Methodology

The flag artwork

Every flag on this site is an SVG from flag-icons by Panayiotis Lipiridis, used under the MIT license and served from this site rather than from a third party. The images are the one part of this site that is not our work and not in question: they are drawn from published specifications and maintained by that project.

The facts

Everything else, the adoption year, the color hex values and what each color is said to mean, the symbolism paragraph, the designer credit, and the written comparison note, is a single dataset of 195 entries that was researched and written by an AI tool. No person has gone through it. There is no editor, no vexillologist, and no author byline, because there is nobody whose checking a byline would stand for.

The one thing on a flag page that is neither researched nor written is the “Most similar flags” list. It is computed at build time from the flag images themselves, by a method that is written out in full on the look-alike page. That makes it reproducible rather than trustworthy: it is a measure by a disclosed weighting, and a different weighting would name different neighbours.

Since then, 17 of the 195 entries have been checked against a primary source: the document that actually constitutes the flag, meaning a flag act, a constitution, or a government-published specification. Not an encyclopaedia, and not this site’s own earlier guesswork. Those pages carry the source, a quote from it, the date it was read, and a superscript on each fact the source backs. Everything without a superscript is unchecked, including on those pages.

That number has grown. It began at eleven and is now 17, after a round of EU and Western-European flags whose constituting law is published as readable text — France, Italy, Norway, Sweden, Croatia and Lithuania — were each held against that law: their colours confirmed in its exact wording, Norway’s missing white restored, and every colour-meaning the law does not state removed rather than left to look verified. This primary-source check is a separate, stricter exercise from the secondary-reference reconciliation described just below — holding an entry against the constituting law is not the same as attaching an encyclopaedia to a claim — and the two are counted and marked apart, never merged. In short: 17 verified against the constituting law, 178 referenced.

That checking was automated. It compared the entry against the text of the document and nothing more; no person read either. The quotes are published so you can do what we did not: judge the comparison yourself.

The other 178 entries have since been given secondary references: for each, the country’s flag article on Wikipedia. These are marked deliberately unlike the checked pages — a blue (r) rather than the crimson [ ], under the heading References — not verification. A reference means a claim can be read in a reputable place, not that the constituting law says it. Where the article stated no meaning for a colour, that colour now carries none.

Attaching those references was not a rubber stamp, and the result is the most telling number on this site. Of 698 factual claims examined across the referenced pages — what each colour is said to mean, and each symbolism paragraph — 368 matched the source as written, 260 had to be corrected to match it, and 70 were removed because no allowed source stated them at all. Nearly half of what the AI pass had written, in other words, either contradicted the reference or could not be found in it. Adoption years and the lists of colours themselves almost always held up; the meanings and the histories are where the errors clustered.

The removals were the condition of the exercise: a fact that no primary or secondary source supports does not belong on the page, so it is gone rather than flagged with a warning. Among the corrections were a flag whose colours had been assigned to the wrong islands, a “blood of martyrs” story attached to a colour the source explains as a purple-dye heritage, an emblem described as an antelope that the source calls a stylised human figure, and one country’s entry that described a flag its own artwork had already replaced.

None of this makes the 178 referenced pages verified. A secondary reference is not a primary source and it is not the law: it tells you a claim is attributable, not that it is correct. The 17 checked pages remain the only ones held against the document that actually constitutes the flag.

The pages are a static build: each country page is generated from that one dataset file at build time, so what you read is exactly what is in the data, with nothing added per page.

What we know about the error rate

Rather than guess at how reliable the data is, we checked a sample. A representative 20 countries were cross-checked against Wikipedia, Flags of the World, and official sources. 13 of those 20 contained at least one factual error. That figure describes the original sample, taken before the revision work: it is the state the dataset shipped in, not a measurement of it today. It has not been re-measured since, because a second sample would cost what checking the entries costs, and checking them is the more useful of the two.

The errors were not subtle. Romania’s flag was dated to 1848 instead of 1989, a gap of 141 years. Ireland’s was dated 1919 instead of 1922. Flag colors were missing outright: the green script on Iraq’s flag, the white star on Timor-Leste’s and on Myanmar’s, the bird and stars on Papua New Guinea’s, whose triangle description was also the wrong way round. Belize’s and Seychelles’ blues had the wrong hex values. Libya’s black was given a meaning it does not have. Grenada’s stars were miscounted. Iraq’s 2004 and 2008 redesigns were conflated. A claim that Chad formally protested Romania’s flag at the UN was simply invented.

Those 13 were corrected, and two of them, Ireland and Timor-Leste, have since been checked against their constitutions. That check caught one more error the sample had missed: Timor-Leste’s black was explained as something the white star overcomes, a role the constitution gives the star nowhere. Correcting a page is not the same as verifying it, which is why the two are counted separately here.

The remaining 178 countries have not been checked against a primary source. They have, however, been put through the secondary-reference pass described above, which corrected 260 claims and removed 70 — so they are no longer the raw AI output the 20 sampled countries came from. What they are is a dataset reconciled with Wikipedia, which is itself a secondary source that can be wrong. Treat any page without a crimson superscript as unverified, and check anything that matters against a primary source before you rely on it.

We publish this number because a site that names its own error rate is more useful than one that implies it has none.

Corrections

If you find an error, and statistically you will, please write to hello@hellointernet.com. Corrections are applied to the dataset and go live with the next build.