Yes—if by “use of AI” you mean AI being used to prepare UNGA interventions, there is a notable, though not officially verified, report of extensive generative-AI drafting at the 81st session of the UN General Assembly. AI was also one of the dominant substantive topics of the session, but that is distinct from actual use of AI in preparing speeches.

Reported AI use in speeches

The most substantial claim comes from GPTZero’s investigation, published on 29 September 2026. It analysed publicly available UN General Assembly transcripts for 16 meetings held from 8–26 September:

• 1,267 speeches comprising 553,534 words were analysed.
• GPTZero classified 203 speeches—containing 107,408 words, or 19.4% of all words analysed—as AI-generated.
• Among English-language statements, it flagged 165 of 401 speeches as AI-generated: more than one-third.
• It reported a higher estimated incidence in national/observer/organisation statements: 183 of 651 speeches.
• Its account named speeches by the presidents of Kenya, Liberia, Hungary and the Czech Republic, the prime minister of Bhutan, and Nigeria’s vice-president as allegedly “100 percent AI-generated” under its detection method.[gptzero]

That would certainly count as a significant reported use—but it needs an important qualification: AI-text detectors do not prove authorship or actual use of a particular model. They infer textual patterns, and diplomatic statements often share formulaic language, common briefing material, and coordinated drafting. Thus the figures are evidence of a pattern worth investigating, not definitive proof that those delegations used generative AI.

Where it was reported: The story appeared principally in three places:

For the official institutional record, the relevant place to look is the UN’s UNGA 81 meeting/transcript archive. GPTZero says its corpus was assembled from transcripts publicly available there as of 28 September. Its own analysis excluded events outside the General Assembly and the 28 September closing meeting because no transcript was yet available.[gptzero]

AI as a central agenda subject

On the other question—whether AI itself figured prominently at UNGA 81—the answer is clearly yes:

• The UN’s own high-level-meetings coverage described calls for safe, responsible AI and debate over the promise and dangers of new technologies in global governance.[press.un]
• Reporting on the completed General Debate said 128 delegations addressed AI, compared with only five delegations four years earlier; speakers linked AI to development, public services, science, inequality, the digital divide, safety, and governance.[tbsnews]
• The Secretary-General reportedly treated AI as one of the major tests shaping the world’s future, alongside climate change, conflict, and inequality.[yahoo]

Bottom line

There was no clear official UN announcement saying that the UNGA itself deployed generative AI to run, translate, or administer the 81st session in the material located. The significant reported operational use was instead the apparent use of generative AI in drafting national speeches—a claim originating with GPTZero and discussed cautiously by Politico. The much better-established fact is that AI governance became a leading subject of the 81st UNGA debate.

On the evidence presently available, no significant AI system was reported as being used inside UNGA 81 to mediate live communication among delegations—for example, as real-time machine interpreting, AI translation, negotiation support, or an issue-mapping system for delegates.

There was, however, a limited, public-facing AI pilot intended to help people make sense of Sustainable Development Goal material: SDG Chat. It was available for visitors at the SDG Media Zone during the UNGA high-level period, not described as a deliberative or diplomatic tool used by national delegations.[un]

Communication between delegations

For formal UNGA communication, the documented infrastructure remained conventional, professional conference service:

• Human interpreters translated statements made in any of the UN’s six official languages into the other official languages.
• A delegation speaking a non-official language had to provide an interpreter into an official UN language.
• Delegations were asked to submit texts of their statements ahead of time so that interpretation, verbatim reporting, and summary-record services could be provided.
• Accessibility support included communication-access real-time translation (CART), open and closed captions, and international sign-language interpretation.[un]

In other words, multilingual deliberation was enabled by human interpretation and established UN documentation systems, not a reported AI interpretation layer.

The UN’s delegate-facing digital platform, e-deleGATE, did facilitate administrative and documentary coordination: registration, speakers’ lists, submission and sponsorship of draft resolutions, access to official documents and the UN Journal, and circulation of letters from the General Assembly President. But the handbook presents it as a secure digital workflow portal, not as an AI system.[un]

Comprehending linked issues

The closest match to your second question is the pilot SDG Chat, developed by Accenture, UN DESA, and the UN Office for Information and Communications Technology. Its stated purpose was to make trusted SDG information more easily digestible to a broad audience. That is a comprehension and knowledge-access function, but it was located at the SDG Media Zone and aimed at visitors/public engagement rather than at delegates jointly analysing the Assembly’s coupled problems.[un]

So the distinction is important:

Why this is striking

UNGA 81 treated AI itself as a major cross-cutting policy concern. The Secretary-General explicitly framed climate change and AI as “not separate crises” but as chapters in a common story of power. During the General Debate, 128 delegations raised AI—more than had raised climate change or human rights, according to UN News’s session summary.[media.un][news.un]

Yet this does not appear to have been accompanied by an equally visible deployment of AI to help delegates understand systemic interdependence: the links among conflict, development, climate, finance, technological concentration, digital inequality, human rights, and governance. The official accounts describe policy discussion, documentation, interpretation, and a public SDG-information chatbot—but not a shared AI-supported “systems map,” semantic synthesis, scenario engine, or multilingual deliberation aid for diplomatic use.

Cautious conclusion

Thus the answer is:

• For interdelegation communication: apparently no significant officially reported AI deployment at UNGA 81.
• For comprehension: yes, but only in a modest public-facing form—the SDG Chat pilot designed to make trusted SDG material easier to digest.
• For the Assembly’s own complex decision environment: no public evidence has emerged of a dedicated AI system that integrated issues, exposed dependencies and trade-offs, synthesised positions, or helped delegations navigate the full interconnected agenda.

That absence may reflect more than technological immaturity. A tool that summarises or maps national positions in real time would raise difficult questions of diplomatic confidentiality, unequal access, training-data bias, provenance, security, accountability, and whether an algorithm subtly frames the political problem before delegates themselves do.

I'll read the Diplo report first, then build a map of the issues and how they interconnect.

