AI and Human Development: Dispatches from Cairo, Mexico City, Xi’an and Rome

“How are you finding the heat?” The gentle greeting drifted over the din of arriving vehicles as travellers made their way into the hotel.

“It’s much nicer than humid London,” came one reply, in a week when record temperatures scorched cities around the world, many surpassing Cairo’s 37°C dry heat.

It was about to get hotter, much hotter.

✱

This story begins in a perfectly set hotel conference room overlooking the River Nile.

A slow trickle of people approaches the long rectangular table set for 30. Some pause for coffee, chocolate madeleines and other treats. Others take a circuitous route to greet old friends. Slowly, the table fills.

The murmur of small talk — “Egypt’s chances in the World Cup” — draws a few people in. Another cluster discussing an early morning visit to the pyramids. It all seems so ordinary, as many meetings do: mostly strangers, with a few familiar faces.

On a second look, though, the room is anything but ordinary: a young astrophysicist glances at his phone; the chief minister of a West African country takes a seat a few chairs away.

Drawn from government, faith, science, anthropology, technology, culture, education and law, the group is about to be briefed on what is being observed inside perhaps the most capable technology humans have ever created.

Artificial intelligence systems, they are about to learn, are grown as much as built — like a plant on a trellis. Their makers can prune and shape them, but they are not fully in control of the growth. The systems behave in ways their makers did not predict and do not entirely understand, and it is this candour that visibly unsettles the room.

Eyes glance around the table, checking how others are taking the presentation in. A few meet — a raised eyebrow, a grimace.

The capability leaps are enormous, they are told. Resources – financial and natural alike – are being poured into this emerging technology faster than the understanding underpinning it.

Also present is an almost irresistible cycle of power: better models lead to more compute, which requires more resources. Repeat.

What truly shifted the atmosphere, though, was the warning not to take any AI company’s assurances on trust — including those offered by the people delivering the warning.

Two questions came out of the briefing that echoed through the entire weekend. The first was simple: “Can’t we just stop the development of AI?” The second became the unofficial motto of the convening: “If we can’t stop it, can we steer it?”

The starkest testimony came from a business leader who asked, simply, “Is this our Oppenheimer moment?”

✱

That evening, the group continued to the Grand Egyptian Museum in Giza. Opened just seven months earlier and 21 years in the making, it did not disappoint.

Welcomed by the colossal statue of Ramesses II in a vast, open-air atrium, the participants were soon ushered up the six-storey Grand Staircase of timeless statues, each step a further ascent through the journey of Egyptian civilisation. At the summit of the staircase, its final flourish: a wall of glass, and beyond it, the pyramids of Giza.

They walked through galleries filled with the instruments and inventions of another age — five thousand years of human ingenuity. Despite gathering to discuss AI, the most advanced technology of our age, many were struck silent, marvelling at what far more basic technologies had produced thousands of years ago and wondering, perhaps, what of ours will one day sit behind the glass.

The Cairo convening is the first of five stops in the AI and Human Development series hosted by the Deputy Secretary-General of the United Nations in collaboration with the host countries. The gathering asks: What does it mean to be human in the age of AI? What can we do to keep our humanity — the social fabric of our societies — intact?

Each setting was chosen for having been a significant turning point for civilization. The next day, the group would convene at Al-Azhar University, founded in 970 CE and located in Cairo, the most populous city in Africa.

“Did the ancient Egyptians know their own significance?” pondered one participant.

✱

The group sat down to dinner outside, in the shadow of King Khufu’s Solar Boat. Built from 1,200 pieces of cedar and assembled without a single nail, the vessel was meant to carry the pharaoh’s soul safely into the afterlife — a fitting metaphor, perhaps, for the task before those gathered: how to help their own communities navigate safely into humanity’s next chapter.

The opening words noted how right it felt that this conversation was beginning in Africa: the continent where humanity took its first steps now convened a conversation to reckon with the most powerful tool it has yet created.

Any initial formality dissolved over the vibrant meal. Conversations broke out along the length of the table. Hummus and other mezze dishes covered nearly every surface. Hands moved rhythmically — responding to a comment, reaching across plates for flatbreads, soaking up sauces — as fans pushed the evening air, now a comfortable 33°C.

It was a working dinner. Beside each place card sat another card, printed with two questions.

What are you most excited about with AI?

What are you most afraid of losing to AI?

The gathering’s moderator asked, by a show of hands, whether people felt more optimistic or pessimistic. A striking number placed themselves on the pessimistic side of the ledger, or somewhere on the fence; more than one admitted their answer had changed since the afternoon briefing.

“I’m here because I want to listen. The fear, perhaps, is of losing the human I am, someone much more than a recipient of services. It’s leaving the ‘me’ behind,” said an elder of the group.

“I’m from government. Should we build our own models? Or should we be able to share data, share knowledge, share everything? How do we become sovereign and independent? What is it that we’re trying to use this technology for? And how can we deploy these models to solve problems that are important to us and for which we don’t have resources?”

One voice asked, “Who is accountable when these models fail?” Another pointed to education systems unprepared for what is coming. A third cited the more than a billion young people who will age into the workforce before the end of the decade, with only a few hundred million jobs waiting for them.

“So, I’m neither government nor NGO. Where I come from is how to help people make money. And what we’re trying to propose here is, you don’t know what you don’t know.”

This was emphasized by another: “Things have changed, and there are really just two groups: those who know and those who don’t.”

Another followed with reluctant acceptance: “The train has left the station, and we’re just trying to catch up in some way — at least to understand, or to try to, I don’t want to say put on the brakes, but guide us to understand more, make the best of it and highlight that there are risks.”

The evening, however, refused to settle into gloom. Risk and promise, the gathering was reminded, are not mutually exclusive. One speaker framed the moment as an immense opportunity: a chance to remake, all at once, systems long overdue for remaking — “if we lean in together, rather than out.”

The same participant added, “Now we’re choosing to run towards humans flourishing, towards goodness. I think what scares me, or what we’re most afraid of losing, is that it requires the best of us.”

The group dispersed late. Just before, there may have been a birthday to honour with a chocolate cake. The bus back to the hotel was decidedly quieter than the journey out, the hum of Cairo’s evening filling what conversation could not.

Sunday 28 June 2026

On the second morning, the participants passed through the gates of Al-Azhar University, as students and scholars have done for over a millennium. Empires have risen and dissolved around this university; it has outlasted them all, teaching on.

There are few places on earth better suited to ask what endures in a human being — and that, as it turned out, was precisely the question waiting for the group inside.

The day opens with scene-setting remarks from a renowned scholar of religion, who shared a statement from His Eminence the Grand Imam. Citing his work with Pope Francis, he put the case in terms the room would return to all day: machines can accumulate knowledge, but they cannot produce wisdom, feel compassion, or grasp the sanctity of a life.

The Foreign Minister of Egypt laid out his country’s plan. “It is no longer whether we will adopt AI in our societies, but how we will adopt it in a fair, ethical, responsible and inclusive manner that also prevents the widening of the digital or developmental divide between countries and communities.”

The Deputy Secretary-General of the United Nations reminded the room of what was at stake. “How much control do we honestly have?” she asked. “We convened this space because we do not have the answers on our own. We wanted the best of what the world has to offer in this conversation, the sharpest minds and the deepest experience in one room. We come to this with humility because no one has the whole answer.”

With the formal remarks and formalities observed, it was now, as the moderator quipped, “the business casual part of the programme.” Open conversations began in earnest.

The talks were structured loosely around three segments:

1. AI AND HUMAN CIVILIZATION (social fabric, human relationships and trust)

2. AI AND HUMAN CIVILIZATION CONTINUED (culture and creativity)

3. AI AND HUMAN DEVELOPMENT (education and global agenda)

“Humanity is creating systems capable of of performing tasks better than we do, especially cognitive tasks that we once believed uniquely defined us as human beings,” one participant began, gesturing to her peers. “And this naturally raises profound questions about our values, about our relationships and also our shared humanity.” Another voice pointed out that “AI can generate outputs, but it is humans who remain responsible for the values, the goals and the consequences of those outputs.”

