Like many kids of the late 60s and early 70s, I grew up watching The Twilight Zone. One of my favorite episodes was “Nick of Time,” about a young married couple (Don and Pat) who are stranded in a small Ohio town after their car breaks down. They end up in a diner, where their table has a novelty fortune-telling machine called the Mystic Seer which, for a penny, will answer a yes-or-no question.

So the couple begins asking questions. Will Don get the promotion he has been waiting to hear about? The answer suggests yes, and soon afterwards the phone rings and he finds out that he got it. Will their car be fixed by three o’clock? Again, the answer seems to fit. Then Don asks whether they should leave town. The Mystic Seer warns them about danger. They leave the diner anyway, are nearly hit by a car, and head back to ask more questions. By this point, they are hooked.

The parallel to what is happening today is uncanny. Generative AI arrived in most of our lives only a few years ago, and we started asking it to rewrite emails, summarize documents and explain Excel formulas. Increasingly, people ask it whom to date, whether to leave a relationship, what career to pursue, what someone meant by a text message, why they are unhappy, whether their reaction to something is reasonable, and what they should do next.

The useful machine

The difficulty with dismissing AI is that it is useful. Studies show that AI reduces cognitive work1. Writing emails, synthesizing large amounts of information, and taking notes have all become easier. That convenience, however, can change how much thinking we do ourselves.

In 2025, researchers at Microsoft and Carnegie Mellon surveyed 319 knowledge workers about 936 real examples of using generative AI at work and found that greater confidence in GenAI was associated with less critical thinking, while greater confidence in one’s own ability was associated with more2. They also found that AI changed the character of the thinking people performed: information gathering became information verification, problem solving became the integration of AI-generated responses, and carrying out tasks became supervising them.

Further compounding this problem is automation bias. Most people view artificial intelligence as more intelligent and accurate than they themselves are3. ChatGPT can tell me which career it thinks I should pursue, explain its reasoning based on what I have told it about myself, respond to my objections, and produce a revised answer. If I remain uncertain, I can supply more information and ask again. The process can feel less like using a tool and more like consulting an adviser.

This is essentially what Don does with the Mystic Seer. His first questions concern things he cannot know, such as whether he will receive his promotion or when the car will be repaired. Once the machine appears to get those things right, the questions change. He asks whether he and Pat should leave the diner, whether they should leave town, and whether something bad will happen if they do. Somewhere along the way, he has moved from using the machine to obtain information to consulting it before making decisions.

Nick of Time

At the end of the episode, Pat finally persuades Don to leave the diner. They collect their repaired car and drive out of town. Another couple enters and sits down in front of the Mystic Seer. They have clearly been there before and have already organized their lives around its answers. The man asks whether they can leave town that day. The machine tells them they cannot, so they stay.

In this 1960 episode, the Mystic Seer had a small collection of pre-written answers and knew nothing about the people sitting in front of it. It could not remember their previous conversations, analyze their relationships, read their emails, absorb years of personal history or generate a persuasive explanation for its advice. Yet the couple believed its answers without questioning them.

The issue with generative AI goes far beyond accepting its answers without critical judgment, although that is troubling enough. When we repeatedly rely on AI to remember, synthesize, interpret, solve problems and formulate arguments for us, we spend less time exercising those capacities ourselves. Our brains adapt to what we repeatedly ask them to do. Cognitive skills, like other skills, are strengthened through use. This matters particularly for children and young adults, who are still developing the habits of attention, reasoning and judgment they will carry into adulthood.

And this brings us back to the couple sitting in the diner. Their problem is larger than believing a fortune-telling machine. They have lost the habit of questioning it. They no longer wrestle with uncertainty, test the answer against their own judgment or decide what they think.

The Mystic Seer cost a penny a question. 60+ years later, we have built a machine that can answer almost any question, knows enough about us to make the answer personal, and is always within reach. The couple in the diner surrendered their judgment (and subsequent actions) to a plastic devil with a spring-loaded card dispenser. How much of our own judgment, our decisions, and our capacity to think for ourselves are we willing to hand over to a machine?

^{Source: NYT}

In 2004, I was living in Namibia and was cut off from daily contact with my professional circle in Washington and New York. Watching my Rolodex collect dust, I was happy when LinkedIn appeared because it felt like a way to follow people’s careers and remain part of a professional world thousands of miles away. While filling out my profile, I did feel a bit uneasy about divulging the dates of every degree and every job I’d held. But I filled it in anyway because being findable by professional contacts was the whole point.

The following year I was invited to join Facebook, back when membership was still restricted to particular universities and I was excited to “reconnect” with my old classmates. As more people joined, Facebook became a way to bridge the miles, to share my life and see pieces of theirs, posting quick reflections as mundane as what we’d had for dinner alongside bigger things like the birth of a baby. And the data and photographs accumulated year after year.

My sister, however, made a different choice. Though my sister did join FaceBook, she joined under a pseudonym with a bunny as her profile picture and she never posted anything personal or shared a photo. I remember feeling annoyed at the lack of reciprocity and thought that was stalker behavior. From her, I wanted the same easy stream of pictures everyone else gave me and desired to see what she was doing without arranging a time to talk. When I approached her about it, her response was: “You know we can always call each other”.

We could, of course, though a phone call asked us to find a time and stay present for the whole conversation. Facebook let me keep up with dozens of people while asking almost nothing of any of us. I could see a friend’s new baby, notice a former colleague had changed jobs, or catch a cousin’s holiday between meetings, and it felt remarkably like staying in touch. My sister’s point, all along, was that intimacy without connection is superficial.

More than twenty years later, my sister has almost no visible footprint online, while I seem to be everywhere, even though I deleted all my Meta accounts years ago. There’s another difference between us as well: I shared photographs of my children.

At the time, I believed I was sharing them with people who cared about us, family and friends scattered across countries and continents. Facebook gave them a window into my kids’ lives, and it gave me the pleasure of having those lives witnessed. My kids were too young to consent, and I made the decision for them anyway, because it felt ordinary. Millions of parents were doing the same thing.

This summer, I finally got around to reading Kashmir Hill’s Your Face Belongs to Us, which tells the story of Clearview AI, a facial-recognition company founded by Hoan Ton-That that scraped billions of photographs from Facebook, LinkedIn, Instagram, news sites, company pages, and countless other corners of the open internet, converting the faces in them into biometric identifiers, a system that could match an unknown person to their own online images.

Someone using Clearview could upload a photo from a security camera or a phone, and the system would search its database and return possible matches, with links to the pages where those images appeared. A face could lead to a name, a workplace, a social account, and from there the rest of a person’s life might be surprisingly easy to reconstruct. The company offered the system to police departments and government agencies, often handing out early access informally. Wealthy investors and private also had access with some reportedly using it to identify strangers in restaurants and at parties.

The photographs in Clearview’s database had originally been posted for entirely different reasons: professional headshots, family pictures, school photos, or simply someone caught in the background of another person’s image. Clearview treated the fact that a photo was available online as permission to collect it and repurpose the face inside it.

When I joined Facebook, sharing a photo meant sharing it with the people I’d chosen as friends. It akin to what a photo in your wallet used to be - something you handed to the people in your life rather than something you put on a billboard. That was how most of us understood it which is exactly why almost nobody stopped to ask who else might be looking, or storing, or selling what we handed over.

What I hadn’t imagined, even accounting for all of that, was that a company I’d never heard of could collect those photographs years later, measure the faces in them, and fold the resulting biometric data into a commercial identification system. Almost nobody using these sites in 2005 could have anticipated that. The technology, the business model, and the eventual customers were all still invisible to us. Consent given for one purpose, to a narrow audience, under one set of rules, quietly turned into permission for every use the technology would later make possible, under a completely different set of rules that nobody voted on.

Turning photographs into biometric data changes the nature of the intrusion. A compromised password can be changed and a stolen credit card cancelled but your face travels with you into streets, airports, hospitals and political demonstrations… connecting your physical presence to whatever records exist about you. A flawed match can implicate an innocent person (and we already know these technologies carry racial bias).

There’s a potentially larger problem: a biometric database outlasts the governments and corporate owners that built it. The photograph I post today of my child at a birthday party sits in a database that could still exist twenty years from now, under an administration none of us can predict, matched against a face at a protest that hasn’t happened yet. My child never chose to be in that database. I put them there, for entirely different reasons, under a government I trust(ed), and I have no way to guarantee who inherits that record or what they decide to do with it. That is the real cost of treating a face as permanent, searchable data. It becomes available to whoever collects it now, and to whoever holds power later.

I feel responsible for the choices I made, though personal regret only goes so far. I made them inside a culture built to make sharing feel natural: platforms that rewarded disclosure with connection, while the consequences showed up later, elsewhere, through companies I’d never had any relationship with.

Even my sister’s caution protected her only so far. She may not post photos of herself, but other people can still take them. A school, an employer, or a newspaper can publish an image she never approved. She can go through the world with no online profile and still not control every camera pointed at her, or every person who uploads a picture with her face in it.

So the answer can’t just be “share less.” More care makes sense, especially with photos of children, but care has limited power over images already circulating, and it can’t reach into a database to pull biometric data back out of a photo once it’s there.

The law has started to catch up, unevenly but genuinely. European regulators have found Clearview’s practices incompatible with data-protection rules. In 2024, the Dutch Data Protection Authority fined the company €30.5 million for building an illegal database of faces belonging to people in the Netherlands, and regulators elsewhere have issued their own fines and deletion orders. These cases establish a real principle: information remains personal even after someone finds it online. Enforcement stays difficult, though, when a company operates from another country and nobody outside it can inspect the database.

The law should draw a clear boundary around the human face. Public access to a photograph should never amount to consent for biometric extraction. People deserve the right to discover who holds their facial data and have it deleted. Children should be able to erase digital histories created on their behalf, and police searches should require judicial authority and leave an auditable record.

When I filled out my LinkedIn profile in Namibia, I wanted former colleagues to find me. When I joined Facebook, I wanted to reconnect with old classmates. When I posted photographs of my children, I wanted distant friends and relatives to watch them grow. Each of these choices had their own purpose and audience.

Twenty years later, companies claim rights that were never part of those choices. They have taken our desire to be seen by people we know and converted it into the ability to identify us anywhere. Being findable was always the point but being permanently searchable was an agreement none of us actually signed.

^{AOC holds jar of brown water from next to Meta’s Georgia data center at EPA hearing. Source: Moneywise.com}

This is 2026, the Chinese Year of the Fire Horse, and true to the name, it has been the hottest summer anyone alive can remember. Across Europe, India, and the American Southwest, record temperatures have killed thousands of elderly people this year (including my children’s grandfather) and sparked wildfires from the Iberian Peninsula to the Cascades to the forests of Uttarakhand. The Rhine (the industrial artery of Central Europe) has fallen to some of its lowest levels on record and is too shallow for the cargo ships that carry fuel, chemicals, and grain across the continent to sail fully loaded. Fields are cracking and crops are failing on a scale farmers say they haven’t seen before.

