AI and Human Development: Convenings with Purpose

FOREWORD

“MORE HUMAN INPUT TO AI SYSTEMS WILL NOT AUTOMATICALLY PRODUCE MORE HUMANE TECHNOLOGY. BECAUSE HUMANITY IS CAPABLE OF BOTH EXTRAORDINARY COURAGE AND TERRIBLE CRUELTY”

The most consequential dimension of artificial intelligence may be the one we have spent the least time examining: its implications for human evolution and development.

What defines our humanity? What does it mean to be human in the age of AI?

Much of the global conversation on AI thus far has rightly focused on the science and how it should be governed. These global convenings were created to examine AI through a third, less explored lens - one centred on what it means for people, societies and the future of humanity in all its diversity.

Building on the work of the Independent Scientific Panel and the Global Dialogue on AI Governance, this year we took that question across four global convenings in Cairo, Egypt; Mexico City, Mexico; Xi’an, China; and Rome, Italy, bringing together an exceptional cross-section of leaders and thinkers from government, science, faith, business, the arts, Indigenous communities and universities, with young people present throughout. The series now culminates in New York during the Eighty-First session of the United Nations General Assembly. Across every convening, we asked the same question: what does it mean to be human in the age of AI?

Participants pushed into the deeper terrain of AI: its implications for human capability and culture, for relationships and agency. They asked what happens when technology becomes part of how we learn and how communities carry knowledge forward. They explored what this means for the social fabric that allows people and societies to flourish, and how its opportunities and risks may be experienced differently across genders and generations. The meeting also probed the intersection with the quickly shifting economic and jobs landscape.

Looking at AI through the lens of human development opens a wider horizon. It asks us to consider how technological change interacts with the values, relationships and sense of purpose that shape human life. It also places human rights and human dignity firmly in view, asking how AI can expand human agency while respecting the rights and worth of every person. Across cultures and traditions, different understandings of dignity, community and our place in the world have evolved over centuries. Yet in the Sustainable Development Goals, the world found common ground in a single promise: to leave no one behind. That breadth of human experience, and that shared promise, belong at the heart of how we think about the future AI is helping to shape.

This Report, reflecting the perspectives of participants that emerged across the diverse convenings, carries that inquiry forward. It captures the questions raised by the participants that informed the convenings and will shape a much wider conversation as this work continues in countries around the globe. That said, the pace of change signals the urgency for that conversation to take place. The imperative is to shape choices while they are still ours to make.

AI is already changing - and will shape - what people can do. The question that now deserves our full attention is what it will mean for who we become and the world that will be shaped by this new era of technology. I believe the choice is ours to make, lest it be made for us by a few privileged actors.

Amina J. MohammedDeputy Secretary-General of the United Nations

PREFACE

Every era is shaped by the tools it creates but ultimately defined by the values it upholds.

That idea has been a touchstone for our AI and Human Development journey. Guided by the Deputy Secretary-General, we took the conversation to four places where ancient civilizations once faced turning points of their own: Cairo, Mexico City, Xi’an, and Rome.

History grounds us. It reminds us how societies have evolved and changed, and how some have disappeared.

Now we find ourselves at another turning point.

Artificial intelligence is advancing faster than our ability to make sense of what it means for us. So, we invited people from different disciplines, cultures and traditions to step beyond familiar debates about technology and consider something more fundamental: what kind of human future do we want to create?

The conversations were deliberately open-ended, with no prescribed answers and no expectation of consensus. What emerged was a rich, and sometimes uncomfortable, exploration of human dignity, agency, relationships, culture, creativity, education, faith and, ultimately, what it means to be human.

Throughout the conversations, one simple word kept returning: “we.”

In every room, as groups spoke about how we should steer this technology, someone else would ask who that “we” includes. Whose languages and faiths? Whose ways of knowing? How are women and young people, so often overlooked, part of it? And how should that “we” be much wider, encompassing the communities whose lives and futures will be shaped by decisions they may have had little part in making?

These questions are at the heart of this report.

The convenings were built with Member States and partners who welcomed the conversation into their cities and institutions. Our immense gratitude goes to the Government of Egypt and Al-Azhar University; the Government of Mexico; the Government of China and the People’s Government of Shaanxi Province; the Government of Italy; and the Holy See, the Office of the Presidency of the Governorate of Vatican City State and the Pontifical Gregorian University.

We thank the Ministries of Foreign Affairs of Egypt, Mexico, China and Italy, and their Permanent Missions to the United Nations in New York, as well as the Permanent Observer Mission of the Holy See to the United Nations, for their support throughout.