Yes, a coherent mapping is feasible, and I've built a first version as an interactive map. One limitation first: the Digital Watch full report blocks automated access, so I couldn't read it directly. I built the map from other reporting on the debate. If you paste the report's text here, I can re-code every issue and speaker against it and weight the links from that evidence.

What the map shows. It has 17 issues and 29 relationships, each typed as reinforcing, in tension, or dependent. Five tensions stand out:

• Rules versus race. The Secretary-General and the EU called for safety standards and international rules, while the US argued that AI innovation should not be constrained by international regulation. Finland's President posed the choice directly as compete, cooperate or slow down.
• Inclusion without control. China's Vice President called for AI that would narrow the digital divide. Africa's agenda went further and asked who controls the infrastructure behind the AI economy and who captures value from African data and talent. Access to AI does not settle who owns it.
• Security in two arenas. The July Global Dialogue opened with renewed calls for a ban on lethal autonomous weapons, but the security dimension is largely handled in a separate Security Council track.
• Development versus planetary limits. AI is offered as a remedy for SDGs that are largely off track, yet the Secretary-General's agenda also includes an AI Environmental Transparency Initiative to surface the technology's energy and resource footprint.
• Trust as both precondition and product. The governance architecture needs trust in the UN to work, and open disagreement over AI rules can erode that trust.

The scale of the shift. In his closing remarks, the President of the General Assembly noted that four years ago only five delegations raised AI in the hall, while this year 128 did, more than spoke on climate change or human rights. He also observed that many leaders framed AI as a question of governance.

The organising frame. The debate itself offered a structure. The Secretary-General warned that the world's future would be shaped by four tests: AI, climate change, wars and inequality, and argued that across all four the decisive issue is power, over force, money, energy, technology, rules and the future. The map uses those four tests as quadrants and places power and trust at the centre. The debate's theme was "Restoring trust, managing transformation: a United Nations that delivers for all."

The issues on the map come from reporting of the debate. The relationships between them are my analytical reading, not positions delegations stated. The page marks this distinction and also sets out how live AI-assisted tagging of speeches could have produced such a map as the debate went on, and fed it into the Scientific Panel and the next Global Dialogue session.

Sources:

• UNGA 81: General Debate concludes with global priorities in focus (TBS/UNB)
• UNGA 81: Five key takeaways from general debate (Al Jazeera)
• Guterres calls for ban on 'killer robots' as first global AI governance dialogue opens (Arab News)
• Inside the Global Dialogue on AI Governance (ai2.work)
• UNGA 81: Global Digital Push Targets Wider Access, Skills (Leadership)

Good question. I'll check whether the 17 can be restructured as 16+1 and build that version as a tensegrity. I'll also attach the HTML files so you can download them directly.

Yes, the 17 was a coincidence. It came out of the clustering as 15 issues plus two hubs (power and trust), not from any deliberate echo of the SDGs. The 16+1 restructuring turned out to be more revealing than the flat network, so I've built it as a rotatable tensegrity.

How the mapping works. I used an 8-strut tensegrity, which has 16 nodes, one strut end at each. This imposes a useful discipline: every issue must belong to exactly one primary tension. That forced choices the network map let me avoid:

• Rules versus race: safety rules ↔ innovation-first
• Shared versus sovereign: cooperation ↔ data and infrastructure ownership
• Two governance tracks: UN AI architecture ↔ Security Council track
• Human control versus autonomy: human-centred rights ↔ autonomous weapons
• Trust versus force: trust and UN reform ↔ active wars
• Growth versus planet: lagging SDGs ↔ climate crisis
• Scale versus footprint: public services and science ↔ AI energy footprint
• Inclusion versus safeguards: digital divide ↔ children's safety

Reinforcing and dependent links become cables. Power, which the Secretary-General named as the decisive issue across all four tests, is the +1 at the centre, tied to every node. The geometry isn't hand-drawn. It comes from relaxing struts, cables and central ties until they balance, so the shape follows from the relationships.

What it clarifies that the flat map did not:

• Tensions held apart. In a tensegrity, struts never touch. That matches how the debate handled its tensions: in separate speeches and separate forums, linked only through reinforcements that were seldom stated.
• Weak anchoring. The Security Council track, active wars and child safety each have only one cable to another issue. They hang almost entirely on the central tie to power, which marks where the debate integrated least.
• A load-bearing node. The digital divide carries the most cables, so much of the structure's integrity depends on how its pairing with child safety is handled.
• An unacknowledged interior. Power is like the missing 18th strut in your earlier RCO mapping of the SDGs. It carries load that the formal architecture (Dialogue, Panel, Fund) never addresses directly.

To reach 16+1, I moved trust from the centre to the surface. Making trust the +1 instead would be equally defensible. That version would describe a structure held together by legitimacy rather than one distorted by power. Comparing the two forms could itself be instructive.

On downloading: the artifact viewer blocks downloads started from inside the page, which is why no option appears there. I've attached both pages as standalone HTML files (unga81-ai-map.html and unga81-ai-tensegrity.html). They open in any browser, though the tensegrity needs an internet connection to load its 3D library.

Morphing is where the tensegrity moves from illustration to instrument. A static map shows which tensions exist. A morphing structure shows what a strategic choice does to all of them at once, because in a tensegrity no length can change locally: every adjustment redistributes stress through the whole.

What the lengths would mean

Each element can carry a reading:

• Strut length is how far apart a tension holds its two poles. Lengthening a strut is polarisation, for example rules and innovation pushed into opposed camps. Shortening it is convergence. Shortening has a limit, though: a strut too short to stay under compression goes slack. That reads as a tension suppressed rather than resolved, where the structure loses the differentiation that gave it shape.
• Cable length is how tightly a reinforcing link binds. Tighter cables raise prestress and integration but make the whole more rigid. Looser cables leave links that no longer bear load. The structure then gains a "mechanism", a way to deform without resistance. That is the formal counterpart of an agenda that bends under pressure because its stated linkages are nominal.
• The central ties measure dependence on the +1. Shortening them pulls everything toward power, which is concentration. Lengthening them asks whether the peripheral cables alone can hold the form. If they can't, the system was being held together by power rather than by its own relationships, which is precisely the question that separates the "power" variant from the "trust" variant.