If only we’d been warned. Oh, right… we were. Climate scientists had been issuing exactly these warnings for forty years, with models, with data, with public testimony that grew more urgent every year. And they were met with a coordinated, well-funded campaign of denial (even some saying ‘drill, baby, drill’). The fossil fuel industry spent billions seeding doubt, financing front groups dressed up as research institutes, discrediting scientists, and buying the political silence needed to keep drilling. The playbook was methodical, and it worked. And to our peril, decades of necessary action were lost.

Meanwhile, ordinary citizens continued to make individually rational choices within an economic system that rewarded consumption. Cheap flights, larger homes, faster delivery, higher meat consumption, and ever-expanding material expectations became the normal trajectory of prosperity. Each person’s contribution seemed insignificant, and collectively those small choices reinforced the very system everyone depended on. It’s a version of the tragedy of the commons.

We have delayed responding to one environmental crisis while living with its consequences. At the same time, a new resource-intensive industry is expanding at remarkable speed. Data centers require land, electricity, and water on a scale that many communities are only beginning to feel. And, therefore, we need to be more involved in the decision making. The question are: will decisions be based on transparent evidence, public participation, and democratic oversight? Or will communities understand the costs only after construction is complete? Unfortunately, we have already seen what happens when private incentives outweigh the public interest. Wealth becomes concentrated while environmental and social costs are distributed across everyone else. That is a pattern we must avoid while there is still time.

Why Data Centers Use Water

Racks of servers operating around the clock generate enormous amounts of heat, and without cooling, that heat would quickly damage the chips that power them. Data centers rely on five main cooling approaches, and each one trades water for energy differently.

1. Air cooling uses fans and chillers instead of water, but draws more electricity to run them.

2. Evaporative (open-loop) cooling cuts electricity use by cooling through evaporation (it’s the most water-intensive of the five.)

3. Closed-loop water cooling recirculates a sealed supply, cutting freshwater use sharply, though it costs more to install.

4. Direct liquid and immersion cooling sends coolant straight to the processors or submerges them in it, which is highly efficient and uses comparatively little water.

5. Free cooling simply draws on cool outside air, rivers, or seawater where the climate allows, minimizing both water and energy at once.

There is no perfect solution. Every cooling system involves a trade-off: reducing water consumption generally requires more electricity, while minimizing electricity use typically increases water demand. In practice, this means air cooling is often preferred in water-scarce regions, while evaporative cooling remains attractive where electricity is constrained or expensive.

From Where Does the Water Come?

Most cooling water has historically come from potable (drinking-quality) sources. One analysis estimated that approximately 57% of data centers’ direct water use is from potable water supplies. A broader industry assessment measured the question differently and put the figure at 80–90%, drawn from “blue” water sources such as lakes, rivers, or aquifers… the same sources that often supply community drinking water.

For example, Loudoun County, Virginia (home to the world’s largest data center hub) supplied roughly 1 billion gallons of water to data centers in 2023, mostly relying on treated potable water because reclaimed water capacity was insufficient. This is why local residents raise legitimate concerns: if a data center is using city water, that is water also serving homes, especially during a drought.

A growing number of data centers are shifting to recycled water. Google uses reclaimed or non-potable water at over 25% of its campuses, and AWS announced in 2023 that 20 of its data centers cool with purified wastewater. These initiatives leave more drinking-quality water for the community, but for now, these centers remain the exception rather than the rule.

Scale in Context

A medium-sized data center may consume roughly 100 million gallons per year which is equivalent to about 1,000 U.S. households annually.

Large hyperscale facilities can use 1 to 5 million gallons per day under peak conditions. At the upper end, that is as much water in a single day as a town of 30,000–50,000 people.

Nationally, all U.S. data centers combined consumed approximately 449 million gallons per day as of 2021 which is roughly 0.3–0.4% of total U.S. daily water usage.

About 40% of U.S. data centers are located in areas of high or extreme water stress, which means that these communities really suffer from the impact of the bid on their scarce water supplies.

What Do We Do?

If we choose to build data centers, then governments, together with the citizens they represent, should decide where they are located, what standards they must meet, and what obligations they owe to the communities and ecosystems around them.

Rivers, aquifers, and forests don’t recognize property boundaries and water drawn down upstream runs dry downstream regardless of whose land it crosses. Ownership has always carried that kind of obligation. What happens on one piece of land rarely stays within its borders; it ripples into watersheds, ecosystems, neighboring communities, and generations not yet born.

The good news is that we already know what better data centers look like. What’s missing isn’t the technology but the will to make responsible practice the norm instead of the exception.

What Good Looks Like

Responsible data centers already exist, built with technology available today. They draw on reclaimed water rather than drinking-quality supply wherever possible, and cool with closed-loop or direct-liquid systems rather than the thirstiest open-loop methods. Wastewater gets treated and monitored before discharge, and increasingly, that data gets published rather than kept private. A growing share of these facilities also run on renewable power, which cuts not just carbon emissions but the water that conventional generation consumes in its own right.

None of these measures alone solves the problem, but together they make one thing clear: the environmental footprint of AI infrastructure isn’t fixed. It’s a product of the choices being made in engineering, in business, in policy, and in what communities are willing to accept.

Sources:

https://www.gasilov.com/insights/data-center-water-usage

https://pmc.ncbi.nlm.nih.gov/articles/PMC12827721/

https://geoinfo.nmt.edu/publications/openfile/downloads/600-699/647/Presentations/16_AnaPrivette.pdf

https://securewater.illinois.edu/data-center-expansion-in-virginia-closing-critical-gaps-for-informed-water-planning-and-permitting/

https://www.nature.com/articles/s41545-021-00101-w

https://www.eesi.org/articles/view/data-centers-and-water-consumption

^{image from the Guardian}

Before we ask whether AI can strengthen democracy, regulate democracy, or even threaten democracy, we should ask a much simpler question. What is a citizen?Humanity has never agreed…and yet, we have built systems that assume it has.

We Assume Democracy Is Universal

We talk endlessly about aligning AI with democratic values. Transparency. Fairness. Accountability. The conversation is everywhere. Almost nobody asks the prior question.

Alignment to what world?

Every political system rests on a set of assumptions about what exists before reasoning begins: about what a person is, what authority is, what creates obligation. Philosophers call this ontology. And many of us do not examine our own assumptions and we, instead, mistake it (our assumptions) for reality.

We have assumed that democracy is a universal destination. We have not asked whether we all started from the same place. Before AI can align with democratic values, it first has to represent the world those values inhabit. And that world has never been singular.

Four Answers to One Question

Ask four civilizations the same question, and you will receive four different answers.

What is a citizen?

Athens

In Athens, citizenship belonged to only a small fraction of the people who lived there. Women, slaves, and foreign residents were excluded outright. For the men who qualified, citizenship wasn’t a status held in reserve. It was a duty performed continually in public. A citizen who withdrew from public life hadn’t made a private choice. He had failed at something the city treated as a moral obligation. Citizenship was an activity, not a possession. It was also a privilege reserved for a narrow group to begin with.

Ubuntu

Travel instead to a village shaped by the philosophy later named Ubuntu, and citizenship begins somewhere else entirely. Umuntu ngumuntu ngabantu: a person is a person through other persons. This is less a moral instruction than an account of human existence. A person does not first emerge as an independent self and later enter into relationships. Personhood is formed within relationships. Family, community, and mutual recognition do not surround the individual. They constitute the individual.

Confucian thought begins from a different premise again. Citizenship is neither a legal status nor primarily a bundle of rights. It develops through the cultivation of character, sustained by obligations to family, community, and the broader social order. A good society depends upon people learning to inhabit these roles with wisdom and restraint. Like the Athenian tradition, however, this moral vision rested upon inherited hierarchies of age, gender, and social position. Those hierarchies shaped both responsibility and authority, leaving many people without the power to define their own place within them.

Liberal democracy, drawing on Locke and later the American founders, begins from almost the opposite assumption. The individual exists prior to the state. Rights belong to the person before governments exist, and political authority acquires legitimacy only through the consent of those it governs. This remains one of the most influential political ideas in modern history. Yet it, too, emerged alongside profound exclusions. Women, enslaved people, Indigenous peoples, and colonized populations were denied the very rights the tradition described as universal. The distance between principle and practice was woven into the institutions themselves. Closing that distance required generations of political struggle led largely by those who had been excluded from the promise.

These traditions cannot be reduced to competing opinions about citizenship. Each begins from a different understanding of what a human being is. In one, the individual precedes society. In another, relationships give rise to personhood. In another still, the self is shaped through the disciplined fulfilment of social obligations. Philosophers call these foundational assumptions ontology. They sit beneath political institutions, moral systems, and legal arguments, quietly defining what counts as a person before any discussion of rights, duties, or government begins. Most of us inherit an ontology long before we learn to question it. We experience it as common sense because it has already determined the boundaries of what seems imaginable.

AI Inherits Someone’s Democracy.

Facebook in Myanmar

Facebook's content moderation system in Myanmar reflected a particular understanding of human behaviour. Harm was treated as something produced by an individual speaker through an individual post. Each post could be evaluated against a set of rules, and each decision made in isolation. That approach worked reasonably well for direct threats or explicit hate speech. It had far less to say about the dynamics of collective incitement. Hatred spread through thousands of seemingly ordinary posts, each reinforcing a broader social atmosphere without necessarily violating any single rule. By the time violence emerged, the cumulative process that had made it possible remained largely invisible to the system. The United Nations' Independent International Fact-Finding Mission concluded that Facebook had played a determining role in the spread of hatred that fuelled atrocities against the Rohingya. The failure did not begin with flawed moderation decisions. It began with the assumptions built into the system itself. A model designed around isolated individuals could not readily perceive harm that emerged from relationships, networks, and collective behaviour.

COMPAS in the United States

A similar pattern appears in COMPAS, the recidivism algorithm used in American courts. The system represented each defendant as a measurable unit of future risk. Past behavior became variables. Variables became a score. That score informed decisions about bail, sentencing, and parole, assigning weight to crimes that had not yet occurred. In 2016, a ProPublica investigation found that Black defendants were nearly twice as likely as white defendants to be incorrectly classified as high risk. The debate reached beyond statistical accuracy. COMPAS represented people as predictable risk profiles, abstracted from the historical, social, and institutional conditions that had shaped the data from which those profiles were calculated.

Facebook in Myanmar and COMPAS in American courts seem to inhabit different worlds. One moderated speech. The other informed judicial decisions. Yet both rested on the same decision. Before either system processed a single piece of data, each had already adopted an account of what a person is. That decision determined what could be represented, what counted as evidence, and what disappeared from view. No dataset arrives free of these assumptions. Every model of human behavior begins with an ontology, whether its designers recognize it or not. There is no neutral way to represent a person, because representation always begins by deciding what kind of being a person is.

Translation Has Always Been Part of Democracy

Does this mean universal principles are hopeless? Post-apartheid South Africa suggests not, but it points to a different way universality actually works.

South Africa’s Truth and Reconciliation Commission had a Western model of retributive justice available and chose not to use it. It built its process around Ubuntu instead, treating the humanity of the perpetrator and the humanity of the victim as bound together. Archbishop Desmond Tutu put it plainly: his own humanity was caught up in the humanity of others, so that when someone else was dehumanized, he was dehumanized too. Transparency, in this framework, wasn’t just a safeguard against corruption. It was a tool for collective healing, because concealment itself was treated as an ongoing harm to the relationships the commission was trying to repair.