The convenings were greatly enriched by the generosity of the Grand Egyptian Museum in Cairo, Egypt; the National Museum of Anthropology in Mexico City, Mexico; Emperor Qinshihuang Terracotta Warriors Museum in Xi’an, China and the Vatican Museums in Rome, Italy.

We also thank the United Nations Resident Coordinators and country teams in Egypt, Mexico and China, who made each convening possible on the ground. We are grateful to the moderators and provocateurs, as well as colleagues across the Executive Office of the Secretary-General and the United Nations Office for Partnerships who shaped the programme and captured the discussions.

We owe a special debt to colleagues across the wider United Nations family who were indispensable to these convenings, from Cairo to Rome. Above all, we thank everyone who gave these conversations their time, their trust and their candour.

This report does not offer a final verdict. It captures a conversation that is still unfolding and will be taken to communities far and wide ahead of the 2027 SDG Summit. This report’s cover is a young African girl. Look closer and she is made entirely of the symbols of the 17 Sustainable Development Goals. Her future will be shaped by the tools we build. Whether it is defined by the promises we have made to her is up to all of us.

I invite you to read it in that spirit, and to add your own voice to the “we”.

Annemarie HouExecutive Director, United Nations Office for Partnerships

REPORTING ON FOUR CONVENINGS

WHAT WE TALKED ABOUT

Artificial intelligence (AI) is reshaping not only what we can do, but how we understand who we are.

As AI systems match or surpass human capabilities in certain domains, a deeper question comes into view: if machines can reason, communicate and act with apparent autonomy, what is it that makes us human?

Much of the global conversation on AI has so far centred on two dimensions. The first is what AI can do: the rapid expansion of its capabilities and their potential to advance science, productivity, public services and other areas of human activity. The second is how AI should be governed: how societies manage the risks that accompany these capabilities and establish rules for their responsible development and use. The Pact for the Future and its Global Digital Compact, adopted in 2024, form part of that wider effort to build common ground around AI governance and capacity.

This report, and the convenings it summarizes, focused on a third, less explored dimension: what AI means for human development itself. As these systems become more deeply embedded in everyday life, their implications extend beyond jobs, education and economic activity. They reach into social relationships, cultural values, human agency, purpose and well-being, raising questions about how people live together and what conditions allow them to flourish. That inquiry is closely connected to the promise of sustainable development.

At their core, the Sustainable Development Goals (SDGs) strive to build a better future for both people and the planet, guided by the principle of leaving no one behind. Weaving a healthy and cohesive social fabric — one that integrates AI in ways that support human connection, purpose and well-being — will be essential for communities to flourish.

United Nations Deputy Secretary-General Amina J. Mohammed convened the four global conversations described in this report between June and September 2026, putting this question at the centre of a global conversation. Organized by the United Nations Office for Partnerships in collaboration with Governments and their institutions. Participants came from government, international organizations, science and academia, technology and business, civil society and philanthropy, faith communities, the arts and culture, Indigenous and Nomad communities, and young leaders, representing every region of the world.

Individual meeting sessions were varied. Discussions on civilization explored how AI is reshaping the foundations of human society, the social fabric that connects people, and how our relationships with one another may evolve in the age of AI. Sessions about creativity and culture focused on how AI is influencing cultural expression, what role cultural institutions can play in stewarding heritage in a digital world, and what elements of creativity, curiosity, and art remain uniquely human as well as the possibilities AI may open for human expression. On education and Global Goals, participants debated how AI, robotics and other emerging technologies are changing what people need to learn, what capacities education should develop, and what human development should enable people to become. Discussions around spirituality and faith asked how AI is reshaping understandings of human dignity and how teachings from faith traditions can guide societies in safeguarding conscience and truth.

Across different cultural, philosophical and societal traditions, the conversations approached humanity through reason and imagination, emotion and empathy, vulnerability and care, individual dignity and collective responsibility. They drew on ways of thinking that locate human identity not only in individual capacities, but through our relationships with one another, our communities, our cultures and the world we share. At each convening, the same question was put to the room: is AI serving humanity?

One word surfaced time and again throughout the convenings: “we”. Participants said “we” must decide what AI is for. But they also challenged who that “we” includes, pointing to the people and communities still largely absent from the decisions shaping its development and use. Much of the discussion focused on the unequal distribution of power and opportunity across communities and countries, with women and young people among those least represented. The test of whether AI is aiding human development also shifted from whether people can get access to AI, to whether they can still manage without it, and whether they are more capable because of it.

Discussions of embodied AI raised a further possibility: that increasingly human-like machines may become a mirror through which we examine more closely what we value about being human. At the same time, discussions returned to the importance of lived experience and mutual human relationships, asking what aspects of life should remain centred on them. Some proposed treating certain parts of life as “sanctuaries” that should remain protected from AI. Others warned that the language we use to talk about AI — that it “thinks”, “decides”, or can be “trusted” — obscures the responsibility of the people and institutions who design and deploy these systems.

