Key takeaways
- GPTZero flagged 30 of 174 national statements at the 2026 UN General Debate as AI-generated, up from 6 of 174 in 2025.
- Nine leaders, from Sierra Leone to Fiji to Cambodia, used variations of the same "trust cannot be restored by declarations alone" sentence.
- Four delegations repeated nearly identical rising-seas lines that mirror text on the UN's own website, per Politico's reporting.
- Detector verdicts on individual speeches deserve caution, especially for non-native English speakers, but the sentence-level overlaps are visible without any detector.
- The lesson for professionals is the same one the UN just learned in public: when everyone drafts from the same models, everyone sounds the same.
Every September, world leaders get fifteen minutes at the United Nations podium to say what their country stands for. This year, a noticeable number of them used the slot to say the same thing, in nearly the same words. An analysis by the AI detection company GPTZero, published September 28, flagged 30 of the 174 national statements at the General Debate as AI-generated. Last year the same test flagged 6. Politico covered the findings the same day, and the reporter's side-by-side tables did the arguing on their own.
World leaders getting help with their speeches is old news. The speeches matching each other is not.
What the analysis actually found
GPTZero researchers Avikam Mangla and Edward Tian compared this year's General Debate statements against each other and against past sessions, filtering out standard diplomatic boilerplate. In a companion analysis covering 768 speeches across the wider session, they classified 107 as AI-generated, about 16.5 percent of all words spoken. Among speeches delivered in English, roughly one in three was flagged.
The more striking evidence needs no classifier at all. Nine leaders built a sentence on the identical frame that trust in the UN cannot be restored "by declarations alone." Sierra Leone's president said trust grows when promises are honored. Montenegro's president said trust is built when commitments are honored. Fiji's president said trust is restored when commitments are kept. Somalia's president said it will be restored when institutions act consistently. Each version swaps a word or two and keeps the skeleton.
A second template turned up in three speeches, a line describing every nation "regardless of its size, wealth, or influence," with the final word swapped for power or military power depending on the speaker. GPTZero checked past sessions and found that construction in neither the 2025 debate nor the 2022 one, the last before ChatGPT launched. And per Politico, delegations from Tanzania, Canada, Ireland, and Norway delivered nearly identical sentences about rising seas that closely mirror text published on the UN's own website.
The flagged speeches also share the small habits readers now recognize. Phrases like "taught us a simple lesson" appeared in 6 of the 30 flagged speeches and in exactly one of the 127 rated human. That is the pattern we catalog in the AI tells readers notice in 2026. The giveaway is a texture of stock constructions arriving together rather than any single forbidden word.
Some delegations are repeat customers. GPTZero's wider sweep found the speeches of Kenya's William Ruto and Sierra Leone's Julius Maada Bio flagged in both 2025 and 2026, which suggests settled workflow rather than one-off experimentation. And the trend line only points one direction. The count went from 6 flagged statements in 2025 to 30 this year, with the tooling getting cheaper and the drafting calendar getting tighter every session.
The caution the coverage skipped
A detector score on any single speech deserves skepticism, and this is a case where the skepticism has teeth. Many UN statements are written by non-native English speakers, worked over by committees, and translated. Research has shown detectors misread exactly that kind of prose. One widely cited Stanford study found detectors flagged essays by non-native English writers at dramatically higher rates. GPTZero's own trust-template table includes three leaders whose speeches rated as human despite using the shared sentence, which is the honest detail in the analysis and worth repeating.
So treat the per-speech verdicts as estimates. The template sentences are different. Ten near-identical constructions across nine governments is not a probability score. Anyone can read the quotes side by side and see the same scaffold underneath, and that scaffold did not exist in the transcripts from three years ago.
Why everyone suddenly sounds the same
Nothing about this requires a conspiracy. Speechwriting teams in different capitals asked similar tools for help with the same assignment, a short address about multilateralism, trust, and climate. The models obliged with their default moves, and the defaults converged, because the same handful of systems trained on the same public record produce the same favorite sentences. It is the same mechanism that fills LinkedIn with interchangeable posts, which LinkedIn now demotes on sight, and the same flood dynamic we mapped in what AI slop actually is. The UN version just came with heads of state attached.
There is real irony in the messenger. GPTZero built its reputation flagging student essays, and we have examined how its judgments hold up elsewhere. This week its classifier was pointed at presidents, and the presidents fared about like the students.
The lesson if you write for a living
A UN speech is the most expensive writing assignment on earth. A nation gets one televised chance per year to sound like itself, and this year dozens of them sounded like the same helpful assistant. That is the reputational risk in miniature for everyone else. Your proposal, your pitch, your keynote, and your LinkedIn post are all competing against a thousand documents drafted from the same defaults, read by audiences who now recognize those defaults instantly.
The fix is the one the flagged speeches skipped. Start from the model's draft if you want, then make it yours. Cut the template sentences, add the details only your experience supplies, and rewrite the rhythm until it sounds like a person with a point of view. Before anything important ships, run it through a free AI detector to see how it reads to a classifier, then put the flat sections through WriteHuman's AI humanizer, which restructures the phrasing and cadence that make prose read as default output. It is free to try on the homepage without an account, three humanizations a month at 250 words each. What it cannot supply is the part the UN speeches were missing too, which is something only you would say. Bring that, and no template can touch you.
Frequently asked questions
Do world leaders actually use AI to write their speeches?
Yes, some of them pretty clearly do. GPTZero flagged 30 of this year's 174 General Debate speeches, and nine leaders used nearly the same sentence about trust. Nobody has admitted it, but leaders have always had ghostwriters. The difference is that the ghostwriter used to be different in every country, and now it's giving everyone the same lines.
How many UN speeches were flagged as AI in 2026?
30 of the 174 General Debate statements. Looking at the whole session, GPTZero flagged 107 of 768 speeches, which works out to about 16.5 percent of every word spoken. For speeches given in English, it was roughly one in three.
Can a detector really prove a speech was AI-written?
No. A detector score is an educated guess, and detectors guess wrong more often on non-native English writers, which describes a lot of UN delegations. Three leaders used the same trust template and still scored as human. The convincing part isn't the scores anyway. It's reading the speeches side by side and seeing the same sentence skeleton, one that didn't exist in any transcript before ChatGPT.
Is it wrong to use AI to draft a speech?
Using AI to draft isn't the problem. Reading the draft out loud without changing anything is. Everyone else skipping the editing is getting the same default sentences from the same tools, and that's how nine presidents ended up saying the same thing. The embarrassment wasn't using AI. It was sounding like everyone else at the one podium where a country is supposed to sound like itself.