“My humanity is caught up in your humanity, and when your humanity is enhanced... my humanity is enhanced. Likewise, when you are dehumanized, inexorably, I am dehumanized as well.” — Archbishop Desmond Tutu

This didn’t stay a one-time gesture. It became part of the permanent architecture of South African law. In S v. Makwanyane, Justice Mokgoro used Ubuntu to argue the state couldn’t kill a criminal simply to get even. In Government of the Republic of South Africa v. Grootboom, the same philosophy grounded the state’s obligation to provide housing. These cases aren’t a compromise between a universal standard and a local culture, as if South Africa had watered down human rights to make them more palatable. They show that the universal standard only became real, legally and morally, once it was translated into an idea of personhood the society already recognized. Democracy has never been exported anywhere fully intact. Everywhere it has taken root, it has been filtered through existing ideas of what a person is and what justice requires. What made South Africa unusual wasn’t that this translation happened. It’s that the people doing it did so openly, instead of pretending the transplant was seamless.

AI governance often assumes translation isn’t necessary. South Africa proves it always has been

Alignment Begins Too Late.

Long before artificial intelligence, political traditions disagreed about what kind of beings human beings are.

Mencius argued that people possess an innate capacity for compassion and justice that government should cultivate. Xunzi began from the opposite premise. Left to themselves, people are driven by appetite and self-interest. Goodness is not discovered but produced through education, ritual, and institutional discipline.

Political theories differ not only in their institutional arrangements but also in their underlying ontologies of the human person. Those ontological commitments define the concepts of rights, obligation, authority, and citizenship that follow.

That is why the language of “aligning AI with human values” is incomplete. Values are already the product of earlier commitments about what a person is, what authority is, and what obligations exist between people. Those commitments differ across societies, and no model can inherit them neutrally because there is no neutral account of personhood to inherit.

Before We Can Teach Democracy to Machines

Artificial intelligence has not created this problem. It has made visible a question that political philosophy has examined for centuries. There has never been a universally accepted account of personhood, citizenship, or political obligation. Democratic institutions are built upon particular conceptions of the human person and of legitimate authority. Because these assumptions are inherited rather than explicitly articulated, they often remain embedded within institutions without becoming objects of political scrutiny.

The same is true of AI.

A system trained to represent citizens must first represent persons. A system asked to support democracy must first inherit a particular understanding of what democracy is. The challenge, then, is not to discover a universal ontology of citizenship. There may not be one. It is to recognize that every AI system already carries assumptions about personhood, and that those assumptions belong in public debate.

Before we ask whether artificial intelligence can strengthen democracy, we should first ask a simpler question:

Bibliography

1. Aristotle. (1998). Politics (C. D. C. Reeve, Trans.). Hackett. (Original work ca. 350 BCE)

2. Boyd, R. (2013). Boundaries, birthright, and belonging: Aristotle on the distribution of citizenship. The Good Society, 22(2), 215–232. https://doi.org/10.1353/gso.2013.0011

3. Coeckelbergh, M. (2024). Artificial intelligence, the common good, and the democratic deficit in AI governance. AI and Ethics. https://doi.org/10.1007/s43681-024-00492-9

4. Government of the Republic of South Africa and Others v Grootboom and Others [2000] ZACC 19; 2001 (1) SA 46 (CC); 2000 (11) BCLR 1169 (CC). Constitutional Court of South Africa.

5. Kim, S. (2011). Confucian constitutionalism: Mencius and Xunzi on virtue, ritual, and royal transmission. The Review of Politics, 73(3), 371–399. https://doi.org/10.1017/S0034670511000386

6. Larson, J., Mattu, S., Kirchner, L., & Angwin, J. (2016, May 23). How we analyzed the COMPAS recidivism algorithm. ProPublica. https://www.propublica.org/article/how-we-analyzed-the-compas-recidivism-algorithm

7. Lee, S.-S. (2021). Paternalistic gratitude: The theory and politics of Confucian political obligation. Dao, 21, 1–22. https://doi.org/10.1007/s11712-021-09802-y

8. Locke, J. (1980). Second treatise of government (C. B. Macpherson, Ed.). Hackett. (Original work published 1689)

9. Metz, T. (2011). Ubuntu as a moral theory and human rights in South Africa. African Human Rights Law Journal, 11(2), 532–559.

10. Morrow, S., & Ryan, S. (2023). Confucius, wisdom, and political participation: Benevolence and timeliness in the Analects. Philosophy Compass, 18(1). https://doi.org/10.1111/phc3.12895

11. S v Makwanyane and Another (CCT 3/94) [1995] ZACC 3; 1995 (3) SA 391 (CC); 1995 (6) BCLR 665 (CC). Constitutional Court of South Africa, 6 June 1995.

12. Sung, W. (2016). Mencius and Xunzi on xing (human nature). Philosophy Compass, 11(11), 632–641. https://doi.org/10.1111/phc3.12363

13. Tutu, D. (1999). No future without forgiveness. Doubleday.

14. United Nations Human Rights Council. (2018, September 12). Report of the Independent International Fact-Finding Mission on Myanmar (A/HRC/39/64). Office of the United Nations High Commissioner for Human Rights. https://www.ohchr.org/sites/default/files/Documents/HRBodies/HRCouncil/FFM-Myanmar/A_HRC_39_64.pdf

15. Wong, D. B. (2025). Human nature in the ethics of Mengzi and Xunzi. In The Oxford handbook of Confucian philosophy. Oxford University Press. https://doi.org/10.1093/oxfordhb/9780199945498.013.5

A Boundary That Felt Like Nature

In 2007, a team of researchers studying a photosynthetic bacterium reported something unexpected1. They had been examining the Fenna-Matthews-Olson complex, a protein that transfers energy during photosynthesis, when they observed evidence of quantum coherence persisting within a living system. Three years later, further experiments suggested something even more surprising. The warm, noisy molecular environment surrounding the complex, long assumed to destroy coherence almost instantaneously, appeared to sustain it just long enough to improve the efficiency of energy transport.

The discovery was remarkable, although not for the reason it is often presented. Biology had always obeyed the laws of quantum mechanics. The surprise was that biology and quantum physics had become so intellectually separated that few researchers expected quantum coherence to play a meaningful role in a living system. The FMO experiments did not change the laws of nature. They challenged the way scientific disciplines had divided nature into separate domains of inquiry.

Categories Begin as Tools and Become Assumptions

Categories have an unusual characteristic. They begin as analytical tools and gradually acquire the authority of reality itself. What starts as a useful way of organising knowledge slowly becomes an assumption about how the world is organised. The distinction is subtle, yet profound. Nature does not recognise the boundaries between chemistry and biology. Evolution has never consulted a university curriculum. Cells do not know where physics ends and biology begins. Those divisions belong to us.

There is nothing unusual about this. Human beings understand complexity by dividing it into parts.2 Modern science owes much of its success to this process. We isolate variables, define disciplines and narrow the field of inquiry until phenomena become tractable. Physics concerns itself with matter and energy. Biology studies living organisms. Psychology examines behaviour. Economics models exchange. Each discipline develops its own methods, vocabulary and standards of evidence. Without this process of categorisation, scientific progress would have been impossible.

Ian Hacking showed that classificatory systems shape the phenomena they purport to describe — his “looping effects” capturing how categories acquire the authority of reality itself (Hacking, I. (1999). The Social Construction of What? Harvard University Press). Thomas Kuhn similarly observed that disciplinary paradigms constrain the questions scientists are prepared to ask, such that the boundaries of inquiry come to feel like the boundaries of the possible (Kuhn, T. S. (1962). The Structure of Scientific Revolutions. University of Chicago Press). Feminist and postcolonial scholars have sharpened this point considerably. Sandra Harding and Donna Haraway demonstrated that what counts as a legitimate scientific question has always depended on the standpoint of those empowered to ask it, and that claims to a view from nowhere have historically served to naturalise existing authority (Harding, S. (1986). The Science Question in Feminism. Cornell University Press; Haraway, D. (1988). Situated knowledges. Feminist Studies, 14(3), 575–599). Postcolonial historians of science have further shown that disciplinary organisation was entangled with colonial institutions, shaping not only what was studied but whose ways of knowing were recognised as science at all. If disciplinary boundaries came to feel like nature, that is partly because the people with the authority to draw them also had the power to make them feel that way.

Artificial Intelligence Is Exposing This Habit of Mind

Artificial intelligence is exposing this habit of mind in a way few technologies have before. For much of its history it was treated as a problem for computer science. Today, almost every important question about AI extends beyond computer science into philosophy, psychology, law, economics, political science and education. The technology has outgrown the category in which we first placed it.

Consider a question as apparently straightforward as whether a language model understands language. The architecture belongs to computer science. The concept of understanding belongs to philosophy. Language itself draws linguistics into the discussion. Cognitive science asks whether understanding requires particular mental processes. Neuroscience asks whether those processes depend upon biological mechanisms. Psychology contributes theories of representation and meaning. None of these disciplines possesses the entire question, yet each possesses part of it.

The same pattern appears elsewhere. Questions about autonomous decision-making become questions of ethics, law and political legitimacy. Debates about AI in education quickly move beyond educational technology into developmental psychology, sociology and philosophy of learning. Conversations about automation inevitably become conversations about economics, labour markets and political institutions. Even discussions about creativity now involve computer scientists, artists, neuroscientists, psychologists and legal scholars attempting to describe a phenomenon that has never belonged exclusively to any one of them.

The Questions No Longer Belong to One Discipline

Perhaps this explains why conversations about AI often feel strangely incomplete. Participants are rarely discussing the same object. They are examining different aspects of the same phenomenon through disciplinary lenses that were never designed to overlap. The engineer asks whether the model performs. The philosopher asks what it means. The lawyer asks who bears responsibility. The economist asks how incentives change. The educator asks what happens to learning. Each perspective is internally coherent. None is sufficient on its own.

This is not a weakness of expertise. Expertise remains indispensable. The difficulty arises when disciplinary boundaries become intellectual boundaries, when the organisation of knowledge begins to constrain the questions we are prepared to ask. AI does not respect those constraints. It moves effortlessly across them, carrying concepts from one field into another before our institutions have had time to follow.

The Questions Now Confronting Us Increasingly Belong to the Whole

Universities still organise knowledge into departments. Funding bodies evaluate proposals within disciplinary panels. Academic journals reward specialisation. Governments divide responsibility for AI between ministries responsible for technology, justice, education or economic affairs. These structures evolved because they solved a particular problem. They allowed knowledge to deepen. They encouraged precision over breadth. They remain extraordinarily effective for questions that can be isolated within a single domain3.

The photosynthetic bacterium never crossed the boundary between biology and quantum physics. That boundary existed only in the way we had organised knowledge. The discovery did not overturn the laws of nature. It revealed that our categories had become more rigid than the world they were intended to describe.

AI may be doing something similar. It is not simply asking us to build new technologies. It is asking whether the intellectual architecture through which we have organised knowledge still reflects the world we are trying to understand.