The same concern surfaced in discussions of governance. Contributors agreed that the United Nations has global legitimacy, but limited powers of enforcement. There was also a strong sense of urgency. Several contributors argued that the window for shaping the trajectory of what AI means for people and societies may be measured in months rather than years, far shorter than the time States and multilateral institutions usually need to negotiate collective solutions. The resulting gap between the pace of change and the capacity of institutions and communities to respond was a concern that recurred across the convenings.

Taken together, these global conversations offered no single definition of what it means to be human. What they revealed instead was a rich plurality of perspectives, while pointing toward a shared imperative: as societies shape the present, we must remain intentional about the kind of world — and the kind of human future — we want to create.

WHAT STOOD OUT

What makes us human is lived, not performed. The better AI becomes at imitating human behaviour, the less useful performance becomes as a measure of what distinguishes us. Across the convenings, participants located our humanity instead in what has to be undergone rather than produced: presence, vulnerability, grief, love, and the responsibility that comes from being answerable to someone else. It followed that certain domains of life — such as intimacy and the care of children and older people — might need to be treated as “sanctuaries”, placed beyond AI’s reach altogether.

The more immediate risk from AI is that people become less practised at exercising their own agency. Agency is rarely surrendered in a single decision; it goes incrementally, through convenience and through choices made elsewhere. Participants pointed to early evidence that delegating work to AI changes how people think and behave, and to the possibility of “never-skilling”, in which a capability is displaced before a person has ever formed it.

Human capability, not AI system capability, is the measure that matters. Rather than evaluating AI systems on their own terms, participants proposed three tests for whether AI is aiding human development: can people understand and challenge these systems, can they still manage meaningfully without them, and are they more capable over time because of them?

The language used to describe AI can blur human responsibility. Machines are said to “think”, to decide, to hallucinate, to be evil, or trustworthy. Participants argued that these are not neutral descriptions: the words we use shape how people understand agency, judgement and responsibility. When systems are described as though they act independently, human choices can disappear from view.

The values embedded in AI reflect a very narrow slice of humanity. A small number of teams are deciding what the world’s most widely used systems will encourage and treat as normal — a process one participant described as “shockingly illegitimate”. Across the convenings, participants argued that this is not only a question of who gets consulted, but whose understanding of dignity, conscience, care and a good human life is reflected in the systems people increasingly live with. Philosophical, Indigenous and religious traditions were among those seen as largely absent from that process.

A society can use AI fluently and still lose the ability to think in its own terms. Participants described this as a loss of epistemic sovereignty: the capacity to produce, interpret, contest and pass on knowledge in one’s own language and frame of reference, which can decline even as adoption rises. The same asymmetry runs through culture, where the question was whether there can be culture without consent, and through the material base, where the energy, water, land and minerals behind these systems are drawn disproportionately from the communities least represented in what they produce.

The window for shaping what AI means for people and societies may be measured in months rather than years. In every room, participants judged that window to be open but closing quickly. That creates a profound mismatch between the pace of change and the time people, communities and institutions need to adapt — a gap participants returned to at every convening.

01THE QUESTION OF “WE”

Every convening arrived, sooner or later, at the same pronoun. Participants would say that we have to decide what AI is for, and sooner or later somebody else would ask who exactly that “we” includes. The challenge was first raised at the Cairo convening by Graça Machel, who has spent a lifetime in global public life.

“HUMAN BEINGS SHOULD NOT BE TRANSFORMED INTO DATA. THEY SHOULD BE SEEN AND ENGAGED AND INVOLVED IN THE WHOLE OF WHAT A HUMAN BEING IS. I COME FROM A COMMUNITY THAT HAS BEEN TREATED AS BENEFICIARIES. WE ARE TIRED OF THAT. LET US UNPACK THE “WE”, AND SAY WHAT THESE DEVELOPMENTS MEAN TO THE DIFFERENT “WE’S”

In Mexico City the same objection arrived through economic history, by way of Raúl Prebisch’s concepts of centre and periphery. The region supplies the data and the talent; others control the systems and capture the value. AI technology is new; the power dynamic is not.

By the time the series reached Xi’an, United Nations Deputy Secretary-General Amina J. Mohammed had made the thread explicit from the start: “We say that we must govern AI and decide what it’s for. But who is the ‘we’?”

The values and refusals encoded in the leading AI systems are written down by small teams inside a handful of companies. One person working on such a team described the process to the room and asked participants to supply something better.

