This directory complements the main article. Follow the internal reading links or copy a source URL from the footer into your browser. The bibliography uses the same source numbers as the article, so a citation can be traced across the site.
AI learning paths
-
Start with the concepts
Read the definition, compare AI with machine learning, and follow the support-routing example. OECD supplies a broad framework; Google’s learning materials explain the mechanics.
Read the foundations → -
Explore generative AI
Distinguish a language model from a chatbot. Read the Transformer and RAG papers for original contributions, with the introductory article as a guide to terminology.
Explore models and retrieval → -
Understand agents
Anthropic’s workflow-versus-agent distinction is a useful practitioner framework. Evaluate the system’s actual tools and permissions rather than relying on its label.
See the system components → -
Read the numbers carefully
Use Census and Eurostat for business surveys, Pew for public-use research, and Stanford for a broader compilation. Keep the population and question attached to each percentage.
Browse the statistics → -
Study work and productivity
Read the ILO exposure framework alongside the published support-agent study. One estimates possible task exposure; the other measures outcomes in one workplace setting.
Separate exposure from outcomes → -
Investigate limitations
NIST’s framework and generative-AI profile organize risk questions. The article offers concise examples, not a substitute for evaluating a particular deployment.
Read about limitations →
How to read an AI statistic
| Evidence type | Question to ask | Example in this reference |
|---|---|---|
| Survey | Who was asked, what was the wording and when was the fieldwork? | Pew’s February 2026 measure concerns U.S. adults and named chatbot use. |
| Business estimate | Which firms, sectors and time period are covered? | Census and Eurostat do not sample identical business populations. |
| Field study | What task, sample and outcome were measured? | Issues resolved per hour among 5,172 support agents. |
| Exposure model | Does this estimate task susceptibility or observed employment change? | ILO/NASK’s exposure estimate is not a count of jobs lost. |
| Forecast | What year, scenario and system boundary does it cover? | IEA’s 2030 value is a projection for all data centres. |
| Publication count | What is counted, and what can publication delay conceal? | WIPO counts published patent families, not new products. |
Percentages from unrelated populations cannot be averaged into a meaningful universal rate. An observation made in 2025 remains a 2025 observation even when cited in a 2026 article. If a source revises its methods or questions, the apparent trend may partly reflect that change.
Source selection and editorial method
This edition prioritizes original papers, statistical agencies, research institutions and the organizations responsible for the systems being described. Vendor materials are useful for explaining their own frameworks and products; they are not treated as independent endorsements. Short examples and diagrams are original teaching material, clearly separated from measured study results.
The main page groups related concepts under explanatory headings: training and inference, machine learning and deep learning, language models and retrieval, then applications and limitations. The glossary links back to those explanations. Statistical observations retain their dates and qualifications instead of being presented as live counters.
The source-check date records this edition’s research review. This site does not automatically refresh its statistics. Later revisions should record the original source, changed observation period and any methodological changes before updating the page.