Before I began my career in international development, I worked in investment banking, which paid my way through graduate school. While studying at Columbia, I came across an opening for an office manager at Médecins Sans Frontières in New York, essentially the same role I was already doing at Goldman Sachs, though at a fraction of the salary.

During the interview, after we’d covered my experience, I was asked what I thought would be the biggest challenge if I got the job. I knew, of course, that interviews tend to reward performance over candor, and that the expected answer isn’t always the honest one. But I believed that an organization built on humanitarian principles might value integrity over a well-rehearsed response. So I said the biggest challenge would be getting used to the pay cut. The interviewer’s expression told me it was the wrong answer.

What Institutions Recognize

Every institution that hires people faces a version of the same difficulty. It has to make consequential decisions about individuals it barely knows. Since future performance can’t be observed in advance, it relies on proxies: interviews, résumés, references, credentials, personality assessments, and increasingly AI systems, each attempting to infer qualities that resist direct measurement.

That an institution needs proxies isn’t surprising. What’s more interesting is the next step: deciding which proxies count as evidence of merit. And the deciding factor, in practice, tends to be recognition rather than capability itself. Institutions don’t simply reward capability. They reward capability they know how to recognize.

That distinction is small to state and large in its consequences. Recognition depends on legibility: on whether a candidate’s judgment, competence, and potential show up in a form the institution already knows how to read. A candidate whose talent doesn’t arrive in a familiar shape can be just as capable and still register as a worse fit, not because the institution is being dishonest, but because it has no mechanism for seeing what it isn’t already looking for.

Why does a one-hour conversation get treated as a window into judgment? Why does confidence so often read as competence? Why are some of the most honest answers to interview questions also the ones most likely to leave the wrong impression? Perhaps interviews aren’t really designed to reveal who a candidate is but rather whether the candidate is understandable to the institution.

Research supports this assertion. A 1998 meta-analysis by Schmidt and Hunter, synthesizing 85 years of personnel selection studies, found that unstructured interviews have a mean validity of roughly 0.38 in predicting job performance, lower than work sample tests, structured interviews, or even simple cognitive ability assessments. A 2013 study by Dana, Dawes, and Peterson at Yale and Carnegie Mellon went further, finding that unstructured interviews didn’t just fail to add predictive value but they actually diluted the value of better information that was already available.

(Schmidt & Hunter, Psychological Bulletin, 1998; Dana et al., Judgment and Dec)

The Boundaries of Originality

Organizations consistently say they’re looking for independent thinkers: people with creativity, initiative, curiosity, and a willingness to question assumptions. Yet interviews often reward candidates who correctly identify the expected answer, read the room, and demonstrate fluency in the institution’s unwritten rules. Whole industries exist to teach candidates exactly those rules: how to structure a behavioral story, how to carry oneself, how to signal confidence and cultural fit convincingly.

There’s a reason for this. Institutions rarely want unlimited originality. They want originality that’s legible on their own terms, originality that improves an existing system rather than questions the system itself. In practice, the interview can end up testing fluency in a particular social language more than it tests the answer to the question actually asked, a language whose grammar rewards candidates who already know the rules.

Years into my career, a performance review described me as “not only an out-of-the-box thinker, but in the stratosphere,” a line I’ve always appreciated for the contradiction it points to. Across my career, organizations valued my ability to question assumptions, connect ideas across disciplines, and approach problems from unexpected directions. Those qualities became assets once I was already inside the institution. The hiring process had measured something else: largely, whether I knew when not to question the assumptions built into the interview itself. I don’t think this is accidental. Legibility, once again, is doing the work. The version of me that got hired was the version that could perform originality in a shape the institution already knew how to recognize as originality.

The research supports the suspicion. A study by Mueller, Goncalo, and Kamdar in the Journal of Experimental Social Psychology found that expressing creative ideas during an interaction was, somewhat counterintuitively, negatively associated with perceived leadership potential, even in roles that explicitly called for creative problem solving. The effect held across multiple studies. Organizations that say they’re hiring for creativity may, without intending to, be selecting against the illegible version of it.

(Mueller, Goncalo & Kamdar, Journal of Experimental Social Psychology, 2011.)

What AI Inherits

Much of the public conversation about AI in hiring has focused on algorithmic bias, and for good reason. But bias is only part of a larger story, and legibility is the more precise name for what’s actually happening.

The deeper difficulty is that AI can’t distinguish between a successful employee and an employee who successfully embodied the institution’s existing preferences and biases. If every promoted executive made excellent decisions, the model learns excellence. If every promoted executive simply looked and sounded like previous executives, the model learns institutional preference instead. It has no way of knowing which pattern it’s tracing, and in many cases neither do the people who built the institution in the first place.

Some organizations now route candidates through AI-mediated interviews before they ever speak with a human being. These systems analyze language, speech patterns, pauses, facial expressions, eye movements, head movements, and other behavioral signals believed to correlate with confidence, engagement, or trustworthiness. The precision of that measurement says nothing about the validity of what’s actually being measured. A finer instrument isn’t the same as a more accurate one if the assumptions underneath it go unexamined.

AI-mediated interviews break confidence down into measurable variables: a pause counts as hesitation, eye contact counts as engagement, head movement counts as confidence. None of that is established fact. Someone decided those behaviors mean what the model assumes they mean, then wrote that assumption into code. Once it’s running as software, it looks like science. It isn’t. It’s an institution’s preferences, expressed in numbers instead of opinions.

The clearest illustration came from Amazon. Reuters reported in 2018 that the company had quietly shelved an AI recruiting tool it had been developing since 2014. The system had trained on a decade of résumés submitted to Amazon, most of which came from men, a reflection of the technology industry’s existing gender composition. The model learned to treat male candidates as preferable, penalizing résumés containing terms like “women’s,” as in “women’s chess club captain,” and downgrading graduates of two all women’s colleges. Amazon’s engineers removed those specific signals, but concluded they couldn’t guarantee the system wouldn’t find other ways to reproduce the same pattern, and the tool was eventually abandoned. (Jeffrey Dastin, Reuters, October 10, 2018.)

The Amazon case is instructive because the tool worked exactly as designed. It studied the institution’s own record of success and reproduced that logic at scale. When the engineers looked closely enough, they found they couldn’t cleanly separate the institution’s preferences from its theory of merit, because the two had always been entangled. The tool hadn’t learned bias so much as it had learned legibility, absorbed directly from a decade of records about which résumés the institution already knew how to recognize as promising.

The problem was not that the machine had learned the wrong lesson. The problem was that it had learned the right one.

Every model eventually converts human ambiguity into measurable variables. Someone decides that one behavior signals confidence while another signals uncertainty; someone decides that a particular communication style predicts future performance. Those decisions become mathematics, and mathematics acquires the appearance of objectivity. The numbers feel precise even as the assumptions underneath them stay thoroughly human.

Professionalism presents itself as universal, but historically it’s been anything but: the suit, the accent, the handshake, the confidence that reads as competence rather than performance, the tone that’s firm but never aggressive, an answer structured the way a STAR-format training seminar teaches. These are bourgeois conventions, artifacts of a particular class culture that took shape alongside twentieth-century managerial capitalism and gradually hardened into the default picture of what professionalism looks like. That doesn’t make the culture wrong. It makes it particular, one way of doing things among many that happened to become the way institutions learned to recognize competence. AI doesn’t know the difference between particular and universal. It inherits the specific and reproduces it as though it were a fact about human ability, encoding a class dialect with the same confidence as any other pattern in its training data.

HireVue, one of the largest providers of AI-mediated video interviews, with more than 19 million interviews conducted for over 700 employers, quietly discontinued its facial analysis feature in 2021. The company had used algorithms to infer traits from candidates’ facial expressions. That this approach was considered not just plausible but worth deploying at scale says something about how deeply those assumptions had taken hold, well before anyone asked whether the assumptions were sound. HireVue removed the feature after internal research found that advances in language analysis had made visual analysis redundant. The move was framed as a technical refinement. It also amounted to a quiet acknowledgment that the assumptions underlying facial analysis hadn’t held up as well as hoped. (Roy Maurer, SHRM, February 3, 2021.)

Candidates who fall outside the legible register aren’t necessarily less capable. They’re often just harder to read, for the institution, for the interviewer, and now for the algorithm too. Someone navigating an interview in a second language. An autistic candidate whose communication style diverges from neurotypical norms. Someone from a high-context culture where directness reads as impolite. Someone from a background where the conventions of professional self-presentation were simply never taught. Any of these candidates might be marked down not for what they can’t do, but for how they come across, a distinction an institution’s own hiring data has no way of preserving once it’s been folded into a model.

The institution’s theory of merit

What AI ultimately makes visible is an institution’s own theory of merit: the embedded answer to a question most organizations have never explicitly asked, which is what kind of person succeeds here. That answer doesn’t live in any policy document. It emerges through thousands of individual decisions accumulated over years, shaped by who got hired, who got promoted, who was described in a performance review as a natural leader versus a poor cultural fit, until the pattern hardens into something that feels less like a choice and more like common sense.

AI learns that common sense directly from the data. It’s trained on historical outcomes, and those outcomes carry within them every assumption the institution has made about capability, potential, and fit. Left unchecked, it has no way to distinguish between what a person is genuinely able to do and the behaviors an institution has historically learned to recognize as evidence of that ability. It can’t separate the person from the representative, the individual from the type an institution has, over time, come to expect, because that separation was never made explicit in the records it learned from.

Perhaps that’s AI’s most useful contribution to hiring so far; It forces institutions to examine assumptions that had become so familiar they’d stopped looking like assumptions at all, and to ask whether a given process rewards capability, conformity, judgment, institutional fluency, or some blend of all four.

Every institution develops its own working definition of merit. It rarely writes that definition down but it emerges through thousands of decisions about who gets hired, who gets promoted, who is described as a leader, and who is told they aren’t quite the right fit. Over time those decisions stop looking like decisions but start to look like ‘the norm’. And AI learns this…and treats it like common sense.

It doesn’t know whether the institution has identified the qualities that actually matter or simply become very good at recognizing certain kinds of people. It only knows that those people were rewarded.

What to do about it

The more interesting question was never whether AI can identify the best candidate. It’s how our institutions came to define “best” in the first place, and whether they’re willing to look closely enough at their own hiring data to find out. Part of the answer, at minimum, is that most institutions have been answering it by accident, folding two different questions into one interview and calling the result merit.

One practical fix is to pull those two questions back apart. Institutional fit and capability are related but distinct, and treating them as interchangeable is where many hiring processes quietly lose precision. An institution that’s clear about what it’s selecting for, someone who understands its culture, communicates in its idiom, and can represent it credibly, is then free to ask a separate and more rigorous question: can this person also do the work? These two assessments call for different methods, different questions, and different kinds of evidence. Conflating them produces something that feels efficient but is, in practice, less honest than either assessment made on its own terms. A process unwilling to be clear about what it’s measuring can’t be made fair, no matter how sophisticated the technology behind it becomes, and no matter how well that technology has learned to recognize what the institution already valued. Perhaps the interview was never designed to discover the best candidate. Perhaps it was designed to discover the candidate the institution already knew how to recognize.