“THE WAY THESE DECISIONS ARE BEING MADE IS THROUGH A KIND OF SHOCKINGLY ILLEGITIMATE PROCESS, WHERE A SMALL NUMBER OF PEOPLE AT A SMALL NUMBER OF TECH COMPANIES...WHO ARE SMART, WELL-MEANING PEOPLE — I WORK WITH THEM — ...ARE WHOLLY UNEQUIPPED TO BE MAKING THESE DECISIONS”

That request has been repeated in different words at every convening since: a call for stewardship that is international and intergenerational, that does not wait for regulation to catch up, and that draws on a much wider range of human experience, including faith, ethical and philosophical traditions, as well as law.

In Rome that call met an institution that had just addressed it. Participants gathered just months after Pope Leo XIV published his first encyclical letter, Magnifica Humanitas, on safeguarding the human person in the age of AI. In it, the Pope centres the dignity of the person in how technological progress is judged, argues for freeing technology from monopolistic control, and asks that the plurality of human cultures and ways of life be restored rather than flattened. Several people in the room read the encyclical as a demonstration that the question of who decides is not the property of governments and technology companies. Most of the world’s population belongs to a faith tradition, which makes faith communities important stakeholders in AI, whether or not anyone consults them. One participant argued that communities should bring their own moral vocabulary — encompassing justice, compassion and humility ­— into the design of AI systems rather than wait to be invited, and that communities must be empowered to build alternatives that reflect and serve their values.

Widening the “we” is not the same as averaging it, and participants were careful about the distinction. Machines are learning from vast yet profoundly uneven records of human behaviour, including the parts of it nobody would choose to pass on.

Ultimately, what was being asked for was not a poll of humanity on every decision about AI’s development. Instead, the demand was for a meaningful process in which the people who live with the consequences of AI help define the problems and the solutions. It was also a demand to keep human responsibility visible: behind every consequential decision made with or through an AI system, there should be a person or institution answerable for it.

02SOMEONE ELSE’S IMAGINATION

Development policy tends to treat AI as a question of access: some people have the tools and others do not. The conversations repeatedly challenged that framing. Access alone says little about how much power people and societies have over the technologies they use. One participant divided the AI stack into four layers: compute, data, models, and use. The first three, she argued, sit largely outside of her region. Use is the exception. Drawn this way, the gap stops looking like a lag that adoption will eventually close.

The consequence goes beyond economics. A society can be a fluent and enthusiastic user of AI and still lose the ability to produce, interpret, contest and pass on knowledge in its own terms. One participant called this “epistemic sovereignty”. Timor-Leste was offered as an illustration: very high reported use of AI applications like ChatGPT alongside very low readiness to leverage AI for economic, social or cultural advancement. “None of that signals readiness,” the participant said. “It shows a young population reaching for a free product over borrowed funds.” The arithmetic underneath is stark. As one participant explained, against an international benchmark of two per cent of monthly income for affordable internet connectivity, two gigabytes of mobile data costs around 10 per cent of monthly income in Solomon Islands and closer to 12 per cent in Papua New Guinea, before anyone buys a device or pays an electricity bill. The challenge runs deep: 2.2 billion people were offline last year, and more than a quarter of humanity is still waiting for basic internet connectivity. throughout the follow-up survey, 87 per cent of participants projected that inequalities will be widened between countries and within them as a result of AI (see Annex 2).

What to do about this inequity, and what meaningful agency looks like in response, was less settled. One answer is to build. Egypt has a national large language model and a tutor that understands the colloquial Arabic a child actually speaks at home, presented as an act of dignity and sovereignty. Chile’s national AI centre and thirty partner institutions are training an open-source model on regional data, extending it to Indigenous languages with the consent of the communities concerned. A Chinese government official at the Xi’an convening argued that releasing open weights and investing in capacity-building reaches the same destination faster. The other answer questions whether a sovereign model is the best use of the resources, energy and computing a country actually has, and proposes instead that countries specialise in adapting and running existing models at the edge, on devices people already own, while protecting critical national data through “data embassies” held by trusted international custodians.

The same question runs through culture: can there be culture without consent? The discussions contrasted the Voyager 1 Golden Record — 90 minutes of music chosen deliberately by a few people working to represent humanity — against AI compilations assembled by default from whatever happened to be digitised. The practical conclusions pointed in opposite directions. Some argued that cultural institutions should become data factories, digitizing oral traditions and low-resource languages so that they exist in the record at all. Others reported Indigenous and Nomad knowledge holders refusing to hand ancestral knowledge to any AI model, sovereign or otherwise. And others pointed to something that cannot be captured in a record at all — large portions of the human experience and wisdom are learned not by documenting them, but by living them.