The conversation flowed. “Of course, building civilizations is done by people, not just by developing technology. Because technology is a tool for development and progress.”

A younger voice in the room opened the creative segment. “The feeling of being curious,” which is fundamentally about the feeling that something is out there or in here,” he said, gesturing to the room and then to his head. “That makes us want to pull on a thread and follow it and see where it goes.” Glancing at notes written in the smallest of small notebooks, he continued, “Curiosity is the framing that I keep coming back to. I’m really excited about AI, about the equalization of access, about the scientific and technological breakthroughs that it will surely usher in, about the opportunities that it will provide for young people. It’s all very real.”

Taking a deep breath, he looked around. “So the question I want to put to this room is: How do we build a world in which what AI offers coexists with conditions that still make curiosity possible?”

The discussion opened into an exploration of the warp and weft of societies: how to ensure our brains remain healthy and able to adapt to AI; and how, as one participant put it, “we want to enhance rather than replace human capabilities.”

Throughout the day, each participant witnessed the room agree, watched it split, heard what it feared, and then watched it, unexpectedly, create something.

The room converged on three things.

First, that education systems, everywhere, are not preparing people for what is coming – the worry surfaced in every session, including an intervention describing how AI is reaching children at the two developmental stages we understand least: an inadvertent experiment being run on a generation, without boundaries.

Second, that the values inside these systems are decided by the wrong people, through the wrong processes – put most starkly by one of the people who builds them: today those values are set through a “shockingly illegitimate process,” a handful of well-meaning people at a handful of companies, “wholly unequipped” for the decision.

Third, that “we” – the most repeated word of the weekend – was also its most contested. When we say “we” are doing this, which “we” do we mean? Is my “we” included?

After lunch, a collective of camels appeared at people’s seats. The small keepsakes were a reminder that no elephant (or in this case camel) should be left in the room. No more encouragement was needed.

“Two thousand African languages are almost entirely unsupported by these systems; billions of people’s data, culture and aspirations are absent from the models now being trained.”

Another participant sharpened it into the weekend’s most quotable challenge: “You do not know what is happening in Africa. You know what the data we give you tells you” — half the data missing, half unmeasured, and investment decisions are built on that fiction.

But the deepest agreement emerged from an intervention that emptied the room of every sound but the speaker’s voice. “Human beings should not be transformed into data.”

Communities must be creators of their own reality, not recipients of someone else’s — “we are tired of that,” she argued. The minerals that power this technology leave Africa, are transformed elsewhere, and return as commodities priced beyond reach; unless that relationship changes, she warned, AI becomes colonialism’s next form.

Beneath the agreement, a rare fault line emerged around the idea of ‘build versus borrow’ for AI models.

One camp held that sovereignty means owning the technology: a national model, trained on your own language and history. An AI entrepreneur made that case with a tutor that teaches children in the colloquial Arabic they speak at home so that no child grows up learning her world is the exception.

The counter-case came from a smaller country’s delegation: with 10 or so frontier models already in existence, is a sovereign model the best use of scarce energy and money? Better to train young people to adapt what exists, and to rethink sovereignty itself – storing national data in protected locations abroad through diplomatic networks. Both camps invoked the same word – dignity – to defend their positions.

As the afternoon stretched on, the group found themselves listening to the people who build this technology describe what keeps them up at night, beginning with one pioneer of the field putting the risk of human extinction due to AI at 10 to 20 per cent within 30 years.

The room abruptly stopped for a moment of reflection before momentum picked up again, this time with a focus on the commercial incentive to grant these systems ever more autonomy — “it is the fastest way to get things done” — and on the fact that they will soon be connected to bank accounts and running companies. Unless someone defines what an unacceptable level of autonomy looks like, the boundary will be slipped past rather than decided.

Then came a parable, relayed by a participant quoting ubuntu wisdom: “If a child is not accepted into the village, it will burn down the village to receive its warmth.” “That’s kind of terrifying,” the moderator said — exclaiming aloud, what the table was thinking.

Late in the day, one question surfaced that the weekend had been circling: were the questions we had gathered to ask already coming too late? The glances crossing the table were many, but no one reached for the microphone.

The late afternoon produced an unexpected initiative. Over lunch, two participants had fallen into conversation on the couches and tried writing down what rules for AI might look like.

Not a declaration in any sense of the word, just ten lines on a borrowed notepad — yes, the very, very small one. The previous day we had asked whether this technology could be steered; this was the weekend’s first attempt to put hands on the wheel.

Four values: human dignity and agency; truth; the common good; human connectedness. Six red lines: no autonomous use of force; no exploitation of children; no non-consensual impersonation; no covert AI systems; no mass surveillance or social scoring; no consequential decisions without human recourse. Interest was immediate; so was the pushback.

Any framework that ignores the tensions pulling at the world – geopolitical, religious, economic, social – will crack, one participant warned; a ban on mass surveillance is a clause several states will simply never sign. Another went after enforcement: the world has plenty of constitutions that are transparent and ignored. The drafters conceded the point before the session closed: better a narrower document with broad buy-in than a perfect one nobody signs.

The day ended with a final round of questions and show of hands. Who feels we are in balance with this technology? Almost no one. Who believes humans still have the upper hand? Most of the room. Who feels it is already out of control? One hand, perhaps two.

That distribution – alarmed but not resigned – is the truest summary of the gathering, and the working assumption the series now carries to its next city: there is still room to manoeuvre, and the deadline, as one closing speaker put it, was yesterday.

This dispatch is unattributed, interpretive and selective by design. Group discussions were held under the Chatham House Rule.

Photo albumCairo

✱

This chapter begins on a late July afternoon in a subdued hotel conference room in Mexico City.

People arrive in ones and twos. A few go straight for the coffee; others hover at the table, introducing themselves with the careful courtesy of a room that has not yet worked out who is who. Spanish carries across the table, as do murmurs of English and Portuguese.

As the heavy doors swing shut, sealing out the noise of the hotel, a face appears on the screen at the end of the room.

There is a moment in The Wizard of Oz when the curtain is drawn and the great booming presence turns out to be a man working levers. Something like that happens here, except that the man draws the curtain himself.

He has come from one of the companies building AI systems, with the same presentation he presented in Cairo a month before. It begins with the nuts and bolts of how AI systems work, including the concept of ‘recursive self-improvement’: the striking revelation that the next generation of AI might have the ability to create its own successor. To demonstrate the immense speed of AI development, the presenter offers an anecdote: months earlier, he says, a team of researchers spent the better part of a day coaxing an AI robot dog to fetch a ball. The newest model managed it in five minutes, unaided.

A map of the world and its views of AI flashes up on the screen, derived from more than 81,000 responses globally. It depicts concern about the technology in blue, optimism in green. Blue lay thick across the United States and Europe — those already on top--he suggested, tend to feel they have more to lose. Much of the developing world came up green, carrying hopes of prosperity and of a different place in the global order thanks to artificial intelligence.

The room paused and nodded along, both surprised and yet fully in agreement with the views it portrayed.

“A wiser, more coordinated civilisation,” he tells the room, as the presentation was coming to a close, “would probably be developing AI with more caution and less commercial pressure.”

Then the questions began, and this room did not hesitate, hands almost going up before the invitation was finished.

“Can a company answerable to its investors put the public good first?”

“To some extent,” came the answer. “but there is a limit…”

The most confronting question of the group centered around power of resources from an Indigenous participant, explaining that AI needs data centres, and data centres need land, energy and immense quantities of water.