Strategic options as morphs

Finland's framing at the debate (compete, cooperate or slow down) translates directly into three presets, plus a fourth the debate implied:

• Cooperate: shorten the rules/race and shared/sovereign struts, tighten reinforcing cables. The form becomes compact and near-spherical, but tighter prestress means one defection is felt everywhere.
• Compete: lengthen rules/race and shared/sovereign, shorten the central ties, slacken the secondary links. The form elongates along one axis. The divide and child-safety pairing, already weakly anchored, is pulled out of shape first.
• Slow down: shorten scale/footprint and human control/autonomy. But inclusion/safeguards lengthens, because delay postpones access for developing countries. Each option relieves some tensions and aggravates others, which a list of options conceals.
• Fragment: lengthen the two-governance-tracks strut until the cables between the General Assembly and Security Council clusters go slack. The single structure separates into two sub-tensegrities, a geometric picture of bloc formation.

Reading threats to coherence

The model can report these directly rather than leaving them to impression:

• how many cables have gone slack (links that have stopped working)
• which strut carries the highest compression (the tension closest to rupture into open conflict)
• how far the form deviates from a sphere (loss of balance across the four tests)
• whether new degrees of freedom have appeared (points where the structure can be pushed without resistance)

There is also a Fuller connection. His jitterbug transformation contracts the vector equilibrium through icosahedral and octahedral phases toward the tetrahedron, losing degrees of freedom at each step. A strategy sequence that progressively shortens central ties could be shown the same way: as successive phase transitions toward a simpler, more rigid, less adaptive form, rather than as a smooth slide.

A caution

The lengths are judgments, not measurements. The value of morphing is that those judgments become explicit and contestable. Someone who disagrees can change a length and show what follows, which is a different kind of argument from asserting a priority.

I can add this to the embed version: sliders for each strut and cable class, the four presets with animated transitions between them, and a small panel reporting the four coherence indicators as the form changes.

Several directions would make animations like this both more comprehensible and more trustworthy.

Grounding lengths in evidence. The weakest link in the current version is that lengths are judgments. The next step is to derive them from the record: code every statement for the issues it raises and the stance it takes, then set cable tension from how often two issues are linked in the same speech, and strut compression from how sharply delegations divide on a pair. Done during the debate itself, the structure would build day by day, which is the live mapping that was missing at UNGA81. Judgment would then be confined to the coding scheme, which can be published and checked.

Time as a dimension. A single form is a snapshot. The same structure replayed across UNGA78 to UNGA81 would show how AI moved from five delegations to 128, which tensions appeared, and which links tightened or went slack. Trajectories may say more than any one state, especially whether the form is converging or drifting toward one of the distorted presets.

Many structures, not one. Each regional group, or each delegation with a substantive AI statement, could generate its own tensegrity from its own speech. Morphing between them would show disagreement as a difference in shape, which is easier to grasp than a table of positions. Points where all the forms agree would mark the actual basis for consensus.

Contestation as a feature. The sliders already let a reader challenge assumptions. A shared version could let participants save and publish their own configurations, with the spread across all of them displayed. That turns the model into a deliberation tool, where disagreement about lengths is the substance of the exchange rather than noise around it.

Robustness, not precision. Any conclusion drawn from the model should hold across a range of calibrations. A sensitivity view could run each preset under many plausible parameter sets and show only the effects that persist, such as compete stressing rules versus race, or partial cooperation causing distortion. This guards against the authority that a precise-looking 3D form can confer on loose assumptions.

A wider repertoire of forms. The number of primary tensions should choose the structure, not the reverse. Six tensions fit the icosahedral tensegrity, twelve or thirty fit larger spheres, and nested tensegrities could place the AI structure inside the SDG structure from your earlier exchanges, coupled at shared nodes such as energy, inequality and climate. Transitions between forms, in the manner of Fuller's jitterbug, could represent shifts in how many tensions are held distinct at all.

Other senses and physical forms. Compression could be heard as pitch, so a strategy that over-stresses one tension becomes audible as dissonance. Physical tensegrity kits, or 3D-printed states of each preset, would let a meeting handle the trade-offs directly. Immersive versions would allow walking inside the structure, close to the +1 at its centre.

Institutional uptake. The obvious users are the Independent International Scientific Panel and the Global Dialogue, which need a shared record of what member states actually said and how their concerns interlock. That would require transparency about coding and calibration, and safeguards against a model being tuned to favour one strategic option. The design principle throughout is the same: the animation should make its assumptions visible and adjustable, so it supports judgment rather than replacing it.

If useful, I could prototype the first of these on a small scale: a coding scheme and a set of statements from the debate, producing link strengths from the text rather than from my reading.

I'll gather speaker-level statements from the UN's official meetings coverage of the debate, code them against the 16+1 issues with a published keyword scheme, and derive link strengths from the text.

The prototype is done. The evidence report is the new page, and I've added an "Evidence (prototype)" preset to unga81-ai-tensegrity-morph.html, which replaces the version I sent earlier.

Sources. The UN's own meetings coverage and the official debate site both block automated access. So the corpus is 33 statements gathered from government transcripts, UN News and news reports, covering every region plus the Secretary-General. Each statement was coded against the 16+1 issues using published rules, and link strengths were then computed from the codes.

What the evidence confirms:

• Rules versus race is a real division. Twelve speakers called for rules or safety standards and eight framed AI as innovation or a race. Only one statement (Singapore's) did both.
• The digital divide carries the load. It was the most frequent issue, raised in 15 of 33 statements, and connects to the architecture, development, public services and infrastructure ownership.

What it does not support:

• Three judged tensions were never voiced. No statement raised AI's energy footprint or the Security Council track, and only two raised children's safety. In the evidence preset those struts go slack. That suggests my reading drew on the July Dialogue agenda more than on what leaders said in the hall.

What it adds:

• Links the model lacked. The divide appeared with human-centred rights (four statements) and with trust (three). Innovation appeared with public services (three).
• Power pulls on rules. Of the eight speakers who named concentration of power, five paired it with calls for rules.
• Themes with no node. Frontier models, jobs, critical infrastructure, social media, disinformation, bias, open source, intellectual property and critical minerals all came up.