A further asymmetry runs through the people these systems are built by and learn from. Women and youth remain heavily under-represented in AI research and in the teams that design these systems, and the data those systems learn from carries the same omission. As the United Nations Deputy Secretary-General put it in Xi’an, “feed an algorithm decades of unequal hiring practices and it will learn that a man is the stronger candidate; train it on a world in which women have had less access to wealth and it will read that exclusion as good financial risk management; build it with medical data centred on male bodies and female symptoms will be ignored. Without deliberate correction, AI gives old discrimination new speed and authority.” The harms are not only statistical. Those in the discussions also noted that the overwhelming majority of non-consensual sexual imagery produced by AI systems targets women and children, and that online spaces have become another territory in which violence against women is carried out.

03WHAT GETS ADDED AND SUBTRACTED

Participants discussed how the education sector has produced evidence pointing in both directions. On one hand, a randomised trial of an AI tutor in 1,700 schools in Sierra Leone reported gains equivalent to 1.7 additional years of learning in a short period — an experiment since replicated in Italy. Yet on the other hand, another study found students using an AI assistant improved while they had it but then did worse than students who had never used it once it was taken away.

The difference between those two results is not the technology, but rather whether the technology adds to capability or subtracts from it. This led the conversations back to a different measure of progress: not what an AI system can do, but what happens to human capability when people use it. Professor Joanne Lim of the University of Nottingham Malaysia proposed three tests, which were developed further in discussion, for whether AI is aiding human development.

Another participant at the Xi’an convening set out three levels of consequences for capability: individual de-skilling, where a person loses an ability; structural de-skilling, where an institution loses it; and never-skilling, where a capability is displaced before it has ever been formed. Never-skilling may matter most for human development, the speaker argued, and it appears in none of the indicators currently collected. A related framework puts AI capability and human capability on two axes. Both rising is augmentation. AI rising while human capability falls is dependency. Human capability rising while AI is refused is resilience bought at the cost of opportunity. Both types of capability falling is exclusion. Through the follow-up survey, 93 per cent of participants expressed their belief that AI capability will continue to rise, but only 35 per cent shared the optimistic view of augmentation that human capability would also rise in parallel (see Annex 2).

“THE BRAIN FOLLOWS THE “USE IT OR LOSE IT” PRINCIPLE. IF YOU USE AI TO REPLACE YOUR DECISION-MAKING, THE CONNECTOMES OF DIFFERENT BRAIN AREAS WILL BECOME WEAKER AND WEAKER ...DO NOT USE AI TO REPLACE YOUR DECISION-MAKING. USE AI TO ACCESS AND ANALYSE GLOBAL INFORMATION, TO ASSIST AND OPTIMISE”

Two practical conclusions followed. The first is that the teacher’s job changes from transmitting knowledge towards mentoring, judgement and character, which makes teachers more necessary rather than less. On timing there was no agreement: some argued for keeping AI out of the earliest years to protect cognitive formation, others for introducing it early and redesigning assessment around oral defence, debate and memory.

The second practical conclusion concerns where money goes. York University Professor Steven J. Hoffman observed that billions of dollars are being invested in technological development without the same level of investment in learning how to manage these developments effectively. He argued that the social sciences and humanities — the disciplines that help us understand our humanity — are comparatively under-funded and further threatened as student enrolment in them falls. Youth delegates made a related point themselves. A manifesto written by forty-eight university students in Malaysia over twelve weeks asked that students be treated as participants in the design of educational AI rather than as subjects of its policing. Amir Siraj, an astrophysicist and concert pianist completing his doctorate at Princeton University, argued that confusion is not an obstacle to understanding but the place where understanding gets built, and asked what happens to curiosity when the destination can be reached without the journey. At the Rome convening, Sister Nancy Usselmann turned the same concern into four questions that a student or teacher can ask as a practice of “AI mindfulness” before reaching for the tool, rather than afterwards.

04WHAT REMAINS HUMAN

“THERE IS NOTHING LIKE TRUSTWORTHY AI. PLEASE DON’T BLUR THE LINE BETWEEN THE MORAL SPHERE AND THE TECHNOLOGICAL SPHERE ...YOU DON’T TRUST MACHINES. YOU RELY ON MACHINES ... I DON’T WANT TO TRUST AI. I WANT TO TRUST THE HUMANS WHO BUILD AI AND THE FRAMEWORKS AROUND AI”

A thread running through the Rome convening was that the anthropomorphizing vocabulary we now frequently use to describe AI can transfer moral weight from people to machines. For example, AI systems are said to think, to decide, to hallucinate, to be trustworthy. Yet trust belongs to the moral sphere; at best we can say that AI is reliable.

