Method, sources and limits

What this is

An analysis of the statements delivered in the General Debate of the 81st session of the United Nations General Assembly, 22–26 and 28 September 2026, read against every General Debate since 1946.

This is a working prototype. The measures on it have not been validated against human coding, and nothing here is an official product of any United Nations office.

The historical corpus is the UN General Debate Corpus. The 2026 statements were collected from gadebate.un.org as they were published during the debate.

The unit of measurement is the sentence

Every statement is split into sentences, and every sentence keeps a character offset into the text it came from. That is what lets any figure on this site be traced back to the words that produced it. The segmentation is checked by a gate that requires 100% of characters to fall inside some sentence: a word that belongs to no sentence could not be quoted, and would silently drop out of every count.

Where the English comes from

Delegations publish in different ways, and the four routes are not equivalent. Each statement says on its own page which one it took.

A United Nations transcript is an automated transcription of the spoken interpretation, and says so at source. A translated document is machine-translated here. Neither is the delegation's own English, and a quotation drawn from one should be read that way.

Two different measurements, and they are not the same

The globe and the vocabulary pages count words: did this delegation use this term, and in which sentence. It is exact, reproducible and needs no judgement — the same pattern is applied to every statement since 1946, and every result carries the sentence it came from so you can check it.

The topic pages measure something else: whether a sentence is about a subject, which is a reading, not a match. A statement can discuss artificial intelligence at length without using the phrase, and can use the phrase once in a list of six things without discussing it at all. Where both exist, they will disagree, and neither is a correction of the other.

What this cannot show

• Three statements are published entirely or mostly in capital letters — Eswatini, Cameroon and Argentina. Automated topic coding cannot reproduce a quotation from them character for character, so they are reported as uncoded rather than matched loosely.
• One statement's text is out of order at source. The United Arab Emirates' document is a fully justified PDF whose lines were extracted in the wrong sequence, so some sentences read as scrambled. It is published here as it was extracted, and is excluded from topic coding rather than silently coded.
• Counting a word is not measuring a position. A statement that names artificial intelligence in order to warn against it counts the same as one that welcomes it. The vocabulary series and the globe measure attention, not agreement.
• The globe draws no borders. Land outlines only, and one mark of equal size per delegation at its own position. No boundary is depicted, disputed or otherwise, and no delegation is drawn larger than another. Twenty-eight small island and city states have no shape at the resolution used and would have vanished from the map entirely, so their positions are set explicitly — a missing mark would read as “did not speak”.
• The topic layer is still being built. Statements show their text in full; where topics are absent, that means not yet measured, not nothing said.

Data version

Not yet published in this version

These parts of the analysis exist in the plan and not yet in the data. They are listed rather than omitted, so that what is missing is visible.

Country groups

Group membership comes from published rosters, each with its source and the date it was checked. A roster that could not be verified is not used: the G20 is excluded here, because the sources available list summit participants rather than members.