One podium, 191 voices, three questions
Every September each member state takes the same podium, in the same format, within the same week. That makes the general debate the best place to compare what governments are willing to say in public. These speeches are strategic: a state says what it wants others to hear.
This brief asks three questions of the texts. How often do states demand institutional change to the UN and the world order? How often do they acknowledge the risks of artificial intelligence? And how often do they use the second to justify the first?
Three distinctions that guide the reading
Criticism is not a reform demand. Condemning an institution is not the same as asking for a specific change.
Discussing a risk is not acknowledging it. A speaker may raise the threat only to reject it.
Co-occurrence is not a link. Mentioning reform and AI in one speech does not mean one is the argument for the other.
Asking for reform is the norm
162 of 191 texts (84.8%) ask for at least one institutional change. Reform is the rule, not the exception.
But the marker counts the demand, not its direction. The United States welcomes a 15% cut to the UN budget and the loss of more than four thousand posts. Others ask for a stronger organisation. Both count as “UN reform”, so 79% cannot be read as support for a stronger UN.
Criticism is not a programme
152 of 190 texts criticise the current international order. Criticism and reform mostly travel together, but not always.
14 texts criticise and ask for nothing specific. Brazil's speech says the UN is failing one of its core missions and calls the veto a flaw built into the Charter. Yet it names no concrete institutional change, so all five reform markers are zero.
Its credibility has never been lower.Brazil · President Luiz Inácio Lula da Silva · 22 Sep · source, p. 1–2
23 texts ask for reform without criticism. Mexico backs the Secretary-General's reform effort and co-founded a support group for it with Norway, without faulting the existing order.
AI is now on the podium
Artificial intelligence is named in 133 of 191 texts (69.6%). It has become a standard item of the diplomatic agenda.
Two speakers raise the threat in order to reject it. The United States rejects any scheme of global AI control and attributes catastrophic forecasts to its opponents. Argentina calls dystopian scenarios economically meaningless and commits not to regulate pre-emptively.
These are the same people that are now saying that AI is going to kill us all.United States · President Donald Trump · 22 Sep · source
58 texts do not mention AI at all. Among them is Kenya, whose speech demands all five types of reform, as well as Indonesia, Mexico, Sweden, Denmark, Ireland and New Zealand. Silence is not rejection. It tells us nothing about how much a state cares.
Risk, acknowledged — mostly as inequality
Half the debate acknowledges that AI carries risk: 94 of 191 texts (49.2%). The most common worry is that AI will widen the gap between countries.
It can also widen inequality if infrastructure, knowledge and rule-making remain concentrated in a few countries.Sierra Leone · President Julius Maada Bio · 22 Sep · source, p. 13
Loss of control and the survival of humanity are rare: 12 and 6 texts, and only 8 and 2 on a strict reading. Lithuania warns that control of AI “might slip out of humanity's hands”. Côte d'Ivoire and Liechtenstein are the two clear existential cases.
Two topics, one speech, no bridge
88 texts (46.1%) contain both a reform demand and an acknowledged AI risk. On the surface, the two agendas look intertwined.
Read the texts and 60 of those 88 never connect them. Japan calls for urgent Security Council reform and names AI's national-security and disinformation risks, but argues for reform on general grounds and handles AI through its own Hiroshima process.
The reverse also holds. Greece and Ukraine both warn that AI could escape human control, yet neither asks for any reform of the UN or the world order. A severe risk does not automatically become an institutional demand.
…there is already a real possibility that AI – not only people – will begin to decide what happens on the battlefield.Ukraine · President Volodymyr Zelenskyy · 23 Sep · source, p. 9
The narrow bridge: 27 speeches
Only 27 of 190 texts (14.2%) justify institutional change by AI risk. Ten of them are borderline calls; on a strict reading, 17 remain.
Even these 27 do not ask for the same thing. They point to different institutional places:
A binding instrument
Dialogue can bring positions closer, but it does not create obligations.Bahrain · Foreign Minister Abdullatif bin Rashid Al Zayani · 25 Sep · source, p. 6–7
Rules written under the UN
UN reform as the answer
Who writes the rules
Outside the UN
One label, opposite demands
28 texts ask to change the veto. Counted together they look like a camp. Read one by one, they pull in three directions.
End it
Spain, the Maldives, Malaysia and Finland, which says “no single state should have veto power”.
Limit it
France, Estonia and Barbados back restraint in mass-atrocity cases. Lebanon, calling itself realist, asks to limit its use rather than abolish it.
Share it
The African Common Position: permanent seats with all prerogatives, “including the veto for as long as it exists” (Nigeria, Sierra Leone).
Mauritius holds two positions at once: if the veto stays, new permanent members must get it, and it also supports restraint. Russia wants a larger Council and no change to the veto rule, so it does not appear in the veto marker at all.
The texts even disagree on the facts. France says 128 states support the French–Mexican restraint initiative; Barbados puts the figure at 123.
Counting supporters is not an argument
Frequency shows that a topic is on the diplomatic agenda. It cannot show a shared model, coordinated action, or which position is right. 162 texts ask for reform, but those 162 do not back one project. Any institutional design should be judged by how it answers three questions that come straight out of the debate:
Who can stop a decision?
On which issue, on what grounds, and who reviews the stop? The debate contains abolishers, restrainers and extenders of the veto. A model that serves one of them loses the other two.
How are money, voice and access shared?
73 texts ask to reform the financial architecture, and inequality is the AI risk named most often. A model that only addresses security leaves the most common concern unanswered.
What is binding, and for whom?
The same hall holds states that demand binding international rules and states that reject global control outright.
AI risk has not yet become a common argument for reform. 101 texts discuss AI governance and 94 acknowledge risk, but only 27 use that risk to argue for institutional change. And the reform agenda stands without AI: 42 texts demand reform and never mention AI, 29 of them for the Security Council and 18 for the financial architecture.
CIPRF interpretive lens · separate from the findings
CIPRF's research programme on optimalism (Emzar Khvichia) asks the same questions from another angle: how to secure the coexistence of different functions without requiring one model for all. The study takes questions from the doctrine, not its allocation principle, and adds conditions the sources do not supply: equal rights, non-discrimination, a duty to give reasons and a right of appeal. This dataset neither confirms nor refutes the doctrine.
Read it with care
- Texts, not delivery. 158 texts come from PDFs with a text layer, 20 from scanned PDFs recovered by OCR and 13 from automatic transcripts. None was checked against video. In two cases the transcript contains passages that are missing from the submitted text.
- Machine coding. Language models did the coding. 24 of the 130 newly coded texts were re-coded blind by a second model: agreement 96.9% (Krippendorff's α 0.92). The world-order marker is the weakest (α 0.53). Agreement between two models is not proof of accuracy.
- Mention, not weight. A marker records whether a text raises a topic, not how much space it gives it. The direction of a demand was not coded.
- 191 is not 193. Afghanistan and Myanmar are missing, and five cells are unknown. Missing data is never counted as zero; with both absent members set to all-yes or all-no, the share demanding reform stays between 83.9% and 85.0%.
- Description, not cause. 191 texts are the full schedule, not a sample. No number here supports a causal claim.