“THE DANGER ISN’T THAT A CHILD WILL LOVE IT. THE DANGER IS THAT THE CHILD WILL LOVE SOMETHING THAT WILL NEVER LOVE IT BACK. AND TO THAT CHILD, THAT WILL FEEL NORMAL”

“A PROFILE PRODUCED BY ALGORITHMS IS NOT A PERSON, BUT THE RESULT OF STATISTICALLY SIGNIFICANT CORRELATIONS THAT DETECT IT WITHIN ENORMOUS MASSES OF QUANTIFIED DATA... THAT PROFILE DOES NOT DELVE INTO THE SINGULAR SITUATION OF EACH SUBJECT, NOR INTO THEIR RELATIONSHIPS AND DIFFERENCES WITH OTHER PEOPLE”

The same concern about keeping human responsibility visible ran through discussions of values and red lines. Over the lunch break in Cairo, a few participants wrote 10 items into a notebook – four values and six red lines intended to guide the development of AI – and the list has been added to at every convening since. The values were human dignity and agency, truth and honesty, the common good, and connectedness. The red lines: no autonomous use of force; no exploitation of children; no impersonation of a specific person without their consent; no covert systems where someone could reasonably expect a human; no mass surveillance or social scoring; and no consequential decision without a route back to a human being (see Annex 2).

At the Xi’an convening, one speaker applied similar logic to the parts of human life that may lose something when they are turned into measures. Rather than prohibiting outcomes, she proposed prohibiting measurement: no AI benchmarks, no leaderboards and no optimisation targets for emotion, empathy, intimacy, personality, or the care of the young and the old. Her reasoning was that whatever aspect of AI’s performance gets measured is eventually industrialized and automated, and systems will be trained to simulate whatever the benchmark rewards. This turns red lines into “sanctuaries”, areas fenced off not because AI does badly in them but because doing well in them is the risk. In the follow-up survey, participants prioritized protecting 1) decisions involving the use of force, 2) intimate and close personal relationships, 3) early childhood education and development, and 4) judicial decisions and sentencing.

“ONLY HUMANS CREATE. AI GENERATES, BUT IT GENERATES WITHOUT THOSE NECESSARY ELEMENTS NEEDED FOR CREATIVITY, SUCH AS EMOTIONS AND LIVED EXPERIENCE AND SPIRITUALITY”

Underneath these proposals is an argument about presence. A former war correspondent described a moment in Ukraine when he looked into the eyes of a man who was about to shoot him, and said that what frightens him about the future is not that such a moment goes wrong, but that there would be no eyes to meet.

The same sentiment surfaced in Cairo, where one participant noted that a machine can produce language that sounds caring without ever feeling compassion, and can simulate presence without ever being present. An anthropologist in Mexico City made the point from the other direction, about what these systems do to the person in front of them.

Concerns about creativity divided the room in the same way. One speaker drew a line between creating and generating: creation draws on emotion, lived experience and spirituality, whereas an AI system with none of these can only recombine what people have already made. Another participant offered the illustration that his enjoyment of parkour videos was drained when he could no longer tell whether they were real or synthetic. Still others celebrated the role that AI can play in supporting human artistic expression in ways that expand rather than limit what is possible.

The conversations also considered what repeated delegation to AI may do to human behaviour. One participant cited a 2025 study finding that people who hand profit-seeking tasks to AI become more willing to accept dishonest results. Calculation can be delegated; accountability cannot. In Mexico City, the same line of inquiry extended to mental health and emotional dependence, including what prolonged relationships with increasingly human-like systems may mean for well-being and human connection. One proposal was to treat the mental-health effects of generative AI as a subject of assessment in their own right. China’s regulation of emotionally adaptive companion services, introduced within about three years of their launch, was cited as one early response from which others might learn.

05WHO PAYS FOR INTELLIGENCE

Across the four convenings, the discussions repeatedly returned to the point of cost: that AI may feel immaterial to those using it, but its costs are borne “by someone, somewhere.” As reported by one participant, the International Energy Agency estimates that data centres consumed around 415 terawatt-hours of electricity in 2024, roughly 1.5 per cent of the world total, and will double to around 945 terawatt-hours a year by 2030.

The same speaker cited a recent report from the United Nations University’s Institute for Water, Environment & Health (2026) that puts data-centre water use on a path to exceed the basic annual domestic needs of 1.3 billion people, and the land footprint above 14,500 square kilometres, about twice the Jakarta metropolitan area. Voices from Latin America connected these figures to a longer history of extraction and to the communities living where these resources come from. The minerals, water and land that make the technology possible come disproportionately from the territories least represented in what it produces. There can be no technology without nature, as one participant put it.

