning CG (W3C) Prompt API T github.com/webmachinelearning/prompt-api

Google’s Chrome Built-in AI eam + Members of the Web Machine Learning CG (W3C) Prompt API T github.com/webmachinelearning/prompt-api

request to summarise

request summarised content

request to summarise summarised content

ranslat on API i github.com/webmachinelearning/translation-api T T Google’s Chrome Built-in AI eam + Members of the Web Machine Learning CG (W3C)

AI Preferences WG (IE F) A Vocabulary For Express ng AI Usage Preferences i T ietf-wg-aipref.github.io/drafts/draft-ietf-aipref-vocab.html

D lemmas for spec authors Could this feature inadvertently affect privacy? Could this feature result in misinformation? How can we balance privacy and accessibility? Is it possible for this feature to be deterministic? i …

What sort of input may be missing? (re web standards on AI /web? i Q uest on for you

Questions and discussion hidde@hiddedevries.nl hidde.blog/slides

Hidde de Vries
@hdv

1 / 41

Resources

The following resources were mentioned during the presentation or are useful additional information.

UNESCO: Recommendation on the Ethics of Artificial Intelligence
Threat Model related to Agentic AI Web Browsers by Simone Onofri

See especially the ‘What can go wrong?’ section
AI in the browser threat model (by Tom Jones)
IETF draft: A Vocabulary For Expressing AI Usage Preferences

Abstract (from the draft): This document defines a vocabulary for expressing preferences regarding how digital assets are used by automated processing systems. This vocabulary allows for the declaration of restrictions or permissions for use of digital assets by such systems.
W3C CG Draft Report: Translator and Language Detector APIs

Abstract (from the draft): The translator and language detector APIs gives web pages the ability to translate text between languages, and detect the language of such text.
W3C CG Explainer: Translation API
W3C CG Explainer: Prompt API

(Note: a CG, Community Group, does not create standards, it is a less formal group with different IP protection that makes reports that could go on to become a standard in a Working Group)
W3C Group Draft Note: Ethical Principles for Web Machine Learning
W3C Candidate Recommendation: Web Neural Network API
W3C Neural Network API Explained
Announcement of Agentic AI Foundation