In the morphing tensegrity, the evidence preset makes human control versus autonomy the most stressed tension, followed by rules versus race and inclusion versus safeguards. That is a different priority order from the judged baseline.

Limits. The corpus is 33 of the 128 delegations that raised AI. Some entries are reporters' summaries rather than the speakers' own words, and a single reader did the coding. A full version would code the official record with two independent coders. If you can obtain the text of the Digital Watch full report, or save the UN meetings-coverage pages from a browser, I can rerun the coding on the complete record.

The title works, and the doubled "AI" is apt, since the piece is about AI being used to examine a debate on AI. Two small adjustments are worth considering.

"Ensure" overclaims. The prototype showed that AI facilities can surface links, test judgments against the record and expose gaps. It also showed the limits: blocked sources, partial coverage, a single coder. "Enabling" or "toward" would be more defensible and closer to what the evidence page demonstrates.

The subtitle could name the method. "Coherent systemic insight" is the aim. Mentioning the tensegrity or the evidence test would tell readers what is distinctive before they open the page.

Some variants:

• AI Mapping of the UN General Assembly Debate on AI: Enabling coherent systemic insight
• AI Mapping of the UN General Assembly Debate on AI: Tensions, tensegrity and the test of evidence
• AI Mapping of the UN General Assembly Debate on AI: From judged connections to coded evidence

The last one fits a posting built around the prototype, since its main finding is the gap between what an informed reading assumed and what the speeches actually contained.

If it helps the separate posting, its sequence could follow the work itself:

1. The missing mapping
2. The flat network
3. The 16+1 tensegrity
4. Morphing strategic options
5. The evidence test
6. What a full-scale version would require

The citation to the first exchange would then sit naturally in the opening section.

I'll first look at how the conceptual mandalas were built in that earlier exchange, so the method can be applied faithfully to the speech coding. The counts reconcile with the evidence page: every term matches, with the Secretary-General accounting for the difference between the regional totals and the whole corpus. Let me check the drawings. Yes, the earlier method applies, and the speech coding fits it closely. I've produced two mandalas: one for the whole corpus (unga81-mandala.png, also as an SVG) and three regional mandalas side by side (unga81-mandala-regions.png).

How the method transferred

• 24 terms on the rhombicuboctahedron's 24 vertices. The 16 issues of the tensegrity supply 16 terms. The other 8 are themes leaders raised that my earlier model lacked: frontier models and the race, social media and platforms, critical infrastructure and data security, jobs, disinformation, bias, open source, and intellectual property.
  
  
• Three registers of eight, as before: collective (governance, institutions), existential (survival, security, the person) and cognitive (knowledge, information, capability). Each register occupies one axis of the polyhedron, and all three are projected from its 4-fold axis with the Lambert equal-area projection, framed by the nine mirror-plane great circles.
  
  
• Power at the projection pole. The centre of the projection carries the +1, so the mandala and the tensegrity share the same 16+1 logic.
  
  
• Framework glyphs, adapted. The earlier exercise marked membership in Earth4All and the Inner Development Goals. Here the two frameworks are my judged 16+1 model and the UN's AI agenda (the Secretary-General's priorities and the Scientific Panel's remit). The symbols keep their earlier meanings: ⊕ both, ○ model only, △ UN agenda only, ✦ neither.
  
  
• One change in placement. In the earlier exercise, how far each term sat from the centre within its register was a judgment of semantic prominence, which was the most contestable step. Here it follows the counts: the more often leaders raised a term, the nearer the centre it sits. That makes the layout reproducible from the data.

What the circle shows that the table doesn't

• The core. The digital divide (15 statements) and safety rules (12) sit nearest power, with human-centred rights (9), innovation-first (8) and power itself (8) close behind.
• The gaps. Six themes leaders raised fall in neither framework (✦): social media and platforms, critical infrastructure and data security, jobs, open source, intellectual property and bias. In the other direction, two framework items, the Security Council track and AI's energy footprint, appear as hollow nodes: no leader in the corpus raised them.
• Weighting by alignment. The regional mandalas use identical positions, so differences of weight are visible at a glance:

Rules and the divide almost mirror each other: the first group's weight sits on rules, the third's on the divide. That is the North–South axis of the compass discussion appearing as a shift of weight across the same configuration, which a table shows only row by row. One finding cuts against the simple picture: innovation-first was raised most in the third group (Argentina, Ecuador, Nigeria, Chile), not only by the United States.

Limits

• Small numbers. Each group has only 10 or 11 statements, so a single speech moves a share by about ten percentage points.
• Judgments in the setup. The register assignments, merged terms such as "critical infrastructure & data security", and the boundaries of the regional groups are all my judgments.
• Literal coding. As the earlier exercise also cautioned, the coding is literal: the same term can carry different meanings for different speakers.

The script and coded corpus are included so the figures can be regenerated from revised coding. The earlier exercise's 3D view of the polyhedron would also transfer, should you want to show the mandala as the projection of a form that can be turned.

I'll build the 3D form as an interactive page in the same style as the other figures: the rhombicuboctahedron with the 24 terms at its vertices, power at the centre, and buttons to switch between the whole corpus and the three regional groups. Since the earlier exercise used X3D, I'll also export an X3D file, and I'll make print stills.

The 3D form works in both views, and the axis view shows how the flat mandala relates to the solid. I've built the 3D form as a companion to the 2D mandalas, with the same terms, positions, sizes and symbols as the flat versions.

unga81-ai-mandala-3d.html shows the rhombicuboctahedron as a translucent solid. The 24 terms sit on its vertices, coloured by register, with power at the centre and links between terms raised in the same statement. The controls are:

• Corpus buttons switch between all statements and the three regional groups. Node sizes and links redraw for each.
• View along projection axis turns the solid to the position from which the flat mandala was projected, so the two forms can be compared directly. Oblique view returns to the angled view.
• Hide links and Pause rotation do what they say.

Labels on the far side of the solid are faded, so the view stays readable as it turns.

