The premise nobody contested
Six delegations, a UN agency, Google and a youth delegate spent seventy-five minutes agreeing on one proposition: that AI will widen the gap between countries unless something is done about it. That consensus is worth recording, because it is not the room’s natural instinct. Nobody argued the technology would lift all boats on its own. The Chairman of Pakistan’s Prime Minister’s Youth Programme put the caution in the same breath as the promise.
The sharpest statement of the divide came not from a member state but from the UN. Jennifer Louie of UNDP framed the whole session in a single question, then answered it with the only structural number anyone offered.
“This country” was said in a conference room in New York. Louie was careful to call it a description rather than an accusation, and the consequence she drew from it is the one that matters for a development forum: for almost every state in the building, the question is not how to build a frontier model but whether the frontier ever arrives, and who fixes it when it is wrong.
The counting angle: an announcement is not an outcome
This is where the session thinned. The delegations arrived with substantial numbers, and nearly all of them counted inputs. Pakistan reported more than 1.6 million laptops distributed, over 73,000 young people trained in AI, blockchain and data science, and a Digital Youth Hub with more than 800,000 registered users and 1.6 million downloads. China reported manufacturing efficiency up “over 30 percent while cutting defects by half,” digital public services reaching “more than 90% of rural populations,” and youth innovation funds behind “over 10,000” AI startups. Azerbaijan reported a national AI strategy for 2025 to 2028 and a council coordinating 58 initiatives. Tajikistan reported the region’s first national AI strategy, running to 2040, and a regional AI centre in Dushanbe whose construction began on 9 June 2026.
Laptops distributed, users registered, initiatives coordinated, ground broken. These are real programmes with real budget lines behind them, and counting delivery is a legitimate thing for a government to do. But not one speaker closed the loop to an outcome: what a trained cohort earns, what a public service costs or delivers after the model is installed, what moved on any SDG indicator. In seventy-five minutes on AI and the Sustainable Development Goals, no speaker named a single SDG indicator. The word “indicator” was used once in the session, by Kazakhstan, and in a rhetorical sense.
The most telling evidence was an arithmetic collision the room did not notice. Pakistan’s national AI policy was cited twice from the same stage, half an hour apart, against the same deadline. The Chairman of the Youth Programme said that “by 2030, Pakistan aims to train more than 2 million AI professionals.” Google’s representative then described his company’s AI skilling work as “aligned to the government’s visionary national AI policy to train 1 million AI professionals by 2030.” One target, one deadline, two figures a factor of two apart, neither queried. A target that can double between two speakers at its own launch event is not yet functioning as a measurement.
The panellist who named the gap
The most useful thing said in the room was said by the person with the least incentive to say it. Louie, describing what UNDP has learned from roughly 25 AI landscape assessments and work across more than 50 countries, put the measurement problem plainly: governments are being asked to act before there is anything to act on.
That is the session’s own diagnosis of itself, offered from the panel. Her second point is the one governance frameworks tend to miss. AI governance, she argued, is not decided at a summit; it accumulates out of procurement decisions, vendor contracts and software defaults made by people who do not think they are making policy. “Governance is not only decided in one moment,” she said. “We are being decided in many smaller choices.” She asked the room for a show of hands on who felt able to make confident AI decisions for their entire country. One person raised a hand.
Louie also brought the session’s only outcome-shaped delivery figures, and notably they were about resilience rather than innovation: UNDP’s RapidA system, combining AI with satellite imagery and community surveys, has “delivered post-disaster assessments within 72 hours in more than 20 countries,” and in Afghanistan helped restore infrastructure access to 350,000 people after the 2025 earthquake. She was equally direct about the harm side, and it was the one moment where the inclusion language acquired teeth: youth unemployment, she said, is pushing digitally literate young people in the global South into cybercrime, “because they don’t have other jobs, and because they are digitally literate.”
Kazakhstan brought the one checkable number
One delegation cited a figure produced by somebody other than itself. Kazakhstan’s Permanent Representative reported that his country “ranks 10th globally in the online services index and 24th in the eGovernment development index,” a result from the UN E-Government Survey, which is independently published, comparable across 193 states and checkable against the record. It checks out.