When I was growing up in the 1970s, Saturday mornings meant one thing: The Jetsons. Every week, I watched George Jetson make video calls, carry a portable communicator, talk to intelligent machines, and live in a world that seemed impossibly futuristic. At the time, it was just a cartoon.

Decades later, I make video calls every day. I carry a computer in my pocket. I pop a ready made meal in the microwave. Those technologies didn’t appear by accident. Before they were engineered, they were imagined. (Note to product developers: I’m still waiting for the one that dresses me while I stand on a conveyer belt.)

The technology didn’t emerge because someone copied The Jetsons. But generations of scientists, engineers, and entrepreneurs grew up immersed in stories that expanded their sense of what was possible. Science fiction has always done more than entertain us. It has inspired people to build.

Lately, I’ve been wondering what we’re inspiring people to build now. Much of the public conversation about AI is framed around resistance. Slow it down. Pause it. Ban it. The frustration is understandable. For years, some of the largest technology companies have gone unchecked and treated our personal data, our attention, and even our collective knowledge as resources to be extracted and monetized.

At the same time, most of us expressing those concerns are doing so through technology. We work online. We connect over Zoom. We navigate with our phones. We rely on cloud services every day. So, this isn’t a rejection of technology. It’s a rejection of particular choices about how technology has been built.

However, resisting sometimes gets us into a rut. Resisting something tells us what we don’t want. And we get so busy fighting something that we are missing the opportunity to focus our energy on creating something else.

The technologies that shaped my childhood weren’t built because someone spent decades arguing against the present. They were built because people became captivated by a different future. They imagined video calls, portable communicators, intelligent assistants, and yes, even microwave dinners. Some of those ideas remained fiction but many didn’t.

If imagination helped shape the technologies we have today, then it seems worth asking what kind of future we’re putting in front of the next generation of builders. What stories are we telling them? What problems are we inviting them to solve? What values are we encouraging them to embed in the systems they create?

Fortunately, there are already people answering those questions in interesting ways. A few days ago I came across a Swiss company called Infomaniak and its AI assistant, Euria. Your conversations aren’t saved or used to train future models. Your data isn’t resold. The company’s infrastructure runs on renewable energy. Even the heat generated by its servers doesn’t go to waste. It’s fed into Geneva’s district heating network, helping warm thousands of homes.

As children, many of us imagined flying cars, video calls, and intelligent machines. Those stories inspired generations of engineers, entrepreneurs, and inventors. The next generation deserves stories that inspire them not only to build extraordinary technology, but to build it on a different set of values.

Technology has never been value-neutral. Every system reflects the priorities of the people who create it. If we can imagine technology that respects privacy, serves the public good, strengthens communities, and treats sustainability as a design requirement rather than an afterthought, then we can build that too.

The future isn’t something that simply happens to us. It is something we build, one decision at a time. I'd love to see more examples like this. If you've come across companies building AI differently, share them in the comments. Let's give ourselves, and the next generation of builders, more futures to imagine.

For more than two thousand years, Jain philosophers have told the story of several blind men encountering an elephant. One reaches for the trunk and concludes the elephant is like a snake. Another places his hands on a leg and insists it is a tree. A third grasps the tail and is certain it is a rope. Each has encountered the same elephant. Each describes something real. Yet none has encountered the whole.

The story is often read as a lesson in tolerance or pluralism. But, it can also be understood as a lesson in method. Some phenomena resist complete description from any single framework. Every framework reveals something. Every framework also leaves something unseen.

Consciousness has occupied this territory for centuries.

For much of Western intellectual history, consciousness was treated as an exclusively human property. René Descartes famously described animals as automata, complex biological machines that responded to the world without experiencing it. Few scientists or philosophers would defend that position today1. Research in comparative cognition has documented planning in corvids2, problem solving in octopuses, self-recognition in dolphins3, emotional behaviour in elephants, and sophisticated forms of communication, cooperation, and memory across a remarkable range of species. These discoveries have not produced a universally accepted definition of consciousness, but they have repeatedly forced us to reconsider where we believed its boundaries lay.

Plant biology has introduced another layer of complexity. Plants exchange chemical signals, communicate through mycorrhizal networks, alter their behaviour in response to environmental conditions, and exhibit forms of memory that would have sounded implausible only a generation ago4. Whether these phenomena should be described as intelligence, cognition, or consciousness remains an active area of debate. The disagreement itself is informative. Before we can decide where consciousness exists, we have to decide what we mean by consciousness.

That question became unexpectedly public in 2022 when Google engineer Blake Lemoine argued that LaMDA, one of the company’s experimental language models, was sentient5. His conclusion was rejected by Google and by most researchers working in artificial intelligence. The discussion quickly centred on whether Lemoine had mistaken statistical fluency for subjective experience.

The episode raised a more interesting question than whether Lemoine was right.

How would we know?

Human beings never observe consciousness directly…we infer it. We infer it in other people through language, behaviour, memory, emotion, and relationship. We infer it in animals through increasingly sophisticated observations of cognition and behaviour. Artificial intelligence has simply extended that question to another domain. The problem itself is not new.

In trying to answer whether machines are conscious, we have found ourselves returning to a much older question: what is consciousness?

That question has occupied philosophers, theologians, psychologists, neuroscientists, and contemplative traditions for centuries without producing anything approaching consensus. Some understand consciousness as an emergent property of biological systems. Others begin with subjective experience. Some approach it through information processing, others through direct observation of awareness itself. Each framework has generated important insights. Each has also defined, often implicitly, what counts as evidence, what questions deserve investigation, and what kinds of explanations are considered satisfactory.

Perhaps that is why the Jain parable continues to resonate. The elephant was never simply an elephant. It was a reminder that complex phenomena often exceed the explanatory power of any single framework.

Artificial intelligence has not resolved that problem. It has returned it to the centre of public conversation. Every discussion about machine consciousness quickly becomes a discussion about human consciousness. Every attempt to decide whether a machine understands, reasons, or experiences forces us to examine the assumptions through which we have been making those judgments all along.

Artificial intelligence may or may not one day become conscious. That remains an open question. What already seems apparent is that, in trying to answer it, we have been compelled to examine our own theories of mind with unusual clarity. The debate has made visible the philosophical, scientific, and cultural frameworks through which we have always understood consciousness, often without realizing we were relying on them.

Perhaps that is the mirror.

Not because it tells us what machines are, but because it reveals the assumptions we bring to one of humanity’s oldest questions. The elephant has been standing in front of us for millennia. Artificial intelligence has simply given us another place from which to look.

When people invoke Plato’s Cave today, it is usually to make a point about misinformation. The cave has become shorthand for fake news, algorithmic echo chambers, and political polarization. The lesson, we are told, is simple: question what you see.

I suspect Plato was asking a more difficult question.

The prisoners do not choose the shadows projected onto the wall. Someone else does. Their understanding of reality is mediated through a system they neither designed nor can inspect. The problem is not simply that they mistake shadows for reality. It is that they have no way of asking who decided which shadows would be shown in the first place.

Read this way, Plato’s Cave is not simply about perception. It is about institutions.

Every civilization has developed institutions responsible for deciding what knowledge should survive. Religious authorities once occupied much of that role. Universities gradually inherited it. Today, publishers, academic journals, funding bodies, curriculum committees, accreditation agencies, and professional associations all participate in the same enterprise. Together, they perform one of civilization’s most important functions: they organize the overwhelming complexity of human knowledge into something that can be preserved, taught, and expanded. Without these institutions, every generation would have to begin again.

This is not an argument against institutions. It is an argument for understanding what institutions actually do. They do not simply preserve knowledge but they also determine what becomes legitimate knowledge. No curriculum can teach everything. No university can fund every idea. No journal can publish every paper. Every institution must decide what passes through the filter and what remains outside it. Those decisions are rarely malicious. They are often thoughtful and necessary. But they are still decisions.

Science offers perhaps the clearest example. We often describe scientific progress as a steady accumulation of evidence, each discovery adding another brick to an ever-growing wall of knowledge. The history of science tells a more complicated story. New ideas compete with established paradigms. Funding influences what is studied. Journals determine what is publishable. Professional communities decide which claims deserve serious attention.

Thomas Kuhn famously argued that science progresses through periods of stability interrupted by paradigm shifts. Before continental drift became plate tectonics, before Semmelweis convinced medicine that physicians should wash their hands, before Barbara McClintock’s discovery of transposable genes became accepted biology, each existed outside the prevailing consensus. Eventually these ideas entered the canon. Initially, they existed beyond it.

This is not evidence that science is broken.

It is evidence that institutions face an unavoidable tension. They must preserve reliable knowledge while remaining open to knowledge that challenges their own assumptions. That balance has always been difficult. Artificial intelligence introduces a new dimension to that challenge. Large language models are frequently described as having been trained on “the knowledge of humanity.” The phrase sounds universal but it isn’t. They (LLMS) are trained primarily on humanity’s institutional record of knowledge. Books that were published. Research that secured funding. Papers that survived peer review. Archives that were preserved. Languages that accumulated sufficient written material to become part of the digital record. Before an idea reaches AI, it has often already passed through multiple layers of institutional judgment.

Artificial intelligence does not simply inherit our knowledge.

It inherits our institutions’ decisions about what counted as knowledge.

That has profound consequences.

One of AI’s greatest strengths is its ability to synthesize and reason across enormous bodies of accepted knowledge. It can identify patterns no individual could hope to assemble alone. Yet this strength also reveals one of its limits. AI’s centre of gravity lies within institutional consensus because institutional consensus overwhelmingly constitutes its training data.

Before every revolution in knowledge, the genuinely transformative idea and the merely mistaken one often look remarkably similar. Distinguishing between them has never been a statistical problem. It has required judgment, skepticism, experimentation, and, occasionally, a willingness to question the institutions that define legitimate knowledge itself.

This brings us back to education.

For centuries, schools have served many purposes. They have transmitted culture, prepared citizens, expanded literacy, developed expertise, and opened pathways into professions. One of their central functions has been to introduce each generation to humanity’s accumulated knowledge.

Artificial intelligence changes the economics of that task. For the first time in history, access to organized knowledge is no longer scarce. If AI becomes increasingly capable of transmitting institutional knowledge, then education may need to reconsider where its greatest value lies. Perhaps its purpose is no longer primarily to teach students what humanity knows. Perhaps its purpose is to help them understand how humanity comes to know. How evidence becomes accepted. How paradigms shift. How institutions shape legitimacy. How assumptions become invisible precisely because they have become institutionalized.

In other words, the challenge for education may no longer be helping students leave Plato’s cave.

It may be helping them understand who built it, why it was built that way, and how every generation inherits the responsibility of deciding which shadows belong on the wall.

In the Netherlands, for several years before 2020, a government algorithm called SyRI (Systeem Risico Indicatie) quietly cross-referenced citizens’ tax records, employment history, housing data, and benefits claims to generate a risk score. A high enough score got you flagged as a likely welfare fraudster. The system was rolled out in Rotterdam, Eindhoven, Haarlem, and Capelle aan den IJssel — all predominantly low-income neighborhoods, with Rotterdam carrying the country’s highest poverty rate, as Human Rights Watch documented. If it flagged you, you didn’t receive a letter explaining why. You couldn’t inspect the model. You couldn’t find out which data point tipped the scale. In February 2020, the District Court of The Hague ruled SyRI unlawful — a violation of Article 8 of the European Convention on Human Rights, the right to private life — precisely because there were no meaningful transparency safeguards and no real way for a flagged person to challenge what had been decided about them. The UN Special Rapporteur on extreme poverty and human rights, Philip Alston, had submitted a human rights analysis to the court concluding that SyRI discriminated against the poorest members of society; he later called the ruling “one of the first times a court anywhere has stopped the use of digital technologies... by welfare authorities on human rights grounds.”