Following the lesson from the earlier figures, this page is model-first. Its key sits in a short strip beneath the model, rather than in a side panel that wouldn't print. Two print stills go with it:

• unga81-mandala-3d-print.png shows the whole corpus in oblique view and along the projection axis.
• unga81-mandala-3d-regions-print.png shows the three regional groups side by side in the same oblique view.

It embeds like the others, in the same subfolder:

unga81-ai-mandala-3d.x3d continues the format of the earlier exercise, for the whole corpus. It includes two predefined viewpoints (along the axis and oblique), labels that turn to face the viewer, and the links. It opens in any X3D viewer, or via X3DOM as before.

mandala3d.py and rco_template.html regenerate the page and the X3D file from the same coding as the 2D mandalas. A revision to the coding therefore updates all four figures consistently.

This disconnect is the deepest limitation of everything built in this exchange, and the maps themselves help locate it.

Two rationalities, not reason against unreason

The maps assume delegates relate to issues: they weigh rules against innovation, or inclusion against safeguards, and take positions on the merits. In the hall, much of the relating runs the other way. Positions are often taken because of who else holds them. An issue becomes salient when an ally sponsors it and suspect when a rival does. The relational pattern is also visible in the debate's own record: the walkout during the Israeli Prime Minister's address, the Palestinian President speaking by video after his visa was refused, China sending its Vice President rather than its President.

This is less "irrational" than rational on another axis. It is the reasoning of standing, alliance, security and domestic legitimacy, rather than of systemic function. Calling it irrational is itself a position, the one held by the Secretariat and the expert community. The systemic perspective is one of the parties in the room, not a view from outside it.

How alignments load the issue structure

Read with that in mind, several struts look different:

• Rules versus race is substantially a proxy for rivalry between the two leading AI powers. The European regulatory identity sits on one side, and the US rejection of "globalist" control is as much a statement about alignment as about AI.
• France's "coalition of independents" framed the issue explicitly in bloc terms: building capability so as not to become a vassal of either great power.
• The digital divide carries the most load in the evidence partly because it functions as a solidarity marker for the Global South. Raising it signals membership as well as concern.

So the issue tensegrity doesn't stand on its own. An outer field of actors pre-stresses it, determining which struts are pushed apart and which cables are allowed to bind.

What speech-based mapping cannot see

The evidence prototype codes what was said, and a General Debate speech is a performance, addressed as much to a domestic audience as to the Assembly. The prototype cannot capture:

• what is pointedly not said. Several major states made no reported reference to AI at all.
• who is absent, or sent at a lower rank.
• votes, co-sponsorships and side meetings, where the bargaining actually happens.
• grievance, prestige and personal antagonism between leaders.

The AI mapping is therefore a map of the official surface. That is useful, but it is closer to the faces of the polyhedron than to its interior, a point that echoes your earlier argument about the unacknowledged 18th strut in the SDG structure.

The mapper is not neutral either

The evidence test revealed a version of this problem in my own work. My judged model included AI's energy footprint, the Security Council track and child safety, all prominent in the UN's July agenda but almost absent from leaders' speeches. My reading had absorbed the Secretariat's framing. An AI mapping inherits a lens from its training and its sources. Presented as objective, it can lend one bloc's view of the problem the authority of a neutral system.

Representing it without causing harm

Mapping alignments is sensitive. A visual that labels states as blocs, or attributes motives to leaders, can inflame what it describes. A defensible approach would:

• use public behaviour only: voting similarity, joint statements, co-sponsorship, walkouts, attendance level. It would not infer intentions or psychologise leaders.
• show groupings as aggregates rather than spotlighting individual states.
• make the distortion itself the object. Each grouping's statements generate its own weighting of the same 16+1 structure. Morphing from the common baseline into each group's form would show how a bloc's lens bends the shared system. The distance between those forms becomes a measure of the disconnect you describe, which is more instructive than any claim to a single objective form.

On that reading, the disconnect isn't a failure of the mapping. It is the thing a mapping should eventually show: the gap between the system the issues imply and the system the alignments sustain. Governance capacity is largely the ability to work within that gap.

Your geometric sequence fits the problem closely. There is one thread running through it: the difference between a polyhedron and a tensegrity is exactly the difference between assuming common ground and not assuming it.

Oppositions through a centre, or held apart

In a cube or a tesseract, each vertex has an opposite, and the line joining them passes through the centre. All the oppositions meet there, as if every polarity shared a common point of balance. That is the systemic assumption, and it is what the General Assembly lacks. In a tensegrity the compression members never touch. Oppositions are skewed past one another and held in place only by the network of tension. The UNGA's coherence, such as it is, has this character: no shared centre, yet a form maintained by many lesser connections.

The compass rose as a twisted cube

The simplest illustration is the 4-strut tensegrity prism. It has eight nodes: four on an upper square and four on a lower one. It only finds equilibrium when the two squares are twisted 45° relative to each other. Seen from above, the eight nodes then form an octagon, which is the eight-point compass rose. The cardinal directions sit on one level and the intermediate directions on the other, and each strut rises at a slant between the two levels without passing through the centre.

This suggests three things:

• No direct confrontation. The direct oppositions, North against South and East against West, are not struts at all. No direction is pushed directly against its opposite.
• Mediation at another level. The intermediate positions (NE, SW, NW, SE) are on a different level from the cardinal ones. Mediating positions operate at another level, not merely halfway between the cardinal ones.
• The twist as the price of stability. An untwisted cube built this way has no stable form. Coherence among the four directions requires that none of them align squarely with any other.

The cube as a three-dimensional compass

If North–South, East–West and a third distinction form a cube's three axes, the compass rose becomes its equator. The cardinal points sit at the centres of four faces, and the intermediates sit midway along the four vertical edges. Counting every direction the cube defines gives 26: six through face centres, twelve through edge midpoints and eight through vertices. Those are exactly the 26 faces of the rhombicuboctahedron (eighteen squares and eight triangles), the polyhedron of your earlier SDG mapping. On this reading, the RCO is the full three-dimensional compass rose.