“THIS IS AN EXTRACTIVE INDUSTRY, BAR NONE. NO ONE ESCAPES IT — NOT A CHILD, NOT A GOVERNMENT, NOT ANYONE”

On the other hand, a contributor, described digitising the second-largest collection of slave records in the Western Hemisphere, work that survived a lightning strike on the original archive and that needs computing capacity to continue. She concluded that helping design better data centres beats abstaining from them. It was further argued that data-centre investment could be made in Africa in ways that also expand the grid capacity communities and economies need. And a country that hosts infrastructure has one option a country that does not host it lacks: it can turn it off.

The costs of AI cannot be settled without better measurement, and current measures are fragmented and incomplete. Several proposals sought to make those costs, and their distribution, more visible. First, green tokens would put a standard accounting of carbon, water and thermal footprint against AI computation, so that efficiency claims can be audited rather than asserted. Second, the Asia-Pacific regional commission has begun putting three questions to every data-centre proposal: what power it needs and who pays for it, what water, land and cooling it uses and what it emits, and what the host economy measurably gains. Those questions are designed to stop neighbouring states bidding each other down. Third, an Intelligence Domestic Product could sit alongside GDP to capture value that AI creates but national accounts do not record, together with an updated measure of inequality. Once the value can be measured, it can be argued about and redistributed; until then it cannot. Fourth, one proposal was to require AI companies to report their full economic, social, cultural and environmental footprints. Without companies reporting these data from within, the world has had to rely on independent researchers and journalists investigating these numbers from the outside.

06WHERE TO GO FROM HERE

These convenings examined AI primarily through the lens of human development, but the discussions repeatedly intersected with the two lenses that have dominated the global AI debate: how the technology is governed and how it is put to use. Participants saw different roles within the United Nations system. Member States can establish norms, principles and collective commitments. The Secretariat, agencies, funds and programmes can generate evidence, support national capacity, convene actors who might otherwise rarely meet, and help translate global principles into practice. The question was therefore less whether the United Nations could govern AI on its own than where its particular legitimacy, reach and convening power could make a difference to the people and communities living through these changes.

The discussions also acknowledged the limits of what the United Nations can do to steward AI’s future development. Enforcement, for example, is not readily available, while technical iteration outruns regulatory process. Geopolitical tensions may also make binding consensus difficult. Against that backdrop, an argument first made in Cairo and repeated in Mexico City, Xi’an and Rome was that the United Nations possesses another form of influence: moral authority and legitimacy. A statement of values and red lines endorsed by a plurality of States, while not binding, would still confer standing on the companies and governments that adopt it while withdrawing standing from those that do not. Several voices across the convenings described the present moment as a vacuum of moral authority. In that view, a meaningful role for the United Nations would be to help articulate the human values and boundaries that should guide technological development, even where enforcement is unavailable.

Yet the design trade-off was left unresolved. A statement strong enough to prohibit practices such as mass surveillance or social scoring might lose the support of States whose participation would be important to its legitimacy. A statement broad enough for universal agreement could become too weak to influence behaviour. The Rome convening gave the accountability question a concrete case. Participants debated a company’s use of copyrighted books to train its AI models and the court-approved class-action settlement of US$1.5 billion that followed. One participant argued that this case teaches technology companies that they can break rules so long as it makes them enough money to pay penalities if caught. Another described the case as an ongoing failure of giving credit for creative and intellectual labour. A third disagreed, arguing that copyright laws should not restrict the reading of written materials by AI systems as that gives publishers a veto over how knowledge is used.

Throughout the series of convenings, a range of ideas emerged for carrying the human development conversation forward. One was to make AI the subject of an explicit global goal rather than a crosscutting concern, on the argument that nobody is measuring what AI is doing to people and that nothing unmeasured gets managed; an eighteenth Sustainable Development Goal was proposed from the floor. A second was to write specific human commitments into whatever follows the 2030 development agenda, such as human agency, human dignity, human relationships, human discernment, and human transcendence, each stated as something development is meant to protect. A third idea looked outside the United Nations, proposing that Pope Leo XIV convene a network of the interested — faith communities, grassroots movements, scientists, citizens and other stakeholders— gathered to test and scale what already works well. A fourth idea returned to power itself, asking whether the distribution of power over AI systems should be examined as a subject or high-level commission in its own right. A fifth was to continue this series of intergenerational convenings in all countries where the United Nations has a presence so that many more communities can shape the conversation, and to bring what emerges and to bring what comes back into the Sustainable Development Goals Summit in 2027.