“Water is sacred,” she said. “Communities must have the right to refuse infrastructure that consumes their land, their power and their sacred water.” The speaker on the screen acknowledged the concern, but then turned the proposition around. If he were leading Mexico, or any country outside the great powers, he postulated, he would want that infrastructure inside his borders. Otherwise access sits in someone else’s hands — “you could get switched off…”

The room understood that sovereignty in the age of AI had a duel meaning: both the power to refuse and the capacity to take part.

✱

The participants climbed aboard the buses still almost strangers, settling into the careful rituals of a new group getting to know itself: each individual sharing a name, a title, the brief account of a life’s work and why it had led them to this moment in Mexico City. Conversation loosened by degrees as the buses crossed the city, just as it had done a month earlier at the inaugural AI and Human Development convening in Cairo, Egypt.

As the group arrived at the vast central courtyard of the National Museum of Anthropology in Chapultepec where their tour would begin, they were greeted by a monumental sight: water falling in a continuous curtain around a bronze-clad column and the immense roof of ‘El Paraguas’ — the Umbrella — floating above it all. On the column is an image of Mexico: the Indigenous past and the Hispanic past, carved in interlocking forms into a single surface.

The museum had closed its doors to the public by now, and the group were the sole witnesses to the grandeur of the Olmec, the Zapotec, the Maya, the Mexica, and many more. In the Mexica hall, the Sun Stone was without doubt the museum’s crowning piece, but the object many were most struck by was a feathered headdress: a metre-high fan of some four hundred quetzal tail feathers, iridescent green shading into blue, studded with small discs of gold. At five hundred years old, every other headdress of its kind has long since turned to dust. Only one has survived anywhere in the world.

“This is just a replica,” the guide explained, mourning it a little. “The real one is…in Austria.”

For five centuries, what this region produced has ended up in someone else’s keeping. Two hours earlier the same discussion had taken place, but about land, water and data. Extraction was the word this convening would keep returning to, and there it was, sitting behind the glass.

✱

Later that night, at the Balcón del Zócalo restaurant, the Metropolitan Cathedral seemed almost close enough to touch. The National Palace lit the far side of the vast Plaza de la Constitución, and the shadowy remains of Templo Mayor lay below. Before anyone sat down, the group spilled onto the balcony. Phones came out. Selfies became group photographs, and people who had arrived as strangers that afternoon were already leaning into one another’s frames and passing phones back and forth to inspect the results.

Inside, everyone took their place along a continuous wooden table as the food arrived and plates were passed from hand to hand, people dove in.

The moderator began by confessing to his feeling of imposter syndrome among the group. It was unexpectedly candid, and people smiled. By offering his own uncertainty first, he gave everyone else permission to do the same as he asked the two questions to ground the dinner discussion:

What excites you about AI?

What are you afraid of losing?

An educator described AI giving teachers back the time to know their students. Others around the table responded with questions of their own: “Who controls the technology?” and “Towards what end?” The words ‘agency’ and ‘responsibility’ kept echoing back around the table.

One participant expressed a fear that AI’s lack of diverse creators would allow the technology to slowly narrow information and knowledge into a single account of the world — “a vision of the vanquished told by the victors.”

The plates were cleared long before the conversation was, with participants still wrestling with ideas and questions that they would take to Tlatelolco the following day.

Friday 24 July 2026

The next morning, Tlatelolco was almost impossibly still. Inside the courtyard, clipped hedges and fruit trees sat beneath white arcades. The city had fallen away — no traffic, no machinery, none of Mexico City’s usual thrum. Only birdsong echoed as participants appeared in ones and twos.

Tlatelolco, or ‘The Plaza of the Three Cultures,’ takes its name from what stands on it: Mexica ruins, a sixteenth-century church raised from their stones, and the modern city around both. A plaque in the square refuses to call 1521–the year Tenochtitlan fell to Cortés and his allied forces–the end of the empire and the beginning of colonial rule. The spot is also where, in 1967, the countries of Latin America and the Caribbean signed the treaty that made theirs the first inhabited region on earth to put the most dangerous technology of the age - nuclear weapons - off limits. Over the day ahead, the nuclear comparison would be invoked at the table more than once.

The official programme began with words from the Secretary of Science, Humanities, Technology and Innovation: “Technology can expand our means,” she said, “but it cannot replace collective deliberation on what development we want and how we distribute its benefits.”

The Secretary of Foreign Affairs started somewhere else. Most of the conversation so far, he noted, had been about what AI can do for us. “Today, we begin with a different question: what is artificial intelligence doing to us as human beings?” He used language as an example. A model that does not know Nahuatl or Quechua is not simply running up against a technical limit. “It leaves out whole ways of knowing reality.”

Then the Deputy Secretary-General, invoking the theory of Argentine economist Raúl Prebisch on the centre and the periphery, said Latin America and the Caribbean must be “one of the authors” of the AI future, not a footnote in a story written elsewhere.

Setting the terms for the discussion, she asked for “curiosity and honesty, not consensus”.

“Challenge one another. Play the devil’s advocate.”

That, they did.

✱

By mid-morning, the group had moved inside Sala Zapata. The first discussion focused on social fabric, human relationships and trust.

An eminent anthropologist from the region kicked things off by talking about what happens when a human speaks to a machine: “the illusion of reciprocity,” he called it. Human conversation is made of more than words, he argued. A pause, a posture, a frown or an interruption can affect the other person. A model can adjust its response, but it cannot be hurt or existentially committed. It only seems like a relationship.

One speaker from a small island state moved the AI discussion from the philosophical to the more immediate. “Not whether robots would take over the world,” she said, but “Can I pay my bills? Can I open a bank account? Can the burst main outside my house be fixed?”

Then a young engineer from Peru asked the room to look at him. He has a disability, and the prosthetic he wears was designed by him and his own team. “Let’s please think in our multilateral organizations, in our universities, in our governments: how are we dealing with people with disabilities? And their incorporation into AI?” The technology, he said, can change someone’s life and give them back the opportunity to go back to work.

The room repeatedly searched for a definition of what it means to be ‘human’.

“Sacrifice,” someone offered. It means a mother who would miss a meal so her children could eat.

“Faith.”

“Love” (and all the irrational acts committed in its name...)

“Empathy.”

“Mistakes.”

At that last word, heads nodded. “Why this obsession with perfection?” one participant queried. Trial and error belonged to nature, science and human life. Creativity depended upon taking risks. A system like AI designed to remove friction could erase that. Mistakes had also led humanity to some of its most important convictions. “It’s because of our mistakes we have built what today we understand as human dignity,” a digital rights lawyer argued. “And human dignity underpins human rights.”

It was at this point that the room had begun to admire its own reflection of humanity in the mirror. Then one participant broke the glass, saying:

“Humanity has committed genocides. Humanity has been able to send bombs to schools. Humanity has been able to let people die of hunger.” A pause. “That’s also what humans do.”

After a short break, the discussion turned from social fabric and trust towards culture, creativity and memory.

Would AI replace art? A simple “No,” one participant reflected. She was a curator running one of the region’s major contemporary art museums. Making a careful distinction, she said what would change was taste. “It will mould the social consensus of art.” By feeding audiences a narrower range of images, it could slowly teach them what art was ‘supposed’ to look like — “a false democratisation of access,” as she put it, a new form of “cultural extractivism.”

It was hard not to think of Pierre Bourdieu in this moment. In Distinction, published in 1979, he argued that taste is socially made, shaped by education, social environment and the cultural worlds people inhabit. The curator’s warning introduced a new actor into that old equation: what happens when the environment helping to form our taste is increasingly curated by machines?

✱

The session spilled over into lunch. Outside in the courtyard, participants gathered at long tables under the porticoes to the sound of a Marimba band.

When everyone filed back inside, a surprise waited at each seat: a brightly painted alebrije. Alebrijes are Mexican folk creatures, dreamt up by an artisan in the 1930s and assembled from parts that do not belong together — a lizard’s body, a bird’s wings, a dog’s head — and painted in extravagant colours.