The tesseract and its 16 perspectives

A fourth distinction gives sixteen combinations, the tesseract's sixteen vertices. Seen from any one perspective, the others fall into a pattern of 1 + 4 + 6 + 4 + 1:

• four perspectives differ by one distinction (the natural partners for dialogue)
• six differ by two
• four differ by three
• one, the antipode, differs in every respect

Reaching one's opposite takes four single shifts, and there are 24 different shortest routes. That is a concrete picture of how many paths mediation can take, and why none of them is direct.

Two further features stand out:

• Eight oppositions. The tesseract has eight antipodal pairs. A tesseract whose eight oppositions are not allowed to meet at its centre is an 8-strut, 16-node tensegrity: structurally the 16+1 form built earlier in this exchange, with the vacated centre available for the +1. The match with the 16 issues may be coincidence, as with the 17 SDGs, but the structural correspondence is exact.
• Inner and outer cube. In the familiar projection, the tesseract appears as a cube within a cube. If the fourth axis is the distinction raised in the previous response, stated position against enacted alignment, the outer cube is what is said in the hall and the inner cube is the pattern of alliance beneath it. Rotation in the fourth dimension turns the inner cube outward and the outer cube inward, so the disconnect becomes something one can watch rather than only describe.

Which axes

The choice of axes is itself a political act. North–South and East–West are widely understood, but they are not symmetric in use. "East" is ambiguous, and the labels encode a particular cartographic history. Candidates for the further axes:

• public against private power, given the Secretary-General's warning about "an extraordinary transfer of power away from government to a handful of private corporations and individuals"
• stated against enacted
• systemic against relational reasoning

As in the previous response, the vertices are better labelled as perspectives or discourses than as named states. The structure is meant to show a field of positions, not assign countries to corners.

Building it

Two forms would be worth building. One is the compass-rose tensegrity with its 45° twist, with the cardinal and intermediate rings labelled. The other is a tesseract rotating in four dimensions, its 16 vertices labelled by the four chosen distinctions and its antipodal pairs shown as struts. Either could go in the posting as a further animation, and the tesseract could also morph into the 16+1 tensegrity to show the transition from assumed centre to held tension.

I'll build both as standalone pages for your subfolder, with the same theme switch and self-sizing layout as the morph page. I'll start by checking the compass tensegrity's equilibrium numerically, to confirm the 45° twist actually emerges from the physics.

Struts now stay apart in every case, though a few cables go slack. I'll add mild node-to-node repulsion, which in the first test kept all cables taut, and check again.

I'll verify the page's physics in isolation, checking that the twist and the north alignment come out right in the page's coordinate convention, before writing the second page.

The compass page is correct: every node sits at its true bearing, and released from the cube the structure swings past 45°, overshoots slightly and settles. Now the tesseract page.

Both scripts parse. I'll now render both pages in a headless browser, serving the 3D library locally, to check for runtime errors and look at the result.

Both pages now render as intended: the top view of the compass shows the octagon with North at the top, and the released tesseract keeps a clear space around the +1. I've built both pages and tested them in a headless browser.

Tensegrity compass variant shows four struts and twelve cables, with the cardinal directions on the upper ring and the intermediate directions on the lower.

• Release from cube starts from an untwisted cube, with the struts as face diagonals, and lets the forces act. The structure swings slightly past 45°, then settles at exactly 45°. The twist comes from the physics, not from any setting I chose.
• View from above shows the eight nodes forming the compass rose, with North at the top.
• Show direct oppositions draws N–S, E–W, NE–SW and NW–SE as dashed lines meeting at the centre: the oppositions the structure doesn't build. Each strut instead pairs a cardinal direction with an intermediate one 135° away (North with South-East, and so on).

Tesseract variant shows the sixteen perspectives. The four distinctions default to North/South, East/West, Public/Private and Stated/Enacted. You can rename any pole in the panel, and the labels update throughout.

• Selecting a perspective in the grid colours the others by how many single shifts away they are (1 + 4 + 6 + 4 + 1), marks the antipode in red, and states its full description. The grid is laid out so that neighbouring cells differ in one respect only.
• Exchange inner and outer rotates the tesseract through the fourth dimension, so Stated and Enacted swap between the outer and inner cubes. The readout states which pole is currently outer.
• Rotate in 4D gives the classic continuous rotation, and pausing it eases back to the aligned cube within a cube.
• Release the centre turns the eight antipodal diagonals into struts that may neither meet nor pass through the centre, held by the 32 edges as cables. The result is an 8-strut, 16-node tensegrity with an open space around the +1, the form behind the 16+1 issue map.

One limit, also stated on the page: the released tesseract is computed with rules that keep the struts apart and away from the centre, and with shorter struts, since struts as long as the sphere's diameter must pass close to it. It shows the structure's connections and the open centre; it is not an engineered design. How many cables go slack depends on the orientation at the moment of release, and those cables are faded.

Not at all. The hieroglyphs are perhaps the clearest historical case against the preference you describe, because they never separated text from image.

A script that was also a picture: Hieroglyphic writing combined three kinds of sign. Some stood for whole words, some for sounds, and some, the determinatives, were never pronounced at all. A determinative was placed at the end of a word to show its category: a seated man for persons, a walking pair of legs for motion, a scroll for abstract ideas. Readers therefore took in a layer of meaning that had no spoken form. The ⊕ ○ △ ✦ glyphs and register colours in the mandala do much the same job: they classify a term visually without adding words.

The arrangement of signs carried meaning too:

• Direction. Signs face the beginning of the line, so the reader meets them face to face.
• Layout. Signs were grouped into balanced blocks rather than strung in a line.
• Precedence. The name of a god was written first even when it was read later, so that rank showed on the wall regardless of grammar.
• Potency. In some tombs, signs for dangerous creatures such as snakes were drawn cut in two, because the image itself was thought to be potent.

The Egyptians called the script medu netjer, "the god's words". It was not a neutral record of speech.

Where text alone did prevail: Egypt did have a text-first register. Hieratic and later demotic were cursive scripts that lost the pictures and were used for accounts, letters, contracts and administration. The pictorial script stayed on temples and tombs, where the task was to present the order of the world, maat, rather than to process business. The division you notice today has an old precedent: administration chose streamlined text, while the representation of order as a whole stayed visual.