SyRI illustrates the governance gap this piece is about. The person flagged (whose life was assembled into a risk score without their consent, whose appeal had nowhere real to go) is the one most affected by automated decision-making. They are also the one most absent from the institutional debates that shape it.

The analysis proceeds from their position, not from the vantage point of regulators or supranational bodies. The deeper risk isn’t which entity, state or corporate, holds the relevant monopoly at any given moment. It’s that concentrated capital and political power tend to recapture whatever regulatory framework is built around them; a dynamic already visible in the lobbying that has progressively eroded the EU’s own AI regulations. It’s worth being precise about what SyRI represents here: not authoritarian overreach, but ordinary administrative modernization within a liberal democracy, built by the Netherlands’ own welfare bureaucracy in pursuit of efficiency.

This matters because it shifts the question away from whether a given government is trustworthy today, and toward whether the underlying architecture (i.e. data collection, automated inference, limited contestability) stays protective when political conditions change. Institutions that function adequately under favorable governance can become instruments of harm under different ones. The absence of structural safeguards isn’t visible until the conditions that made them unnecessary no longer hold.

AI as Democratic Infrastructure: Why Sovereignty Isn't Enough

AI isn’t merely a tool; it’s infrastructure for the future. The European Commission’s Special Eurobarometer 554 (2024/2025) captures the public’s ambivalence: 55% of Europeans expect AI to positively shape daily life over the next 20 years, yet 35% expect negative effects; 62% believe it will benefit the economy, yet 66% think it will destroy more jobs than it creates; 84% say it requires careful management to protect privacy and ensure transparency.

Decades of research, including a 2025 systematic review of 51 studies in Information Systems Frontiers. show that people’s stated privacy concerns often diverge sharply from their actual data-sharing behavior. Wanting meaningful oversight of a system is a different question from willingness to share data day-to-day, and survey sentiment shouldn’t be read as a direct demand for ownership. But when algorithms shape access to jobs, credit, healthcare, and public services, they function less as discrete software products and more as infrastructure governing access to civic life — and infrastructure citizens can’t inspect or contest raises accountability questions that are distinct from questions of efficiency or innovation.

Europe’s answer, gathering speed by the month, is sovereignty: build the European cloud, fund the European model, legislate the European standard, stop renting democratic infrastructure from Silicon Valley on Silicon Valley’s terms. The case for concentrating core digital infrastructure in a handful of U.S. corporations is losing empirical and political credibility, and the policy response is a reasonable first-order correction. The question is whether it addresses the underlying structural problem, or just relocates it.

A European AI champion, however well regulated, can concentrate power in much the same way as today’s dominant technology companies. What matters is not where the infrastructure is based, but who owns it, who governs it, and who can challenge its decisions. Gaia-X shows why that distinction matters. Launched in 2019 by France and Germany as Europe’s flagship cloud sovereignty initiative, its stated aim was to reduce dependence on Amazon, Google, and Microsoft. The major American hyperscalers were ultimately admitted into its governance structure anyway, and a 2025 peer-reviewed study by Andreas Baur in Information, Communication & Society concluded the project had been effectively captured by the very companies it was created to counter. One founding member later called it “a crushing failure... an industrial disaster” that bought the hyperscalers years of additional dominance while Europe’s cloud market position kept weakening.

None of this changed the citizen’s position. The person affected by an AI system still can’t inspect the model, understand how a decision was made, or meaningfully challenge it. Replacing one provider with another doesn’t redistribute power if ownership and governance stay concentrated. The organization operating the infrastructure changes. The relationship between institution and citizen does not.

Gaia-X is a specific case of a broader pattern: capital and influence accumulate around infrastructure that becomes strategically important, public or private, unless the governance is designed to resist capture. Sovereignty by itself is an incomplete answer because it addresses who operates infrastructure but not who owns it, governs it, or whether power is distributed so no single institution becomes unaccountable.

There’s a harder version of this problem that the sovereignty framing doesn’t touch at all: sovereignty has never been a guarantee against a country’s own government turning on its own people. Weimar Germany was a sovereign nation-state. So was Mussolini’s Italy in the 1920s. Both had citizens who understood themselves as living under national self-determination, right up until the infrastructure of the state (e.g. its registries, its bureaucracies, its administrative machinery) was turned against the populations it had been built to serve. “European” and “sovereign” describe who operates a system. They say nothing about who it protects when power inside that sovereign territory changes hands. A citizen-owned data cooperative resists capture by a foreign hyperscaler and by a domestic authoritarian turn for the same structural reason: neither can simply seize what it doesn’t own or control. A sovereign European AI champion, if its data and governance are centralized in the state or in a handful of firms, protects against the first threat and does nothing about the second.

The current sovereignty debate largely accepts that the choice is between different institutions owning the same centralized infrastructure. That assumption deserves scrutiny. The more fundamental question is whether centralized ownership should be the default at all. Cooperative ownership, public-interest governance, and citizen stewardship offer another possibility, and they remain nearly absent from the debate. The question isn’t simply who should own AI. The question is whether ownership itself can be structured differently.

What the Barometer Data Reveals

The barometer numbers establish a baseline, but read alongside how the AI governance debate is unfolding in practice, they point to four tensions any substantive policy response will need to address:

A Third Governance Model: Citizens as Stewards

Most AI policy conversations assume digital infrastructure must be run by either states or corporations. That binary deserves scrutiny. A third option exists: citizens as stewards, not merely users or voters. It's the blind spot at the center of the current sovereignty debate: almost nobody in mainstream governance discourse asks whether this is viable, despite mounting evidence that it can work.

The case for that third option isn’t primarily about fairness in distribution, though that matters, it’s about structural resilience over time. Systems built for administrative efficiency have repeatedly become systems of surveillance and control when the governments running them shifted priorities, and SyRI already makes that case: a functioning liberal democracy still needed a court to stop it. Governments change, coalitions shift, rights expand and contract across electoral cycles. The question that matters isn’t whether today’s government or company can be trusted with AI infrastructure. It’s whether the architecture stays protective when tomorrow’s holders of power can’t be.

History is not gentle on this point, and the sharpest example again involves the Netherlands. Going into the Second World War, the Dutch civil administration was among the most precise and comprehensive population registries in Europe - the same administrative tradition that, decades later, would build SyRI. When Germany occupied the country in 1940, that registry gave the occupying authorities a near-complete record of who was Jewish and exactly where they lived. Historians point to the precision of that civil administration as one specific reason the Netherlands ended up with the highest proportional Jewish death toll in Western Europe, higher than Belgium or France, where records were patchier and evasion easier. The infrastructure itself was neutral administrative machinery, built by a democracy, for entirely unobjectionable purposes, years before anyone occupied the country. It became lethal not because it changed, but because who controlled it changed. The pattern is consistent: comprehensive data infrastructure is not dangerous because of who runs it today. It is dangerous because it persists, ready to be pointed at whoever holds power tomorrow.

These aren’t proposals for a future that doesn’t yet exist. The institutions operating on exactly these principles are already built, already running, and already demonstrating that citizen stewardship works at real scale. The following include some intriguing examples.

In Switzerland, MIDATA is a citizen-owned health data cooperative founded in 2015, where members store medical records on an encrypted platform and decide, case by case, whether to share them with researchers. It’s governed by a member general assembly and an elected ethics board reviewing every access request.

In New Zealand, Te Pā Tūwatawata is a decentralized, Māori-owned data storage network developed by Te Kāhui Raraunga, built so iwi and communities hold their own data under their own cultural governance rather than surrendering it to a corporate or state custodian.

In New York City, the Drivers Cooperative puts thousands of ride-hail workers on the other side of the algorithm: members earn more per trip, receive profits as dividends, and hold one-member-one-vote control over the platform.

In Bengaluru, Namma Yatri is an open-source, zero-commission mobility platform built on India’s open digital protocols, designed to route around the extractive commission model — by mid-2025 it had completed over 100 million rides, channeling more than ₹1,600 crore (roughly $190 million USD) directly to over 600,000 registered drivers, commission-free, according to company data reported by Business Standard and the Economic Times. And at the physical layer itself, Guifi.net in Spain and NYC Mesh in New York are community-owned networks built and maintained collectively by their members rather than telecom monopolies — Guifi.net alone is documented as the world’s largest community network, with more than 37,000 active nodes and roughly 71,000 km of wireless links across Catalonia and the Valencian Community.

Then there’s the public-interest foundation model: the Wikimedia Foundation has run the global infrastructure behind Wikipedia for over two decades at genuine internet scale, with board accountability and transparent funding, without shareholder pressure, ad-driven data extraction, or a profit motive. It’s proof that public-interest governance can run infrastructure used by hundreds of millions of people at once.

The failure was never a lack of alternatives. Cooperative ownership, public-interest technology foundations, participatory data governance, open-source platforms with real community oversight, and data trusts already exist. The failure is that AI governance discourse proceeds as though none of this exists, treating the state-vs-corporate binary as exhaustive when it plainly isn’t. The question worth pressing: do these models remain the exception, or finally become the blueprint for how AI itself is owned and governed?

That blueprint question has a harder edge than it first appears. Cooperatives, data trusts, and community networks aren’t automatically immune to the capture dynamic that hollowed out Gaia-X. They have to be deliberately structured — legally, financially, in their governance rules — to resist being bought out, diluted, or absorbed when incentives shift and valuations rise. A democratic origin story is not a structural defense. The blueprint worth building encodes capture-resistance as non-negotiable: lock-in mechanisms on asset disposition, supermajority requirements for structural changes, legal forms like perpetual purpose trusts or cooperative statutes with asset locks that make absorption substantially harder, not merely inconvenient. Broadly distributed ownership makes concentration harder, but only when the legal and governance protections are built in from the outset rather than added after pressure begins to mount. Citizen ownership doesn’t guarantee benevolent governance. It makes the concentration of power structurally harder, regardless of who’s seeking it.

The Window Is Already Narrowing

Public appetite is real. Political will is shifting. The EU AI Act (published in the Official Journal on 12 July 2024, entered into force 1 August 2024, binding on a fixed phased schedule through August 2027) is already mid-implementation: prohibitions on unacceptable-risk practices applied from February 2025, obligations for general-purpose AI models followed in August 2025. This isn’t aspirational language; it’s a live, legally binding timetable, sitting alongside the Data Governance Act and a wave of digital sovereignty strategies pointing the same direction.