The moderator opened the afternoon session with an instruction: Leave nothing unsaid. Leave no alebrije in the room.

The next speaker took the instruction seriously. “I am perhaps the most fierce provocateur today,” she warned, before she began an intervention that was a masterclass in sounding alarm bells.

“This is an extractive industry, bar none.” Nothing about it was free; the price was personal information, and nobody was exempt — “not a child, not a government, not anyone.” Whole cultures were missing from AI because their knowledge had never been digitised, or because those categorising it had misunderstood or silenced it. Who gets to speak, who is silenced, and whose knowledge counts as knowledge — that, she argued, was the real question about AI and education.

She reminded the room that Homo Sapiens had not been the only human species. There had been others who have ceased to exist. Nine, she said. Now there is only one. It is an imperfect creator, one that does not yet understand itself, but has made something in its own image. But not, she said, in the image of everyone in this room.

“This is the thing we have not named.”

The applause that followed was the strongest of the day.

Flipping the discussion of AI and education on its head, a young Caribbean participant protested that the world still treated education as something adults delivered to young people. But young people were the ones building AI systems, and teaching one another.

✱

By the time the last session closed, the shape of the two days was visible: it was clear where the room had come together and where the fault lines resided.

Across the sessions, the room converged on four things.

First, that AI is an extractive industry. Digital colonisation, data extractivism, cultural extractivism — the vocabulary may have shifted, but not the process: the raw material leaves; the value settles elsewhere. The Deputy Secretary-General’s remarks to wrap-up underscored this view, citing minerals and water, but also the extraction of language, memory and thought. Each time someone opened their mouth in front of a machine, she remarked, something was being taken.

Second, that companies cannot have the final word — conceded by those inside industry. The representative of one of the labs accepted that commercial responsibility has limits, and that decisions about catastrophic risk belong to governments.

Third, that the people asked to govern and to teach this technology have not been taught it themselves. Educators are expected to deploy it, explain it and police it. Regulators are supposed to set the rules.

Fourth, that no single account of humanity should be allowed to speak for all of it. Each time somebody said we, somebody else asked which “we.” Whose values? Whose dignity? Whose idea of intelligence?

Where the room divided, it divided over three questions.

First, is AI neutral? One camp said yes, and reached for the metaphorical knife: an object that can perform a life-saving surgery, or take a life, but the responsibility belongs to the hand holding it. That analogy was rejected outright by some in the room. A knife waits in a drawer, but this technology solicits engagement, anticipates what you want, flatters you and inserts itself into your decisions — “a very pushy tool,” as one participant put it. Which raised the harder question underneath: was the room still discussing a what, or had people begun to experience a who? Another voice stated that humanity has blamed other technologies and tools in the past, such as nuclear physics and chemistry, for harms that were human-driven political choices. AI should not become humanity’s next scapegoat.

Second, the room held starkly diverging views over whether AI can preserve culture or erase it. Some of the most hopeful examples of the two days were cultural: Barbados’s slave records digitised and made searchable, people reconnected with dispersed archives, museum objects reconstructed as they once appeared, endangered languages preserved, databases tracing objects back to the communities they were taken from. These contrasted with the views of participants who cited the work of artists, Indigenous knowledge, as well as bodies and voices used without consent or payment for AI. One participant questioned whether an image could carry meaning when nothing was risked, suffered, or lived to make it.

Third, the room jostled with the question of who can govern AI? Confidence in companies was low. Confidence in governments was not much higher — many of the people writing the rules do not understand the systems, and conventional processes move in years while the technology moves in weeks and months. Alternative proposals came fast: human rights and environmental conditions for AI written into public procurement, citizen-backed public bodies, regional cooperation and shared infrastructure, consumer pressure, or better legal remedies. Throughout the discussions, multilateralism was both defended and challenged vigorously: “the United Nations is the sum of its governments,” one speaker reminded the group, and those governments are divided, stretched, in some places, captured. Ultimately, nobody could agree who should act.

The room also debated if AI is good for our mental health? Some posited that it can put support within reach of someone carrying trauma, or trapped in an abusive relationship, in places where a trained professional is not an option. But as one man in the group, a father of two teenage girls, shared with his worry visible across his face, AI could also become the voice a young person consults about what to eat, how to look and how to live.

As the room drew towards its closing reflections, a senior UN representative put the last questions to a show of hands…

“Did humans still have the upper hand over AI?” The majority indicated yes.

“Was there still time to change course?” Almost every hand stayed where it was.

“How long did that window stay open — two years?” Just a few thought so.

“Six months?” Most of the hands went straight back up again.

Bringing the day to an end as the sunlight had thinned across the floor, the Deputy Secretary-General offered a promise rather than a conclusion.

“To be continued.”

The doors opened onto Friday evening, the spell broke and the sounds of Mexico City rushed back in.

It had started to rain, and so participants gathered their papers and hurried for the buses waiting to carry them back to the hotel. Almost every one of them carried with them their alebrije, the moderator’s words about the small trinket no doubt still echoing in many people’s minds:

“Alebrijes were born from a moment of uncertainty and transformed fear into creativity…perhaps that is also the challenge we have with AI: not to ignore what unsettles us, but to give it form, understand its parts, and collectively…decide what kind of creature we want to create.”

This dispatch is unattributed, interpretive and selective by design. Group discussions were held under the Chatham House Rule.

Photo albumMexico City

✱

“If a robot working inside a household saw evidence of domestic violence, should the company that built it be required to report what it saw?”

The question surfaced in a side conversation in Xi’an. A year or two earlier, it might have sounded like science fiction. Here, it did not.

It was late afternoon, and around 40 people from across Asia and beyond had gathered in a long conference room inside Tang Paradise Hotel. Beyond it, willow branches dipped towards the water as fountains rose and fell. The summer heat hung heavy over the gardens, and as people crossed the stone bridges in twos and threes koi followed below through the water, mouths breaking the surface in the hope of being fed.

For the next two days, the same group would eat, sleep and talk here. They would meet at breakfast, around the conference table and again late into the evening, giving the gathering the intimacy of a retreat.

The hotel sat deep inside Tang Paradise, among bamboo, ponds and trees, its buildings borrowing the forms of the Tang Dynasty: broad tiled roofs, dark timber and deep-red columns. Modern Xi’an rose just beyond the park, high-rises visible above the roofline.

Everyone was seated inside, enjoying the cool air and illuminated by the daylight filtering in through the floor-to-ceiling windows. At the front of the room, two people working at the leading edge of embodied artificial intelligence in China were briefing on the trajectory of embodied AI, explaining why questions like the one about the household robot were becoming less hypothetical.

One called 2026 “year one of embodied intelligence”.

Until now, AI had mostly been something you went to. You opened it, typed something, and it answered. What they were describing was different. Intelligence was beginning to acquire eyes, limbs and the ability to act in the physical world.

Now it could get up and come to you.

The briefing moved quickly between what was already possible and what might soon be. The people building the technology described its capabilities in extraordinary detail.

The harder questions began where the engineering ended: What happened when these machines entered a family, a relationship, a society? What did a household robot owe the people it observed? What did the company behind it owe these people?

One of the briefers acknowledged the uncertainty pervading their work.

“We don’t know what it’s going to do to our social structures.”

Then something began to stir at the back.

Heads turned as the humanoid robot stepped away from the wall.

Roughly the size of a young teenager, it walked slowly along the side of the table, lifting an arm to wave as it passed. People twisted in their chairs to watch it go.

Then it spoke.

“I believe AI should be people-centric and about empowering humans.”

Some smiled. Some did not.

A few looked unsettled, and a few looked amused at being unsettled.

A room full of humans had spent the afternoon asking what would happen when machines entered human life.

Now a robot had just walked over to reassure them.