The distinction is not pedantry. Every other national claim in the room was a government reporting on itself with no external referent, which is exactly the pattern this Forum’s VNR sittings spent a week trying to move past. He also gave the session its most quietly revealing line, framing AI and its infrastructure as “one of the defining indicators of economic development, technological advancement, and national competitiveness.” AI has become the thing you measure a country by, before there is much agreement on how to measure it.
What went unmentioned
A session convened on innovation, inclusion and resilience did not mention the resource cost of the compute it kept asking for. Across the full transcript there is no reference to carbon, to emissions, to a footprint, or to energy at all. “Water” appears twice, both times incidental. This was eight days after a side event in the same building where the Director of UN University’s water institute told an HLPF audience that AI is a physical product with a water, carbon and land footprint, feeding the Secretary-General’s AI Environmental Transparency Initiative. Six delegations called for more data centres and more compute in developing countries. None costed them.
The gap was named, though not from the podium. An AI consultant taking the floor from the audience described what he sees when institutions come to him wanting AI: “they don’t have the internal capacity to assess the positive impacts or the fruitful impacts, and it’s sort of a trend right now that everyone wants to jump onto this AI train without fully understanding the positive and the negative impacts of it.” The room did not return to the point. A young questioner from Myanmar made the other omission concrete, asking how marginalised young people in conflict-affected countries reach the technology at all when their national development has gone backwards. He got sympathy rather than a mechanism.
Our read
Two speakers reached for the same metaphor about measurement, and it is the tell. Pakistan’s Youth Programme chairman closed his keynote by declaring that “the true measure of artificial intelligence will not be the sophistication of its algorithms, but the lives it improves.” A week earlier, at the Forum’s other AI side event, the Philippines had said that “true innovation is not measured by the sophistication of algorithms, but by the breadth of their benefits.” Both are correct. Neither is a measure. The sentence performs the act of measuring while proposing nothing to count, and the fact that two delegations arrived at it independently suggests it is now the standard way of gesturing at accountability without accepting any.
The honest summary is that this was a session where the diagnosis outran the instrumentation. The diagnosis was good: the compute is in one country, the models come from two, the divide is real, and the Permanent Representative of Pakistan was right to close on the observation that “the global AI divide cannot be bridged through declarations alone.” He is also, on the evidence of his own session, describing the problem from inside it. The room brought commitment, budget lines and headcounts, and no way to tell in 2030 whether any of it worked. Louie said the anchor point does not exist yet. The rest of the programme proceeded as though it did.
Why it matters for the SDGs
SDG 9 was the stated anchor and was named from the podium more than any other goal, with SDG 4 (skills and AI literacy) carrying most of the actual programme content and SDG 10 supplying the argument, since the AI divide is an inequality claim before it is a technology one. SDG 17 sat underneath every appeal for technology transfer and capacity building. The measurement question is the one to carry forward. SDG 9 has indicators, several of them, covering research intensity, technicians per capita and mobile network coverage, and a session convened on SDG 9 and AI could have reported against them. It reported laptops instead. That is not dishonest, but it is the difference between accountability and inventory, and the gap between the two is where the 2030 deadline gets lost.
Watch & read
- UN Web TV, recording of the side event (15 July 2026).
- UN E-Government Survey, Kazakhstan country data, the source behind the session’s one externally verifiable claim.
- AI Hub for Sustainable Development, the UNDP and Italy initiative cited from the panel · Prime Minister’s Youth Programme, Pakistan.
- The evidence base the panel kept invoking: the Independent International Scientific Panel on AI and the Global Dialogue on AI Governance, which had concluded in Geneva days earlier.
- The Forum’s other AI side events: AI for social inclusion · Humanity at the center · Digitalization, statistics and AI.
- Full HLPF 2026 coverage.
Quotations are lightly edited from an automated (Otter.ai) transcript of the UN Web TV recording and should be read as close paraphrase. The moderator, the delegates of Kazakhstan, China and Azerbaijan, and all questioners from the floor are cited by role or country. Figures are as speakers and governments reported them about their own programmes and were not independently verified, except where the page says otherwise. Pakistan’s national AI training target was stated twice from the podium, as 2 million and as 1 million by 2030; both figures are reported here as spoken rather than reconciled.