That timetable is already under pressure. Mario Draghi (the figure behind the EU’s competitiveness agenda) has called for a pause on the Act’s high-risk provisions and radical simplification of GDPR; more than 40 European CEOs, including leaders from ASML, Philips, Siemens, and Mistral, formally requested a two-year clock-stop. This is the capture dynamic named in this piece’s title, operating in real time: not a hypothetical future risk, but an active, well-resourced campaign to walk back citizen protections as political and economic incentives shift. The window is narrow not as a metaphor but because this is the mechanism that closes it — concentrated private interests moving quickly to lock in deregulation before democratic governance can take root. The legal schedule holds for now, but the political outcome it represents isn’t settled; it still has to be actively defended. There remains a wide gap between governments talking about rights and citizens exercising power. Closing that gap isn’t a policy task alone. It’s a movement-building one.

From Alternative Models to AI Infrastructure

Consider the person flagged by SyRI. Their benefits were flagged by a model they couldn’t see, built on data they hadn’t consented to share, producing a score they had no mechanism to challenge. What would be different if the infrastructure looked like MIDATA instead of a government black box? They’d hold their own data. They could grant access conditionally, revoke it, see what was inferred and by whom. The score wouldn't disappear. Some form of risk assessment is probably unavoidable. What would disappear is the asymmetry. They’d be a participant in the system rather than a subject of it. That’s the condition democratic AI governance is meant to produce, and it isn’t the default anywhere yet.

The examples that come closest - MIDATA, Te Pā Tūwatawata, the Drivers Cooperative and Namma Yatri returning surplus to the people doing the work, Guifi.net and NYC Mesh treating connectivity as shared infrastructure - aren’t evidence the problem is solved. They show the architecture is viable at the data and platform layer. What’s unproven is whether this model can reach the capital-intensive layer of chips, compute, and foundation models, where ownership concentration is far more entrenched. To illustrate the scale: Nvidia commands an estimated 80–90% of the AI accelerator market by revenue, and the four largest hyperscalers (i.e. Amazon, Google, Meta, Microsoft) collectively spent over $416 billion on AI infrastructure capex in 2025 alone, a 66% increase from 2024. That’s a fundamentally different order of capital concentration than anything the cooperative and data trust examples above were built to contest.

What the existing examples share, beside the model, is difficulty: undercapitalization, dependence on grant cycles that don’t match infrastructure timelines, political environments that reward procurement from large incumbents over patient institution-building. None of them scaled by being easy. They scaled because specific people built unglamorous governance structures and kept them running. And there’s an evidence-based limit beyond funding: a 2022 study by Bunders, Arets, Frenken, and De Moor in the Journal of Co-operative Organization and Management found platform cooperatives are most feasible in taxi rides and professional services, but significantly harder to sustain in food delivery, home care, and microtask work… precisely the sectors where gig exploitation tends to be worst. That’s not a solvable coordination problem; it’s a structural constraint on how far this architecture has actually been proven.

That institutional layer is largely absent from most AI governance conversations, which tend to focus on regulation and standards rather than who will own and operate the alternative. Regulation matters - it can constrain harmful uses and, at its best, enable alternatives: GDPR’s data portability right is a real mechanism making it easier to move data out of incumbent platforms and into cooperatives or trusts. But building citizen-owned alternatives is necessary and not sufficient. They need structural defenses against recapture from the outset: legal non-dilution protections, bylaws that can’t be quietly voted away by a future board, mission-lock mechanisms that survive both a change of political wind and an attractive acquisition offer. Gaia-X showed how fast a public-interest initiative can be hollowed out when incumbents are invited into governance, and the Draghi-backed pressure on the AI Act is a reminder that regulatory gains can be unwound before the implementing infrastructure is in place. What isn’t structurally protected will eventually be absorbed by whatever has the most capital and patience. Regulation alone can’t create a Drivers Cooperative. It can’t build the data trust that would let the person flagged by SyRI own their own profile. Someone has to construct the thing, capitalize it, govern it, and defend it against the much easier option of selling it or shutting it down when the funding runs out.

The concentration of power in AI systems isn’t structurally inevitable. It reflects specific, still-contestable choices about ownership, capital, and institutional design. The flagged person’s situation is the concrete test case: if the infrastructure being built right now can’t eventually give them legibility into a system that shapes their life, then what’s being built doesn’t function as a public good. The institutional models that could change that already exist — built, governed, and defended, at the data and platform layer, by people with far fewer resources than the bodies now drafting AI sovereignty frameworks. The question the policy debate hasn’t been forced to answer: if these models demonstrably work, why aren’t they the starting point? Why are they treated as niche experiments at the margins of a conversation still dominated by the same hyperscalers whose ownership arrangements are the problem being discussed?

^Sources

^{SyRI and welfare surveillance}

• ^{Rechtbank Den Haag (District Court of The Hague), SyRI ruling, ECLI:NL:RBDHA:2020:1878, February 5, 2020}

• ^{Human Rights Watch, “Welfare Surveillance on Trial in the Netherlands,” November 8, 2019}

• ^{OHCHR, “Landmark ruling by Dutch court stops government attempts to spy on the poor, UN expert,” February 5, 2020}

^{Public opinion on AI}

• ^{European Commission, Special Eurobarometer 554, “Artificial Intelligence and the Future of Work,” conducted by Verian, 2024/2025}

• ^{“Unraveling the Privacy Paradox: A Comprehensive Review of Factors Behind the Discrepancy in Online Concerns and Disclosure Behavior,” Information Systems Frontiers, 2025}

^{Digital sovereignty and Gaia-X}

• ^{Politico, “Anatomy of a Franco-German Tech Misfire,” November 2025}

• ^{Euractiv, “Gaia-X: A Trojan Horse for Big Tech in Europe,” November 2020}

• ^{Andreas Baur, “European Ambitions Captured by American Clouds: Digital Sovereignty Through Gaia-X?,” Information, Communication and Society, 2025}

^{Citizen-owned infrastructure examples}

• ^{MIDATA Cooperative, midata.coop}

• ^{Te Pā Tūwatawata, tepatuwatawata.io}

• ^{The Drivers Cooperative, drivers.coop, and Platform Cooperativism Consortium Directory}

• ^{Namma Yatri open data dashboard, nammayatri.in; Business Standard, “Namma Yatri Hits 100 Million Rides, Enables ₹1,600 cr Driver Earnings,” June 6, 2025; The Economic Times, June 26, 2025}

• ^{Guifi.net entry, Wikipedia; Vega et al., “A Technological Overview of the Guifi.net Community Network,” Computer Networks, 2015}

^{EU AI Act and regulatory pushback}

• ^{Regulation (EU) 2024/1689, Official Journal of the European Union, Article 113}

• ^{Euractiv, “Draghi Calls for Deep Cuts to Privacy Rules and Pause on ‘High Risk’ AI Act”}

• ^{Euronews (via AI Transparency Institute), “Draghi Calls for Pause to AI Act to Gauge Risks,” September 2025}

• ^{DIGITALEUROPE, “Our Reaction to Mario Draghi’s Report,” September 2024}

^{Platform cooperatives and AI infrastructure concentration}

• ^{Bunders, Arets, Frenken, and De Moor, “The Feasibility of Platform Cooperatives in the Gig Economy,” Journal of Co-operative Organization and Management, 2022}

• ^{Silicon Analysts, “NVIDIA AI GPU Market Share 2026”}

• ^{Platformonomics, “Follow the CAPEX 2025 Retrospective,” February 2}

They’re right. The progress is real.

Now look at who tops the list. Canada, Japan, the Netherlands, Norway, Switzerland, the United Kingdom. Functioning democracies, strong rule of law, serious regulatory institutions, full alignment with the most sophisticated AI regulation on earth. On every metric the CAIDP measures, they score well.

They would still fail a democratic governance test. And the Index, as built, has no way to detect that failure.

What Twelve Metrics Can See

The CAIDP’s metrics are individual-rights centred. Are people protected from harmful systems? Can they contest an automated decision? Is the government transparent about its AI use? Is data handled lawfully? These are the right questions to ask of a state. They are also, structurally, questions about protection rather than power.

There’s a difference between shielding someone from a system and giving them a say in how it’s built, owned, and governed. The CAIDP Index, the EU AI Act, the OECD Principles, the Council of Europe AI Treaty, the UNESCO Recommendation: every major framework measures the first. Not one measures the second.

Not one asks who owns the compute, the pipelines, the foundational models these systems actually run on. Not one asks whether a community has decision-making authority over the AI shaping its life, or only the right to complain once the decision is made. Not one asks whether a government could walk away from a vendor if the terms changed. Not one asks whether accountability runs to the people affected, or only to a regulator acting on their behalf. Not one asks whether any of this survives the current government, the current vendor, or the current political moment.

These questions mark the line between AI governance and AI democracy. The first produces systems accountable to institutions. The second produces systems accountable to people.

Losing the U.S. Isn’t the Fix

The CAIDP is right to flag the US withdrawal with alarm: Washington opposing the UN panel, an executive order unwinding safe-AI commitments, an exit from UNESCO. But the analysis needs one more step, because the thing being abandoned was never a democratic model to begin with.

The Blueprint for an AI Bill of Rights named affected people. It didn’t give them power. It addressed system behaviour, never system ownership. So losing US engagement is a real problem for global AI safety, and the alternative consolidating in its place, EU-style rights regulation, is more rigorous, more enforceable, and silent on exactly the same structural question.

The real choice underway is between two models of managed AI (one market-led, one rights-based) that both leave the foundational layer’s ownership untouched. Replacing American leadership with European leadership is progress on one axis. Building AI democracy runs on a different axis entirely.

Where the Blind Spot Comes From

There’s a reason every framework shares this gap: they all start from the same place, the liberal individual as the unit of analysis. Rights attach to persons. Harms are suffered by persons. Accountability runs to persons, through institutions built to represent them in aggregate.

That starting point comes from a specific philosophical tradition, one African philosophy has been challenging for decades. As political sociologist David M. Matsinhe puts it, “the Western AI ethicist asks how to protect the individual from the machine, while the Ubuntu philosopher asks what kind of community the machine makes possible.” Not the same question. The distance between them is where millions of people go missing from the frameworks meant to protect them.

Ubuntu starts somewhere else: what does this system do to the web of relationships that makes personhood possible in the first place? Kristy Claassen’s 2026 doctoral fieldwork with Ubuntu-practicing South African communities found something concrete underneath the philosophy: her participants prioritised dignity, care, and community well-being over the transparency-and-autonomy vocabulary that dominates global AI ethics, with women in particular foregrounding collective welfare in ways no standard metric captures. Her conclusion doesn’t leave much room to hedge: expanding the list of values isn’t enough. Privacy, responsibility, dignity: none of it means the same thing everywhere. Meaning is shaped by history, culture, and power, and a metric built for one won’t see the others.

Governance produces systems accountable to institutions. Democracy produces systems accountable to people. Everything else is downstream of that one distinction.

The Ones Who’d Pass and Still Fail

Take the Netherlands. Strong enforcement, meaningful transparency rules, an independent supervisory authority, a genuinely high scorer. Dutch citizens still can’t participate in a single decision about the AI infrastructure their public services run on. There’s no mechanism for community governance over the systems deciding housing, benefits, or employment access. The models underneath Dutch public administration are owned by companies under foreign jurisdiction, and there is no meaningful way to exit that arrangement if the terms change.