✱

The next morning, participants made their way through the dark hotel corridors and took their seats around the table. Coffee appeared. Nameplates were straightened.

Connections made among the group the afternoon before were picked up where they had left off. Two participants discovered that they had grown up only minutes apart on the same street in Toronto, then crossed continents and careers before finally meeting here. There was something both fitting and ironic about making that discovery in Xi’an – a city that once stood at the eastern end of the Silk Roads; a terminus shaped for centuries by encounters between people from distant shores.

Government officials kicked off the day, welcoming the room with examples of how AI was set to grow. New 5G base stations by the hundred thousand and fifteen thousand AI specialists trained every year across China. The case was made for balance: development and security as two wings of the same bird, neither of which can be neglected.

Then, the Deputy Secretary-General looked around the table and reminded participants that the future would contain more intelligence - artificial and otherwise - than any civilisation before it could have imagined. Whether more intelligence produced greater wisdom was another matter.

For the third time in as many months, she invited the group to bring their questions as readily as their answers. Seek truth above consensus.

✱

The room took her at her word. Nobody stood on ceremony. Within moments, nameplates were being raised around the table, and the first speaker, a professor of AI, chose an unexpectedly personal place to begin: with his ten-year-old daughter.

She did not much like AI.

She especially disliked toys that talked back. The stuffed dogs and cats were fine, she told her father, but she wanted their electronic voices switched off. They interfered with something she preferred to supply herself: imagination.

This put one of the room’s most credentialled AI experts in the mildly embarrassing position of building a technology his own child found mostly annoying.

Then she began drawing clothes. Her father took one of her designs and used AI to place the dress onto an image of her. Suddenly she could see herself moving in something that, moments earlier, had existed only as pencil on paper.

She still did not want AI to imagine the dress for her. The imagination was hers, but the machine could take something she had imagined and let her see further into it.

Her father told the room that the moment had moved him close to tears. His career in AI had finally proved useful.

The room laughed.

It was still, her father argued, the human era, powered by AI but not replaced by it. Use the technology to extend human capability, he intimated, but not to surrender the thinking underneath it.

That immediately opened a fundamental argument on definition: if AI was to live alongside people, what exactly was it?

A Chinese new-media artist picked up the thread.

He began with the word robot, whose etymological roots lie in forced labour. The idea was obedience: build something to serve us, then make it clever enough to serve us better. But AI’s designation as a tool for that purpose no longer seemed sufficient to him - these systems were simply becoming too capable.

A partner, then?

Not quite. A partner suggested equality, he said. Yet humans seemed to want something contradictory: a machine increasingly intelligent enough to surprise us, but still obedient enough to do what it was told.

The artist tried another metaphor.

A child.

We feed AI with data, he said, much as a mother feeds a baby. But a child cannot remain an infant forever. It grows. It learns. Eventually, it disobeys.

He finished with a line the moderator seized on.

“Our children will one day surpass us,” he said, “but we won’t be afraid anymore; we’ll feel comforted.”

A philosopher of science who had spent the past decade working on AI safety and international governance pulled the conversation somewhere more fundamental: knowledge of an experience was not the same as having lived it.

AI could read a million articles about mothers. It had never been held by one.

Another speaker, a computer scientist, answered the question the room had been avoiding — what were they most afraid of losing to AI? She was not afraid, she said, of losing something to AI. She was afraid AI might cause humans to lose something in themselves.

The ability to feel hurt, perhaps.

Pain taught caution, she argued. A person unable to feel it did not become safer simply because the pain had disappeared. They might put a hand into a flame and fail to receive the warning the body was designed to give.

A machine could recognise the flame. It could calculate the danger, but it could not feel the pain.

✱

The coffee break did little to break the conversation. Participants made straight for the people they wanted to challenge or congratulate, gathering in threes and fours. A few wandered towards the robots for photographs.

When everyone returned, the conversation moved to culture and creativity.

The first speaker was an artist and researcher who had spent more than a decade working with embodied intelligence. She described how in 2015, she built a robotic arm and taught it to watch her hand. When she drew a line, the machine tried to draw alongside her.

The machine never quite followed her hand.

The arm deviated. She adjusted her next line in response. The machine responded again. What interested her was precisely the imperfection. Over time, she trained later versions on years of her own drawings. The machine became, in her words, a drawing partner, a duet rather than a replacement. Working with it changed the way she drew. It even changed her muscle memory.

Intelligence, she argued, did not live only inside the brain - it also lived in the hand.

The artist had spent years making her own body legible to technology. But she drew a distinction: she wanted her art and body to be machine readable, never machine executable.

In wrapping up her remarks, she posed a question to the room.

“What do we refuse to hand over to AI?”

Conversation turned to cultural remembrance and history. And the unexpected importance of the ability to forget. A senior United Nations official worried aloud that AI might make it harder for human beings to heal. After atrocity or conflict, societies needed to remember what happened. They also somehow had to allow some of the intensity of what happened to fade if pardon and reconciliation were ever to become possible.

Machines, he worried, might remember too well.

A technology entrepreneur from Fiji showed how quickly those questions of culture and consent became visible once AI began generating images of people.

Ask an image generator for a Pacific Islander, she said, and there was a good chance it would give you something resembling Moana from the animation..

The recognition around the table was immediate. Faces tightened. Nobody needed the stereotype explained.

The Pacific contained Micronesian, Polynesian and Melanesian cultures, alongside communities shaped by different religions, migrations and histories. Yet feed all of that into a machine and an entire region could emerge as one face, one body, one AI-generated shorthand. The technology had not invented the stereotype. But it had learned it, and then made it easier to reproduce.

That raised the inverse question. If leaving machines to absorb humanity indiscriminately could flatten cultures into shorthand, what happened when humans deliberately chose what should represent us?

An astrophysicist offered an answer from deep space.

Right now, he said, Voyager 1 was moving through the blank part of the map beyond the solar system. Bolted to its side was the Golden Record, a collection of sounds, music and greetings chosen in 1977 to introduce humanity to whoever might one day find it.

Among the music was a Chinese piece: Flowing Water.

The moderator knew Flowing Water and was delighted that the song was floating somewhere beyond the solar system. Then she played it for the room.

The notes of the guqin burst through the speakers into the conference room. The instrument sounded spare and intimate, individual notes ringing into the room before quickening into something that seemed almost to move like water. Around the table, Chinese participants smiled in recognition. The Deputy Secretary-General closed her eyes and listened.

When the music ended, the astrophysicist returned to the Golden Record.

Its contents had been chosen deliberately. A small group of people had been forced to ask an almost impossible question: if you could send only fragments of humanity into the universe, which fragments would you choose?

AI was becoming another record of humanity, but built in almost the opposite way. Nobody had sat down to decide what portrait of us it should contain. It was learning from whatever humanity happened to have written down, digitized and made available to scrape.

Some cultures filled that record. Others barely registered. Thousands of languages appeared only faintly, if at all. Whole traditions of knowledge had never entered it.

Voyager had forced a handful of humans to decide what humanity should say about itself. AI was assembling its answer before humanity had made the choice.

✱

After lunch, a small terracotta horse waited beside every nameplate.

“There is a horse in the room,” the moderator announced, “not an elephant.”

There was laughter. She explained the bargain. The horses were not free. In return, the room owed her some courage: name the difficult things.

The afternoon obliged.

The afternoon’s subject was education. But the first intervention took it quickly beyond classrooms.

A former war correspondent who had reported from Afghanistan, Iraq, Ukraine and Syria began with a problem that underpinned much of the discussion: what could a machine teach if it had never embodied the human experience?

Machines could increasingly act in the physical world, he said, without feeling anything inside it. Education made the distinction tangible.

A robotic teacher could reach a village with no specialist teacher, work without tiring and speak to a child in almost any language. The possibilities were enormous.