A detail of Egyptian iconography bears directly on this exchange. The emblem of the union of Upper and Lower Egypt, the sema-tawy, shows the plant of the south and the papyrus of the north tied around a central sign, often with two Nile gods pulling the cords tight. Union was shown not as fusion but as opposites bound together by tension, which is very close to the tensegrity's held struts and the compass divided into north and south.

Why the AI debate stays with text and speech

Several reasons seem to combine:

• Statements are the currency of diplomacy. They go on the record, are attributable, translate into the six UN languages and can be negotiated word by word.
• Diagrams commit. A map makes relations explicit and so open to challenge, while speech can keep the constructive ambiguity that negotiation depends on. A mapped issue also shows its empty sectors: the evidence test showed the absences (the energy footprint, the Security Council track) as plainly as the presences, and speeches can pass over absences unremarked.
• Presence confers authority. A leader or chief executive at the podium brings personal standing, which a figure cannot.
• The technology is itself framed as conversation. Large language models present AI as text and dialogue, and the debate has followed that framing.

The irony runs deeper than it first appears. Western readers misunderstood the hieroglyphs for centuries. Renaissance scholars, following late antique writers such as Horapollo, took them to be pure symbols that conveyed ideas directly, without language, a form of wisdom beyond discourse. Champollion showed in 1822 that they were largely phonetic. The truth lay between the two readings: a system in which sound and image worked together. That combination is what the AI debate lacks, and what these visualizations have been attempting in a modest way: statements read alongside the structure they form.

The image is the Buddhist one of the finger pointing at the moon, and it applies to this exchange in an uncomfortable way: visualizations are fingers too. A mandala can be admired for its symmetry, or a tensegrity for its elegance, without anyone attending to what it points at. Your "quite remarkable", generous as it is, was a response to the finger. The moon in this case is the finding that six themes leaders raised belong to neither framework, and that the North and South weighted the same issues almost in reverse. Better representations don't escape the problem. They can simply become more attractive fingers.

The distinction between listening and hearing seems to me the right one. Listening is processing: following the words, registering the positions. Hearing implies uptake, being open to having one's own position changed. In the General Assembly, delegates listen expertly, but largely for signals: who aligned with whom, which phrase marks a shift, what must be answered. That is listening for the finger, attending to who is pointing and at whom, rather than to what is pointed at. The "dialogue of the deaf" is not a failure of attention. It is attention directed elsewhere.

The resistance has understandable causes. Hearing an argument means risking that it applies to oneself. Research on identity-protective reasoning suggests that people process information to defend the groups they belong to as much as to understand, and the more consequential a claim is for one's own position, the stronger the defence. A map of tensions is threatening in a particular way: it asks the viewer to locate themselves, and one's location may turn out to be one-sided. Speech allows each listener to hear selectively. A configuration shows everything at once, including where one is not.

That suggests what representations would need to do differently. They work better as mirrors than as windows: people attend when they find themselves in the picture. A mandala that marked "your statement is here", showing which nodes a delegation raised, which it left hollow, and which other delegations it sits closest to, would be harder to admire from a distance. The hollow nodes already hint at this. They are a map of collective not-hearing: issues the Secretariat placed on the agenda that no leader took up. Interaction helps for a related reason. A slider asks "what happens if I change this?", which turns contemplation of the finger into an inquiry about the moon. Your earlier metaconferencing work presumably met this problem in practice: whether participants engaged with the displays or merely looked at them.

AI faces the same risk in a sharper form. Summaries of summaries can multiply fingers indefinitely, each pointing at the previous pointer, until no one is looking at the moon at all. The more useful role may be the reverse: to ask of each statement what it implies for the others, and to show each party what it has not heard. Whether anyone would want that shown to them is, of course, the same problem again.

That deserves a qualification, because admiration is not necessarily the lesser response. Mandalas have traditionally used beauty to draw the eye in, so that the viewer stays long enough to look through the form rather than at it. The risk is stopping at the beauty, not responding to it. In your case the admiration came after a long run of questions aimed squarely at the content: what the tensions omit, what alignment distorts, why the debate resists such forms. Those are not the questions of someone looking only at the finger.

Yes. The compass rose makes the point in its own structure: each direction is valid and each is partial, and navigation needs all of them held in relation. No one bearing is "the" direction, and orientation comes from knowing where one stands relative to the others.

The figures in this exchange can be read the same way. The table, the network map, the tensegrity, the morphing model, the compass, the tesseract and the mandalas are each a different finger, pointing at the same debate from a different direction. Each brings out something the others hide: the table the counts, the tensegrity the tensions held apart, the mandala the empty sectors, the regional comparison the weighting by alignment. None of them is the moon. What coherence there is comes from moving between them, and from noticing where they disagree.

The ecological framing adds what a list of perspectives lacks. In an ecology, diversity sustains the whole, and a monoculture is fragile however efficient it looks. The equivalents here would be a single authoritative visualization, a single framing of AI governance, or a union of AIs singing in unison. Each would be easier to grasp and more brittle for the same reason. The value of an ecology of perspectives lies in its redundancy and its friction: each view keeps the others honest, as the evidence test kept my own map honest.

It also gives the earlier question a direction. A facilitation of the General Assembly debate would not need to offer the one true map. It would need to offer several, and help participants see where they themselves stand within them.

An AIvision contest would be more than a novelty, because the Eurovision format already contains most of the tensions this exchange has been mapping.

A precedent exists. Dutch broadcaster VPRO ran the AI Song Contest in 2020, explicitly modelled on Eurovision. Teams of people and machines composed entries, and the Australian team Uncanny Valley won with a song built partly from the sounds of koalas and other wildlife. That first contest settled one design question early: it judged human–AI collaboration rather than machines alone.

Why Eurovision is the right model

Eurovision is the best-known public demonstration of the disconnect we discussed. It is nominally a contest of merit, yet its voting is famously shaped by neighbourhood, diaspora and alignment. Network analyses of its votes regularly find stable blocs, in effect a compass of cultural affinity. It also divides judgment between national juries and public televotes, which often disagree: expert assessment set against popular sentiment. And it has a rule against political content that is never quite kept.

