Learn AI / Methodology
How these pages were written
This site tells you to ask where an AI answer came from. It would be a poor advertisement for that advice if it did not answer the question about itself.
The short version
Each term was drafted by a large language model running through a script on a developer's machine, against a fixed prompt and a strict schema. The draft was checked automatically for structure, for required fields, for valid links to other terms, and for reading level. Every external URL was fetched to see whether it answers. A second, different model then reviewed the first model's work, looking for what was wrong with it.
Being precise about what that does and does not mean. A model checking a model is a real check and a limited one. It catches arithmetic that does not compute, definitions that fail to separate themselves from neighbouring terms, quiz answers that are not defensible, and examples that present invented specifics as real. It does not reliably catch a confident, fluent, wrong account of how something works. So: use these pages to get oriented, and check anything that carries weight against the primary sources every page links to.
The published site is a folder of static files on GitHub Pages. It makes no call to any AI service, holds no credential, and could not reach one if it tried.
Where the corpus stands
The pipeline, step by step
- A taxonomy file defines the vocabulary. 307 terms across 16 categories, each with a difficulty, an audience, its prerequisites and the terms it is commonly confused with. Written by hand.
- Resource links are assembled before generation, never by the model. They come from a hand-authored list, from the Wikipedia API (which either returns a real article or does not), or from a constructed search URL. The model is shown the links and writes only the sentence explaining why each is useful. It is never asked for a URL, because a model asked for a URL will eventually invent one.
- One term at a time goes to Azure OpenAI (gpt-5.4-2026-03-05) with a fixed system prompt and a JSON Schema the response must match.
- The answer is validated against the same schema the site uses. A draft missing its analogy limitation, its screen-reader description or a defensible quiz answer is rejected and regenerated with the specific failures fed back to the model. Rejected attempts are kept on disk.
- Every URL is fetched. Links that answer are marked as checked with a date. Links that do not are labelled "not verified" on the page rather than quietly presented as vetted.
- A quality report is generated: reading-level outliers, duplicate definitions, orphan terms, prerequisites pointing at unwritten pages, missing UN examples, broken links.
- A second, different model reviews the first one's work. Currently gpt-5.6-sol, prompted to find errors rather than to approve: arithmetic that does not compute, definitions that fail to distinguish themselves from neighbouring terms, quiz answers that are not defensible, anthropomorphism, and examples that present invented specifics as real. It may rewrite a bounded set of fields, never introduce a new fact or URL, and every rewrite is re-validated against the same schema before it is kept. 263 pages were rewritten as a result.
- A United Nations voice pass rewrites the examples. The first draft read as though written about the Organization rather than from inside it: "a UN team", "a UN office". This pass gives each term a specific person doing a specific job against a real deadline, rotated so that no two of the 307 pages share a scenario. It changes the setting and the register only; the definition and the plain explanation are out of its reach, because those are the part someone with no background reads. 307 pages have been through it. It may cite only document symbols that appear in this site's sourced timeline corpus, and anything else symbol-shaped is refused.
- Publishing is a separate, deliberate step. The site renders one directory, and a term is in it only because someone ran the command that put it there.
The second-model pass raised 865 observations at error level across the corpus, and rewrote 263 pages as a result. The rest are recorded in the repository rather than on the page.
The writing rules
- Define the term immediately. Short sentences, one new idea at a time, no explaining jargon with more jargon.
- Every analogy states where it stops working. This is enforced by the schema, not by good intentions. A term without it cannot be published.
- No anthropomorphism. A model does not understand, think, know, want or decide, and the validator flags drafts that say it does.
- No hype and no doom. Fluent output is never treated as evidence of accuracy.
- Examples come from many regions and many kinds of work. No cats, no dogs, no recipes.
- United Nations examples never suggest putting confidential, personal, survivor, witness or security-sensitive information into an unapproved public AI service. Member States and contested political questions are treated neutrally.
- Where terminology is genuinely disputed, the page says so instead of picking a side.
What this means you should not assume
These pages are a starting point, not an authority. They were drafted by a model and checked by another model, which catches a great deal but not everything. Where a definition matters, in negotiated text, in a procurement document, in a legal opinion, go to the primary source, which every page links to.
Terminology in this field moves. A page written today can be describing yesterday's consensus. Each term records the prompt version it was written under (currently 1.0.0, 1.1.0) so that a regeneration pass can find the stale ones.
This is an independent educational project by one person. It is not a United Nations product, it is not endorsed by the United Nations, and nothing in it represents the views of the Organization or any of its offices.
Privacy
No login, no account, no password, no user database. No analytics, no cookies, no tracking pixels, no third-party scripts, no fonts loaded from anyone else's server. What you search for, what you save and what you get wrong in a quick check stay in your browser and are never transmitted, including to the AI service that helped write these pages, which is not connected to the published site in any way.
See for yourself
The taxonomy, the prompts, the schema, the generation scripts, the validator, the link checker and every published term file are in the repository.