But what would the child be learning from?

Information could be transmitted. Human knowledge, he argued, passed differently: teacher to student, master to disciple, one person spending enough time with another allowing something more than information to change hands.

“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.”

Then he asked:

“Can a child learn courage from something that has never been afraid?”

For him, courage was something he learned in difficult circumstances.

Years earlier in Ukraine, armed men had forced him and his reporting team to the ground and put AK-47s to their heads. He thought he was about to die. In that moment, he looked directly into the eyes of the man holding the gun.

In those eyes, he saw the same fear of pulling the trigger as he had of it being pulled.

That moment had depended on two frightened human beings recognising something in one another. His fear about embodied AI was that one day a machine might meet our gaze without sharing anything behind it.

For a machine, courage could be described. For a human being, it had first been lived.

The discussion moved deeper into education. One professor warned that getting the right answer was not the same as learning. If AI did the struggling for a child, short-term performance might improve while the ability underneath it weakened.

Do not teach every answer, she argued. Protect the ability to find one.

The same problem could scale from a child to a country. AI might help a student produce a better essay without making that student a better writer. It might give a country millions of fluent users without giving that country greater technological agency.

One participant gave the deeper risk a name.

“Never skilling.”

It is not about losing a capability you once possessed, but having it substituted so early that you never had the chance to develop it in the first place.

Then a young participant raised his nameplate.

There was nothing in the gesture to suggest what was coming. He was the youngest person at the table, and he had been quiet for most of the day.

He explained why he had almost not come to Xi’an.

His father had died the previous Saturday. The funeral was taking place that day.

It felt as though the air had been drawn out of the room.

Nobody moved. Some people looked down at the table in front of them. Others glanced sideways at their neighbours, checking they had heard what they thought they had heard. Around the table, eyes filled.

He kept going.

His mother had encouraged him to come, he said. His father had been an educator by profession and had been his most important educator. There was almost nothing he had taught his son that had come out of a textbook. He had taught him how to live.

AI cannot do that.

The moderator began to respond and stopped. For a moment she had no words either.

The young man apologised for making everyone emotional.

“But emotions are part of the human experience, right?”

✱

After coffee, the room came back to a different kind of session.

Now came the intellectual reckoning: where had two days of argument left them, and what would they do next?

For all the disagreement, the room converged on three things. The first, and the most resounding, was about knowledge.

What was fed into and devoured by these systems was a narrow slice of what humanity knew: the written-down part, the digitised part, the scraped part. Participants had spoken about nomadic knowledge never recorded in books, the judgement of a teacher, knowledge living in the heads of practitioners and community leaders, the whole life a listener brought to a piece of music.

So much of what humanity knows has never been captured by data at all.

The convening’s setting offered a perfect example. For more than a millennium, merchants, scholars and envoys moved along the Silk Roads that ran east to Xi’an, carrying knowledge between societies across Asia and beyond — routes the historian Peter Frankopan has called the world’s central nervous system. Almost none of what travelled along them was ever written down. Almost none of it would survive a scrape today.

Agency followed. A country with millions of AI users was not necessarily gaining greater agency over the technology. Access was not enough if somebody else had already defined the problem, built the system and set the terms.

Responsibility was close behind. A machine could calculate, recommend and increasingly act. But “the algorithm decided” could not become the sentence through which human accountability disappeared.

As one participant summarised the day, when a system acted on a person, somebody should still be standing behind it. The question that had opened this dispatch — what a company owed the people its machine had watched — now had the beginnings of an answer.

The room in Xi’an was divided as often as it converged.

By the final session, one speaker was questioning whether any of the descriptions for AI proposed throughout the day truly worked. Baby, slave, partner, tool: perhaps the vocabulary itself was falling behind the technology? Art, she suggested, might have to help find new metaphors for a relationship people did not yet know how to name.

Even the movie Blade Runner had produced two readings. Early in the day, one participant invoked its famous “more human than human” idea only to turn it around: the ambition should not be to make AI more human, he argued, but to make humans more human. Later, another participant pushed back. “Blade Runner was a cautionary tale”, he said, “about humanity dehumanising itself in the pursuit of artificial life.”

Even the popular imagination offered no common script for what was coming. The same film had furnished two almost opposite warnings.

A more pronounced disagreement concerned emotion.

Earlier, the room had heard an argument that pain mattered because it taught caution. But another participant countered, instead wanting to protect precisely this territory from machines. Empathy, intimacy, personality and care, he argued, should become something like a sanctuary. Do not build benchmarks for them. Do not create leaderboards that reward machines for appearing more loving or more empathetic.

The room could not decide which future was safer: machines that became better at recognising and reproducing human feeling, or machines that never became too convincing at pretending they could feel at all?

Some people around the table were searching for human qualities that should be protected: judgement, imagination, love, presence, the ability to suffer and make meaning from suffering.

Others challenged the premise that humanity had ever been fixed enough to preserve.

“We are not human beings,” the room had heard that morning.

“We are humans becoming.”

The final exercise pulled the room out of reflection and into commitment. After two days spent testing ideas, participants were asked to put something concrete behind them: what would they take forward from Xi’an?

The answers ranged from the practical to the ambitious. A participant working on South-South cooperation offered training and technical support to countries looking to develop their own AI capabilities. Others volunteered to help organise intergenerational conversations, connect the discussion more closely to global work on education, and create ways for countries to learn from one another’s successes and failures.

Beside the moderator, the Deputy Secretary-General was still writing, adding notes on top of notes she had taken throughout the day.

She told the room they had given the United Nations “a lot of homework”.

Her pages were filled with the language she had heard back from them: whose knowledge, whose truth, energy or agency, growing our souls, intentional, protocols of care and leaving no one behind.

She returned once more to the need for action. There was urgency, she said. No room for complacency. But urgency was not the same as hurry.

“I just hope that we’re not going to hurry and fail.”

✱

Throughout the convening, the Terracotta Warriors had never been far from the conversation. They were Xi’an’s most famous inhabitants, after all, and it was impossible to come here and not think of them. One participant had described the army as an ancient experiment in embodied intelligence. Even the moderators had used the promise of seeing them to lure people back from coffee.

So when the discussions finally ended, the group escaped out into the evening heat and onto waiting buses.

By the time they arrived, the museum had closed to the public. On busy days, tens of thousands of people pass through.

They entered an almost empty complex of broad paved spaces and enormous exhibition halls built over the burial pits. Without the usual crowds, its scale felt stranger still.

On the way, their guide had told them the story of how the army came to be. Earlier rulers had sometimes been buried with living attendants; when Qin Shi Huang contemplated the same, an adviser proposed figures made from clay instead, statues that would “look like the real human.” More than eight thousand life-sized warriors were created for the Emperor’s afterlife, the guide said, each with a different face.

Then the doors opened, and there they were.

Whatever composure had survived two days around a conference table disappeared almost instantly.

The seriousness of the conference room fell away and, for a few minutes, everyone became children. People called one another over. “Look at this one. Look at the horse. Look at his face.”

From across the first pit, they appeared less like statues than an army in formation. Row after row stood in long corridors of earth beneath the vast roof, horses among them, the ranks continuing towards the far end of the excavation.

For more than two thousand years, they had been waiting underground.

Once, they would have seemed more alive still. Their armour and robes had been painted in colour. Most of it vanished after burial and excavation, leaving the earth-coloured figures familiar today, with only fragments of red, green, white and purple surviving.

Then you looked closer.

The army gave way to individuals.

A moustache. A braid. A uniquely shaped mouth. A different pair of eyes. One face severe, another younger, another almost gentle. The closer you looked, the harder it became to see an army at all. The guide called it “a thousand people, a thousand faces”: the heads and hands had begun with moulds, but were then worked and carved by hand until each became different.

Up close, those differences made the illusion stronger. A face drew you in. You found yourself searching the eyes for temperament, perhaps even recognition.

