UN General Assembly Speeches Now Sound More Like ChatGPT

When heads of state walk up to the UN General Assembly podium, they now reach for ChatGPT's favourite ceremonial words twice as often as they did before 2023. We counted unwavering, steadfast, a testament to, at a crossroads and seven similar words in 956 English statements from the General Debate, the high-level speeches every government delivers in New York each September. From 2015 to 2022 the rate held at about 6 per 10,000 words. In 2024 and 2025 it reached 12.2. As the 81st session's general debate (UNGA 81) opens on 22 September 2026, the language of the world's highest podium has already changed.

Do World Leaders Use ChatGPT to Write UN Speeches?

The speeches now carry the vocabulary ChatGPT favours at twice the pre-2023 rate, and the change appears in two countries out of three. A word count cannot prove that any single speech was drafted with AI, so this study names no country and no leader. What it shows is that the register at the General Assembly podium moved exactly when language models arrived.

Three explanations fit the pattern, and they can all be true at once. Speechwriters may draft or polish statements with an assistant. Delegations may translate texts with AI tools. And writers may absorb the phrasing from the machine-shaped text they read every day, the same diffusion our lab measured in AI vocabulary spreading through scientific abstracts.

ChatGPT Words in UN General Assembly Speeches

Crossroads, unwavering, steadfast and testament rose the most, between 2.5 and 5.3 times their pre-ChatGPT rate (figures below are uses per 100,000 words). These are stock phrases of ceremony: "our unwavering commitment", "a testament to multilateralism", "the world stands at a crossroads". Not every word moved, and the ones that fell are reported too.

The share of statements containing unwavering tells the story in one number: about one in ten every year from 2015 to 2022, then 31% in 2024 and 28% in 2025.

How Much AI Language Is in UN Speeches Now?

The ceremonial AI register has doubled. The eleven words appeared 6.08 times per 10,000 words across the six sessions before ChatGPT and 12.19 times across 2024 and 2025, with 2023 sitting in between as the first session after release. The last column is the share of statements that used unwavering.

The intervals for 2024 and 2025 sit entirely above every pre-ChatGPT year, so the jump is not sampling noise. The 2022 speeches, given two months before ChatGPT launched, are among the lowest in the series.

How We Analysed UN General Debate Speeches for AI Words

We used the statements delegations publish on the UN's own General Debate website for sessions 70 to 73 (2015-2018) and 76 to 80 (2021-2025), keeping only speeches whose sole published text is English. That leaves 956 statements and 1.85 million words, produced the same way in every year. The 2019 and 2020 sessions are not available there in comparable form.

The word list was written down before the first count. Each speech was cleaned of page numbers and line breaks, and every match was counted with a pattern that catches inflections such as fostering and resonated. Rates are per 10,000 words, with intervals from resampling the statements 2,000 times.

Is the Rise in AI Words Proof of ChatGPT Use?

Not on its own, which is why the study was built to rule out the other explanations, and every check points the same way. Five words that were already climbing before ChatGPT were removed by a rule fixed in advance. Transformative is the clearest case: it tripled in 2022 because that year's General Debate theme was "transformative solutions to interlocking challenges".

• Each country against itself. Rates rose in 72 countries and fell in 36, with three unchanged (sign test p = 0.0007). Changes in which countries publish English texts cannot produce that.
• A placebo with no ChatGPT in it. Comparing 2015-2018 with 2021-2022 gives 55 rises and 53 falls (p = 0.92), which is what noise looks like.
• Ordinary diplomatic vocabulary. Peace, security, sovereignty, cooperation and four other staples moved about 7% with no significant majority of countries (p = 0.18), while the ceremonial list doubled.
• Translation. Restricted to countries where English is an official language, the rise still holds: 35 up, 18 down (p = 0.027).

AI Words in Speeches vs Academic Writing

AI vocabulary depends on genre, so the words that mark machine style in research papers are not the ones rising at the UN. The academic words that surged sevenfold in computational linguistics abstracts, such as delve, intricate and showcase, show no consistent movement in General Assembly statements: 48 countries rose and 47 fell. Oratory has its own machine register, built from intensifiers and stock metaphors rather than research jargon. Pivotal is the one academic word that crossed over, rising from 2.3 to 7.4 per 100,000 words.

UN General Assembly 2026 (UNGA 81): What to Expect From the Speeches

If the trend holds, the English statements at UNGA 81 will use these ceremonial words at least 12 times per 10,000 words, and unwavering will again appear in more than one speech in four. The high-level general debate runs from 22 to 28 September 2026, and delegations publish their texts on the same UN website. We will add the 2026 statements to this analysis once the session closes and report whether the curve kept climbing.

Do Words Like "Unwavering" Mean a Speech Was Written by AI?

No. Every one of these words was in use long before language models, and the 2015 statements prove it. A doubling across a whole institution says the register is spreading; it cannot convict a single sentence. Readers who treat stock phrases as proof of machine authorship will increasingly misjudge people who simply write the way the world now writes.

That is also the practical lesson for anyone checking their own drafts. Word lists are blunt, and the words that most often make prose read as AI depend on the kind of text you are writing. A statement, an essay and an abstract each carry a different set of tells.

UN Speeches AI Study: Data, Code and Preprint

Everything is public. The full preprint, with methods, all tests and limitations, is on ResearchGate (DOI 10.13140/RG.2.2.29604.64645).

The per-statement counts, the link to every source PDF on the UN site, and the scripts that rebuild each table and figure are on GitHub. One warning for anyone replicating with an existing speech corpus: a widely used collection translated its 2024 statements into English with GPT-4o, which would quietly inflate any AI-vocabulary count. The delegations' own published texts avoid that trap.