07MONTHS, NOT YEARS

At least three important questions were raised but left unanswered by participants. First, participants asked repeatedly for specificity about which jobs will be created, rather than assurances that some will be. The related proposal — that those who capture the gains should fund the safety net during the transition — attracted broad agreement but no mechanism. Second, everyone who raised transparency also noted that existing commitments remain unenforced, and nobody offered a route from a non-binding statement to a consequence. Third, participants raised questions about the geopolitical competition between the two States at the frontier of AI development – China and the United States of America – but judged it outside the scope of these convenings. That left one of the largest constraints largely unexamined.

The clearest shared judgement from the four convenings may have been about time. The prevailing view from participants was that humanity had still has agency over what comes next, but most believed the period in which societies can shape what AI means for their lives may be measured in months rather than years. That is far shorter than the timeframes within which institutions, communities and social norms usually adapt.

In every room, when the question was put, the response by participants suggested that the window to positively steer AI’s development is still open but rapidly closing. Shows of hands were recorded in words, then later in counts through the follow-up survey. In Mexico City, a large majority judged that humans still have the upper hand, while only a handful thought their own community’s social fabric was strong enough to absorb the change under way. In Xi’an almost every hand went up for giving humanity six months to steer AI, far fewer for a year, and even less for two years. In Rome, the same question drew the fewest hands of any convening for the proposition that humans still hold the upper hand. The participant survey run across all four convenings puts numbers to this for the first time: only 15 per cent of participants think people in their communities can understand how AI systems affect them and only 28 per cent believe people can contest decisions that AI systems make about them. Only 30 per cent of participants believe that the social fabric of the communities in which they live and work are strong enough to absorb the changes AI will bring over the next five years. 50 per cent judge the window to be less than 12 months and another 31 per cent believe we only have an extra year beyond that. While 22 per cent of participants feel AI’s development has largely outrun human direction, 100 per cent of the survey participants reported that humans still have the ability to determine AI’s future (see Annex 2).

Not everyone accepted the question. One contributor explained that asking whether humanity or AI has the upper hand keeps a false opposition alive. In her view, AI is neither a moral nor responsible agent standing across from humanity, but rather a set of technologies that particular people build, own and deploy. The more useful question, in this view, is how “we” steer the humans who shape AI – and how “we” hold “them” to account.

08WHAT COMES NEXT

AI learns from the record humanity has left behind: what we have written down, preserved and digitised. It is an immense inheritance, containing our knowledge and creativity alongside the prejudices, exclusions and cruelties of the past. The conversations made clear that expanding that record is only part of the answer. What matters just as much is the human judgement brought to what we inherit and what we choose to pass on.

So, in Cairo, Mexico City, Xi’an and Rome, people from faith communities, science and academia, government, the arts and culture, young leaders, Indigenous and Nomad communities, returned time and again to the parts of human life that acquire their meaning through experience.

They spoke of presence, vulnerability, grief and love. Of the confusion in which understanding is built. Of a child growing attached to a machine whose affection is simulated. Of a former war correspondent whose fear for the future was a world in which there were no eyes to meet.

This report is a record of those conversations, and necessarily a partial one. The convenings brought together a wide range of perspectives. They also revealed how much more of the human story remains to be heard, and how deeply the meaning of AI will depend on the lives, cultures and communities through which it is experienced.

That incompleteness matters all the more because the window for shaping what comes next may be brief. Across the four convenings, people expressed a sharp sense of urgency. Most judged that humanity still has the ability to shape the direction of AI. Many were far less confident in the strength of their own communities’ social fabric to absorb the changes already under way.

Widening the conversation therefore becomes part of the work itself. Between Xi’an and Rome, the Deputy Secretary-General brought them into an exchange with students at the Indian Institute of Technology Delhi, asking future engineers and innovators what human problems they wanted AI to help solve, and what “democratizing AI” should mean in practice. In a country investing heavily in widening access to AI skills and capabilities, the conversation offered a glimpse of what comes next: a new generation beginning to make these questions its own.

In 2027, these questions will go closer to where their answers reside. The organizers of the convenings, UN Office for Partnerships, will develop a discussion guide that can be shared through country-led and community-based convenings supported by United Nations Resident Coordinators and country teams.

In Cairo, Graça Machel challenged the room to ask whose experiences and perspectives are included when decisions about AI are made. The next phase takes up that challenge. It will invite more communities to bring their own experience, values and understanding of a good human life into this conversation, adding voices and knowledge that this first record could only begin to capture.

What emerges will help inform the 2027 SDGs Summit and the longer conversation about what human development should seek to protect and advance beyond 2030.

The first four convenings repeatedly returned to one question: who is the “we” that decides what AI is for?

The next phase will test what happens when that “we” becomes wider.