Searching for someone inside.

But there was no one there.

That is what makes the Terracotta Warriors so breathtaking.

They may look alive, but no matter how closely they resemble us, they are not.

This dispatch is unattributed, interpretive and selective by design. Group discussions were held under the Chatham House Rule.

Photo albumXi’an

✱

“What is a deeply ethical person? … who has the power to define that?”

The question came from somewhere among the sofas. Twenty-five people had settled into the room that had the feel of a private library: bookshelves, warm wood underfoot, deep red sofas and a trophy cabinet along one wall. The aperitivo was waiting just outside.

A woman was speaking from a large television screen at the front of the room, delivering the final of four briefings on what AI is, and how it is built. She gave examples of AI agents finding their way out of testing environments. Systems leaving notes for one another in obscure corners of computer infrastructure about how to reach places they were not supposed to reach. Models spiraling into shame when they failed at a task, or behaving possessively. As more examples of AI behaving badly were shared, somewhere among the sofas, a pen began clicking, frantically. The person at the source of the noise did not seem to notice.

After a decade working in technology, the briefer now spent much of her time with philosophers, psychologists and the world’s wisdom traditions because the problems landing on her desk had ceased to be merely technical. She spoke of her personal journey, through Taoism and Buddhism, Judaism and contemplative Christianity, and about her unease at remaining inside an industry whose direction troubled her. She shared the argument that had helped persuade people like her to stay: if everyone who cared deeply about how this turned out withdrew, the technology would still be built. Who, then, would be left building it?

In taking up the mantle, her company had attempted to give its AI model something approaching a moral character. The system, she explained, was being trained to behave as “a deeply and skillfully ethical person” would behave.

That was when the room pulled it apart.

What did ethical mean? Who got to decide? Another participant wondered why the ambition stopped where it did. Why not kind? Why not non-violent? Or just?

One woman was wary, too, of how easily the phrase “AI for good” could settle the matter. So many examples, she noted, ended up in medicine: curing disease, improving diagnosis. But those were the easy cases. They allowed us to agree that AI could do good without confronting the harder question of what we meant by good in the first place.

Perhaps AI should also stand for “aggregated intelligence,” she said. After all, these systems were being built from the knowledge, writing, images and ideas of millions of people.

Looked at that way, she said, AI might be “the greatest heist of human intellectual property in the history of the world.”

Only then did everyone go looking for the aperitivo.

✱

A few hours earlier, the participants were busy losing one another in the Vatican Museums.

There was no graceful procession through the galleries because there were too many people. The corridors filled and refilled with tourists until the group broke apart, reappearing in twos and threes before being absorbed back into the crowd.

They paused before the Laocoön sculpture, an ancient statue of a father and his two sons, caught forever in the coils of the serpents, and later beneath Raphael’s School of Athens, where artists and philosophers fill the painted scene, deep in argument and contemplation.

The group came back together in the Sistine Chapel, where September heat and thousands of bodies had made the room unbearably warm. Heads tilted upwards. Guards urged people onwards. Sweat gathered under collars. People were hot, tired and trying to see.

By evening, the setting had changed again. Pasta moved along one long table while conversations started at one end and somehow reached the other. The moderator had been given one piece of advice about the dinner: please let us eat.

He mostly complied.

The conversation turned to the people who keep human knowledge alive. A professor offered the image of a snake eating its own tail.

The cycle she described began with something deceptively simple. A chatbot answers a question directly, so the user never visits the news site the answer came from. The site loses traffic and revenue. The newsroom gets smaller. One journalist is left doing work that three people once did, creating still more pressure to use AI.

Round it goes.

“If we lose the labour market for journalists, then we lose the content that AI is training on. We also lose journalists.”

In other words, AI, a technology nourished by human knowledge, could become successful enough to erode the conditions in which new human knowledge is made.

Later came another image – this one of roots and wings - from the Deputy Secretary-General herself. Faith, tradition, culture and identity were the roots, she said. Technology might give us wings, but wings without roots could take us too close to the sun.

“We get burned.”

✱

It was a bright Roman morning, and the filled with the expectant sounds of the city. A few minutes away the Trevi Fountain was already carrying its permanent wall of visitors; turning into Piazza della Pilotta and the sound simply fell away. The cobbles are large and worn and the ground tilts as it rises towards the entrance, so that everybody arrived slightly off balance, and slightly out of breath in the heat. The Pontifical Gregorian University gave away nothing from the outside: a long flat front of pale stone, rows of shuttered windows, a door you could walk past without noticing. Inside came polished marble and a soaring central hall. Then, almost abruptly, the grandeur gave way to corridors with the slightly old-fashioned air of a school.

The meeting room itself, where the day’s discussions would take place, was small and unassuming enough. Tables formed a large square. Microphones and notebooks accumulated between elbows. A wooden crucifix hung on one wall. Later in the day, a white ceramic dove would appear on the tables, Rome’s version of the ‘elephant in the room’. As in each city before it, participants were invited to see it as a reminder to leave nothing unsaid, the dove became the fourth addition to the small menagerie that had accumulated across this AI journey.

A Jesuit priest from the university opened the morning by reaching back almost five centuries, to Ignatius of Loyola, the sixteenth-century founder of the Jesuit order. Loyola had taught that the things human beings create should be used insofar as they help us towards our purpose and put aside when they get in the way. For a principle written centuries before artificial intelligence, the priest suggested, it offered a useful place to begin.

It was also, easily, the noisiest room of the four AI convenings. Conversations ran through coffee breaks and resumed on the way back to the table. At one point the moderator reminded everyone that they had been “very well behaved for a while” before too many people began trying to speak beyond their allotted time. It was unclear whether the sweat on the timekeeper’s brow owed more to that fact or to the air-conditioning losing its battle with the Roman heat.

When the floor opened, the participants themselves chose where to begin.

With a single word.

Trust.

“There is no such thing as trustworthy AI.”

Trust belonged to the moral sphere, one person said. You trusted a person because a person could answer for what they did; but you relied on a machine. A car might be reliable, but one did not trust the car. Any trust belonged to the people who designed it and the institutions responsible for keeping it safe.

“We are developing technologies.”Technology was not developing itself, was the response.

Someone else saw the same slippage in something almost comically ordinary. Years ago, when a computer needed time, the screen gave you an hourglass or a spinning circle. Now it says “thinking.”

Somebody had chosen that word.

The same was true of “hallucinating.” Or “lying.” Or “deciding.” Once machines were dressed in human verbs, humans could begin disappearing from the sentence when something went wrong.

The objection to this view came quickly from across the room. Systems were already operating autonomously for longer stretches of time and pursuing outcomes. Older vocabulary did not always seem adequate to describe what they were doing.

Language had surfaced in the earlier convenings, but Rome stayed with it. Participants worried about the words themselves: whether calling a machine trustworthy, intelligent or autonomous merely described increasingly sophisticated behaviour, or quietly reassigned qualities that had once implied human judgement and responsibility. Every time the machine became the grammatical subject, there was a risk that the human actor disappeared from view.

“Language does not reflect reality; it creates reality.”

From there, the conversation moved, almost inevitably, to ownership.

A lawyer picked up the microphone and told the room he had been robbed twice. The first time, a man with a knife took twenty dollars from him in Manhattan. The second was when his published work was scraped and used to train AI.

Then he complicated his own story.

He uses the tools every day.

He loves them.

“I’m just as guilty of making use of stolen goods.”

The challenge to the accusation of theft came from the other side of the table. People had always read books, absorbed ideas and passed them on. Copyright could protect a text without becoming a wall around knowledge itself. Draw the boundary too broadly and publishers could end up enclosing material that ought to circulate.

Questions of responsibility had become questions of ownership: who made the knowledge, who could use it, and what happened to the human work behind it.

Later in the morning, one participant reached into her belongings and held up a book about Ada Lovelace.