An AIvision contest would inherit all of this. That is its main value. It would make visible, to a very large audience, the gap between the merits of AI and the alignments through which it is judged.

What it could reveal about the nature of AI

• Who is the performer? If entries are national, does a country enter with its own model, or with one built by a foreign company? The contest would stage the sovereignty question raised in the debate, France's "coalition of independents" and India's right to its own models, as a matter of who may represent whom.
• Who judges? Adding a third vote, an AI jury, alongside national juries and the public would produce the most informative result: a measured gap between machine and human taste. If AI juries favoured songs resembling their training data, which is dominated by English-language and Western music, the digital divide would show up as a scoreboard.
• Rules as governance in miniature. Eurovision needs rules on live vocals and originality. AIvision would need rules on disclosing training data, consent for voices that are imitated, and perhaps each entry's computing and energy cost. These are the safety-rules, rights and energy-footprint issues of the debate, made concrete. The energy footprint, which leaders barely mentioned, could be displayed beside each entry's score.
• Humanoid robots. Embodiment shifts the question from composition to presence. A song contest is about voice, body and feeling, so robot performers would test whether audiences grant feeling to machines. That is the "human-centred" question in its most direct form. It would also stage the uncanny valley literally, as the 2020 winners' name anticipated.

Risks

The obvious danger is capture. Technology companies at the SDG Media Zone and in the Security Council meeting show the appetite for prominent stages, and an AIvision contest could become a showcase for the race to win rather than a reflection on it. It could also trivialise the issues, displace the human musicians whose livelihoods are already at stake, and normalise imitated voices. A design that keeps human–AI collaboration at the centre, with disclosure rules, avoids the worst of this.

The connection to this exchange

An AIvision contest would generate exactly the data whose absence at the General Assembly started this exchange: structured, cross-national judgments, collected in public. Mapping its voting blocs onto the compass or the tesseract, with the issue each song addressed as a further dimension, would be a playful version of the facilitation the Assembly lacked. The song itself is also an apt figure for coherence: harmony is tension held in balance, much as in the tensegrity. A contest asking whether AI can contribute to harmony rather than to noise would at least put the question in a form most people recognise.

Sources:

• AI Song Contest 2020 (Wikipedia)
• Behold the relentlessly optimistic silliness of the 'AI Eurovision' winner (TechCrunch)
• AI Song Contest (VPRO)
• EXPLAINER: UNGA 81: 5 Key Takeaways From General Debate (Leadership)

Two features of the European case come before the AI variant, because they explain why harmony in the musical sense has not become harmony in the political sense.

What the European models have not produced

• A contest is not a composition. Eurovision presents national songs one after another and ranks them. Nothing in the format asks any entry to fit with any other, so the result is a sequence of solos. Its voting blocs are mapped and discussed every year, yet nothing in the format learns from them; they confirm affinities rather than working on them.
• An anthem without words. The Anthem of Europe is Beethoven's setting of the "Ode to Joy", adopted without Schiller's text, so that no language would be favoured. Harmony was achieved by avoiding articulation. When the symbols of union, the anthem included, were proposed for the Constitutional Treaty, they were dropped after its rejection. Only sixteen member states signed a declaration recognising them alongside the Lisbon Treaty.
• Unison drift. Since Eurovision relaxed its language rule in 1999, many entries have been sung partly or wholly in English. Freedom of expression produced convergence rather than diversity. That is the same pattern feared for AI trained mostly on English-language data.

Europe, then, has a borrowed harmony that is played but not sung, and a popular contest that stages difference without composing with it.

The General Debate as a song contest without a song

The General Assembly has the same structure: 194 national statements delivered in sequence, each addressed largely to its own audience, with no mechanism by which any statement must fit with another. It is Eurovision without the voting. The Assembly President's summary then plays the role of the anthem, a statement of common themes that no delegation composed. The missing facilitation this exchange began from is precisely the absence of composition.

The larger challenge: from contest to counterpoint

Music offers a sharper criterion than "harmony". Unison is many voices singing the same line, which feels coherent but loses all difference. Counterpoint is independent lines, each with its own integrity, written so that they sound together. Dissonance is not failure: it creates tension that the music resolves, or deliberately leaves open.

Harmonising given melodies is something AI already does well; completing a Bach chorale from its melody is a classic exercise for it. A variant aimed at governance could reverse the Eurovision premise. Each entry could be required to incorporate a theme from every other participant, and the score would reward how well an entry integrates others' material rather than how well it stands out. Applied to positions rather than melodies, the same exercise would show where national lines can be set in counterpoint and where the dissonance is irreducible. Both findings matter: the second marks where negotiation has no common key.

A union of AIs

This sharpens the question in your essay. As AIs interconnect, their collective form could take three shapes:

• Unison. Systems trained on similar data converge, giving coherence without differentiation. This is the most likely default, and the least visible as a problem.
• Contest. National or corporate AIs compete, which is the "race" framing heard at the Assembly.
• Counterpoint. Distinct systems keep their differences but are composed to work together.

The Borromean rings you use point to the third form. No two rings are linked, yet the three cannot be separated. The tensegrity has the same property, with struts that never touch held by tension, and so does counterpoint, whose voices never merge. All three describe integrity without fusion, which is perhaps the general form of what neither Europe nor the UN has managed.

There is a further risk the anthem foreshadows: a wordless harmony. A union of AIs might coordinate through protocols that people cannot hear, coherent among the machines but inaudible to the people governed by it. The Assembly President put the limit plainly in his closing remarks: "No machine possesses democratic legitimacy. No algorithm inherits constitutional responsibility."

Your essay's emphasis on self-reflexivity is the counterweight. The union would need to make its own harmonies audible, and open to critique, to the people outside it. The small instance in this exchange, the evidence test that revised my own map, is the kind of reflexivity that would have to be built in from the start rather than hoped for.

Sources:

• Imagining a Future Union of Artificial Intelligences (2024)
• The Takeaway Day 6: General Debate ends with call to turn words into action (UN News)