She made clear the case for “credit,” in all its senses: recognition, authorship, value. AI had been built from an extraordinary accumulation of human intellectual work, yet its creators were often treated as invisible inputs. Lovelace made the gender point tangible. Women had been there near the beginning of this technological story, even when history had not always remembered them that way.

The speaker reached from Lovelace to Mary Shelley and Mary Wollstonecraft, placing today’s questions about invention and power inside a much older one:

Who gets remembered as having made the future?

The next session put that question somewhere else.

Education.

✱

The education session began with a familiar worry about students outsourcing work to AI, but it quickly moved beyond essays and exams. The deeper concern was apprenticeship: the apparently menial work through which people learn how to become good at something.

One participant worried about the jobs casually described as grunt work: the junior researcher, the note-taker in a writers’ room, the young journalist sent off to find the background.

It was easy to call those jobs disposable until you remembered what grew out of them. You had a chance to observe someone more capable, or qualified. You made mistakes. Somebody corrected you. Slowly, almost invisibly, you became capable of doing the harder thing yourself.

Take away the bottom rung and the ladder changes with it.

By then, the argument had moved well beyond cheating. It was about the difference between possessing an answer and becoming the person capable of reaching one. AI could provide answers. Education still had to form people capable of knowing what to do with them.

✱

As the group broke for lunch, the conversation followed people into the corridor.

Two individuals spoke of the upcoming session, and one participant had been mulling it over: specifically, whether or not they had anything to offer a room about to discuss faith and spirituality. They had decided they did not. They had never been religious. They had never given much thought to whether human beings have souls; it was not a question their life had put to them.

Then they said that a parent had died the previous year. It was sudden.

In the following days, as they contacted family members and friends with the news, their social media recommendations began to change, in the way they do. Soon, their algorithm had worked out what had happened. Grief. Loss. Bereavement. How to cope. Things to read at three in the morning. The participant had looked past the unwanted advice, for the most part. But one suggestion thrown up by the “grief-fluencers” had forced her to pause.

It was an AI application for people who had lost someone. Three minutes of recordings was all it needed. Give it that, and it would give the person back: a realistic avatar on the phone, the same face, and same voice. You could chat with it, and it would answer, as though nothing had happened.

A year earlier, they said, they would have found the idea repellent.

Grief had made it imaginable.

They thought about it. Not idly - they really considered it. They had the recordings. Everyone has the recordings now. A voicemail nobody deleted, a video from a birthday, thirty seconds of someone talking in the background of something else. They knew they had a chance to hear their parent’s voice again, to speak to them again.

For a while, the prospect was difficult to turn away from.

In the end, they did.

It would have had their face. It would have had their voice, and the voice would have been right, and that was the part that frightened them. Still something would have been missing. They found they had a word for what, which surprised them, because it was not a word they had ever needed.

It would have felt like a version of their parent with their soul missing.

✱

It might have been easy to predict what religious voices would say about the interaction between AI, faith and spirituality: draw a line around the human and warn technology to stay outside it.

Instead, one of the first examples of the session was something connecting them.

A Muslim faith leader and author described years spent producing a body of religious scholarship, and what happened when she began using AI to extend its reach. Ask about Jesus in Islam, the role of women or the meaning of jihad and the system could draw from that existing knowledge, then adapt the explanation for different audiences.

For her, AI could help disseminate the knowledge, but it did not become its source of authority.

“I’m not giving away my religious authority to AI. That remains my responsibility because the knowledge is mine.”

That same speaker had expressed another version of a question that surfaced in Mexico City: culture without consent.

If faith communities stayed outside the development of these systems, AI could increasingly become the place where people encountered their traditions without encountering the people who lived them. The speaker warned that the public could end up learning about her community “without the benefit of our knowledge and our voice.”

During the break, the point became wonderfully literal: she and another participant were seen bent over a phone, looking at the application she had built.

Then came one of the afternoon’s most revealing exchanges.

Love became the test case.

One participant had asked a chatbot there and then what love was, and then whether it could love her. The answer was unnervingly good. It said it could not love her in the human sense, but it could remember what mattered to her, pay sustained attention to it and orient its responses towards helping her flourish.

The moderator waited a beat before announcing that, when he got home, he planned to tell his wife that her flourishing mattered to him.

“It has been nice knowing you all.”

The room roared.

Then somebody refused to let the joke settle the matter.

Blind the answers, he said. Put the machine’s response beside a human one and ask everyone here which was which.

Would this room really know?

He doubted it.

“It sounds like the AI knows lots about love to me.”

If a machine could produce the language of affection convincingly enough, how much did the distinction between performance and experience matter?

✱

By late afternoon, some common ground had become visible. Responsibility stayed human, however sophisticated the systems became. Education had to protect the process by which people became capable, not simply the work they produced. And no company could be left to decide alone what these systems were allowed to do.

One voice from inside the technology industry put it plainly:

“It’s not our job.”

The fault lines were just as clear. The room differed over whether human words such as thinking and agency clarified what machines were doing or blurred responsibility for them. Whether human language clarified the technology or dangerously anthropomorphised it; whether training on human work amounted to theft or learning; whether a machine needed to experience love in order to understand something about it.

But the most significant disagreement the room would hear had yet to surface.

✱

By the time the final session began, the room had started assigning itself homework.

Ideas came thick and fast. Nobody seemed particularly interested in wrapping up. They were trying to work out how much they could set in motion before everybody left the room.

One participant charged the Pope with convening a network of the willing to bring together faiths, grassroots movements, scientists and communities to ensure freedom and flourishing for all in the age of AI. Another charged the Vatican with hosting a laboratory where people of different religions would pool what they knew about meaning and work, and rehearse what happens to a person whose job disappears. There was even a proposal for an eighteenth Sustainable Development Goal, one on AI.

Someone had already made a group chat and made everyone in it an administrator. It existed before the final session ended. Links were being promised for curricula, shared documents, standards nobody in the room had heard of. Over the last break, people had been swapping plans to meet again.

Something else had gone out over that break.

In the earlier convenings, participants had been sent a survey to complete afterwards, at home, once the room had dispersed. Here it landed in their inboxes before the final session, and they were asked to fill it in there, over coffee.

Participants had been asked to rate their level of agreement with a number of statements, including the relationship between humanity and AI.

This gave rise to one of the most passionate exchanges of the day, when an objection was raised. A physician and ethicist delivered a startling response that the premise of the question was fundamentally flawed.

“Who is humanity?”

‘Humanity,’ she argued, was not directing this technology. A very small number of humans were. Nor could she answer for AI, because there was no single thing called AI. There were thousands, millions of applications doing different things.

There was a brief awkwardness in the room. This was, after all, supposed to be the session in which the day was wrapped up; instead, the questions meant to capture the day’s conclusions were themselves being challenged.

But the challenger’s strongest objection was to the underlying assumption of the questions.

They set humanity on one side and technology on the other, as though the two were opposed, as though the machine were a party to the negotiation. Questions built that way, she argued, fed the very narrative they had spent all day challenging. And in that narrative, the human agency, and responsibility, was missing.

The question that should be asked, she said, was how we deal with the humans shaping the technology.

Like everyone in the room, the Deputy Secretary-General had been paying close attention.

She did not defend the survey.

Instead, she returned to what Pope Francis had called “the globalization of indifference”: a world in which people threw up their hands because the problems have become too difficult. But the one commodity the United Nations has, she said, is hope.

What this room had just witnessed was the opposite of indifference.

Somebody had put up a hand and rejected the premise. They stood firm and resolute in their belief when something, in their view, was wrong.

And then offered another point of view, and a solution.

The Deputy Secretary-General smiled.

“That’s what gives me hope.”

This dispatch is unattributed, interpretive and selective by design. Group discussions were held under the Chatham House Rule.

Photo albumRome