Canada ranks at the top of the Index. Serious governance discussion, real civil society participation. The systems running across Canadian immigration and social services sit on infrastructure Canada doesn’t own and couldn’t fully audit without vendor cooperation. Indigenous communities who’ve spent years building data sovereignty over their own lands and peoples have no enforceable claim on how that data feeds AI systems today.

South Korea passed a new AI law this year and shows up as one of the Index’s legislative wins. It handles risk categorisation and transparency well, leaving the question of who owns the stack untouched.

The pattern holds across every high performer: sophisticated protection for individuals, and total silence on who holds power over the systems doing the protecting.

Africa Names the Gap

This is where the structural problem stops being theoretical, because African policy documents say the quiet part out loud. The AU Continental AI Strategy calls for African-owned, people-centred AI. The Rabat Consensus calls for a sovereign path. The Africa Declaration, signed by 54 states in Kigali, insists AI reflect Africa’s own priorities and cultural contexts. These documents name the ownership question directly, in language no EU or OECD text comes close to.

Naming the ownership question is one step; closing it is a much larger one. The AfCFTA Digital Trade Protocol, the continent’s most binding digital governance instrument, prohibits source code disclosure as a condition of market access, which means African governments can’t audit the foreign systems they’re required to accept. The Malabo Convention, the only binding continental data protection instrument, has been ratified by 15 of 55 member states. The Continental AI Strategy has no open-weight requirement, no technology transfer condition, no exit right.

Rwanda, Kenya, South Africa, the continent’s highest scorers on any governance index, are building their AI futures on infrastructure they don’t own, protected by frameworks that shield individuals without handing communities any power. They’d score well on the CAIDP metrics. They’d fail the democratic test. What that gap measures is the distance between the metrics we use and the outcomes we say we want, quite apart from their ambition.

The Models Already Exist

It’s worth being precise here: “community ownership” sounds like it requires inventing something new. Every model below is already built and already working.

Data trusts let a trustee hold data rights on behalf of a community, with a fiduciary duty to act in its interest. Communities don’t own their data individually under this model; they govern collectively how it’s used, licensed, and shared, and any organisation wanting to train AI on it negotiates with the trust, not with thousands of individuals one at a time. The Ada Lovelace Institute and the Open Data Institute have already piloted the model in the UK; Barcelona has used it for urban data governance. It needs no new technology, only the will to deploy it at scale.

Platform cooperatives apply credit-union logic to digital infrastructure: worker- or user-owned platforms, governed democratically, where decisions about what a system optimises for, what data it collects, and how value gets distributed are made by the people the platform actually serves. Stocksy, Resonate, and Up&Go are functioning examples today. The open question is whether that same ownership logic can reach foundational infrastructure, not just the consumer-facing layer sitting on top of it.

Public-interest AI foundations hold infrastructure in trust for the public, governed by multi-stakeholder boards with community representation, the way universities and public broadcasters already are. Masakhane, the African NLP research collective, already operates something close to this model informally, building language models for African languages outside commercial dependency. What it lacks is the institutional structure and public financing to run at the scale this moment requires.

Municipal and regional AI utilities would treat compute the way functioning states treat water or electricity: a public utility, owned collectively, run for public benefit, with democratic oversight of pricing and access. Estonia’s X-Road is the closest working model, state-owned, interoperable, not extractive. The East African Community could, in principle, build a shared AI compute utility owned collectively by its member states, open to any public institution without dependence on foreign vendors.

Indigenous data sovereignty frameworks are the most developed of all of these. The CARE Principles (Collective Authority, Responsibility, and Ethics) give communities veto power over how their data enters AI training pipelines. Te Mana Raraunga in New Zealand and the

Global Indigenous Data Alliance have built governance architecture that’s legally recognised and operationally functional today, adaptable well beyond where it currently operates.

None of this asks anyone to invent a new institution. Every model above has legal precedent and a functioning example, even if only at smaller scale. The frameworks currently governing AI, the CAIDP Index among them, simply don’t require or reward any of it. Feasibility was never the obstacle. A procurement rule favouring cooperative ownership over extractive alternatives, a regulator willing to recognise a data trust as a legitimate governance structure, a development bank willing to fund community infrastructure as a public good: any one of these would shift the incentives. None of them require a revolution. They require policy choices the field hasn’t made yet.

What the Second Instrument Would Ask

A democratic governance tier, sitting alongside the CAIDP’s twelve metrics, would ask questions like these:

On ownership and infrastructure: What share of AI systems used in public services runs on domestically owned or collectively governed infrastructure? Do governments have the technical capacity to operate critical AI systems independently of private vendors?

On community power: Do affected communities hold decision-making authority over the AI systems shaping their lives, or only advisory status? Are there governance mechanisms operating below the level of the state that carry real power?

On exit rights: Can governments, institutions, or communities migrate away from an AI system or vendor if the terms change, without losing essential services? Are technology transfer requirements built into public procurement?

On accountability: Do accountability mechanisms run directly to the communities affected, not only to a regulator acting on their behalf? Do community-level redress mechanisms produce enforceable outcomes?

On data sovereignty: Do communities control how their own data is used in AI training and deployment? Are data trusts, or an equivalent mechanism, actually in place?

On durability: Would the governance framework survive a change of government, a vendor’s exit, or a geopolitical shift in whichever country controls the underlying infrastructure?

None of these questions appear in the CAIDP Index. None appear in the EU AI Act, the OECD Principles, or the Council of Europe Treaty. They appear, partially and rhetorically, in the African regional declarations, and nowhere yet as an enforceable requirement with teeth.

Rights First, Power Second

The strongest objection to all of this doesn’t come from defenders of the status quo. It comes from inside the rights tradition itself: individual-rights protection is what makes community power possible to exercise safely in the first place. A community-owned AI system with no enforceable transparency or contestability is just as capturable as a corporate one, arguably more so, since there’s no external regulator watching for it. The CAIDP’s transparency and accountability metrics are doing exactly this work: building the floor a community needs before ownership can be exercised safely at all.

That objection has a second half, sharper than the first: ownership can be captured too. A national government claiming AI sovereignty can relocate the capture point from a Silicon Valley boardroom to a domestic ministry, without adding a single vote for the communities it claims to represent. Several of the African frameworks cited above make exactly that move, asserting state control and calling it community power. State control and community power are not automatically the same claim, and treating them as interchangeable is where sovereignty language gets soft.

Both points hold, and both point toward tightening this argument rather than abandoning it. A floor of enforceable rights is necessary. On its own, that floor has never produced a community with a genuine say over the infrastructure governing its life: every jurisdiction in this piece already clears the floor and still holds no ownership stake. The objection correctly names a sequencing point, rights protections belong first, and a real design requirement, ownership structures need their own accountability mechanisms rather than an assumption that local control beats distant control by default. What the floor does not justify is measuring only the floor, forever, once the floor is already in place.

Don’t Wait for the Choices to Lock In

The CAIDP Index is the most serious instrument the field has built, and the people behind it are doing essential work. What it can measure, it measures well.

Its starting assumption is that the person, protected, is the endpoint of good governance. A person with no power over the system protecting them is still, in the way that matters, ungoverned.

AI is still in the window where these choices are being made, not yet inherited. Ownership, governance architecture, who gets a vote and who gets a complaints form: none of it is locked in yet. Once the vendor relationships are load-bearing and the regulatory frameworks are set, reversing them will take as long as it’s ever taken to unwind the ownership structure of any utility that came before.

The Index tells us how far the world has come on protecting people from AI. It was never built to tell us whether people have any power over it. Both numbers matter. Right now, only one of them is being counted.

What did we learn?

In this special season reflection, we explore the themes that kept resurfacing across every conversation. Human dignity. Power. Community. Governance. Responsibility. And the surprising realization that the most important questions about AI are rarely about technology.

From Ubuntu philosophy and psychological safety to human rights, humanitarian action, and the future of democracy, we reflect on what happens when we stop asking what AI can do and start asking what kind of world we want to build.

Because in the end, the machine may be the least interesting part of the story.

The future of AI is not being decided by technology alone. It is being shaped by the values, choices, and relationships between us.

In this episode, Dr. Cornelia C. Walther joins Juliet MacDowell and Margot Van Brakel to explore why one of the world's oldest institutions is engaging one of humanity's newest technologies. Together they discuss ProSocial AI, hybrid intelligence, human dignity, moral leadership, wisdom, governance, and the future of human flourishing in an AI-shaped world.

Dr. Walther is the author of Artificial Intelligence for Inspired Action and creator of the Prosocial AI Index, a framework designed to evaluate whether AI systems contribute to people, purpose, profit, and planet.

Topics include:

• The Vatican's intervention in the AI debate
• AI and human dignity
• Intelligence versus wisdom
• The Prosocial AI Index
• Hybrid intelligence
• Moral leadership in an AI age
• Human flourishing and the future of society
• Designing AI that serves people and communities

David M. Matsinhe, political sociologist, international law scholar, and human rights strategist, joins us to ask the question global AI governance has avoided: will institutions ever permit Ubuntu to do what it would actually do if taken seriously? Not enrich the existing framework. Replace its foundation.

Because a person is a person through other persons. And until AI is built around that truth, it is built around a fiction.

#Ubuntu #AI #AIGovernance #AfricanPhilosophy #HumanRights #DigitalEthics #ResponsibleAI #ProSocialAI #GlobalSouth #BetweenUsAndTheMachine

Following recent warnings from the WHO and growing reports of AI-induced psychological harm, Juliet MacDowell and Margot van Brakel speak with Brittnee Savage, founder of the BAI Group and creator of the CASE framework: Context, Attunement, Safeguards, and Escalation.

Brittnee explains why most AI failures are not model failures, but governance failures. The conversation explores AI as infrastructure rather than just a tool, the need for governance architecture before deployment, and why psychological safety must become a normative requirement in AI systems.

Topics include:

the “Digital Therapist Test” for AI systems

escalation design and governance failures

human-attuned AI governance

AI in healthcare, taxation, and public systems

the difference between “human-centered” and “human-attuned”

why regulation, education, and public participation must happen simultaneously

the role of ordinary citizens, schools, and institutions in shaping AI governance

The episode also traces Brittnee’s path from AWS and biopharma into governance work focused on human flourishing, privacy, and safety.

A grounded discussion on AI governance, responsibility, and the social infrastructure needed for the next phase of intelligent systems. 

Anisa Abeytia has spent over a decade in humanitarian crisis zones working on policy and now advises the UN on AI governance. She joins us to explain why "model drift" isn't just a technical glitch—it's narrative volatility that can shift who gets blamed for an airstrike depending on whether you ask in English or Amharic.

We cover:

• Why AI's "scaling" architecture is designed to fail at human complexity

• The three field-tested strategies from humanitarian diplomacy that AI governance ignores

• Her work building "Huma"—an alternative AI trained on paid community conversations, not scraped internet data

• Why democracy requires resisting AI-driven apathy

Writer and organizational thinker Nico Decock joins us to challenge the premise that organizations just need to optimize their existing systems. Instead, he argues that AI is forcing us to confront a deeper truth: most institutions weren't built for the complexity, speed, or nature of work today.

We get into why HR doesn't need a new model—it needs a new engine; the critical difference between automation and genuine transformation; and why institutions keep rewarding the people who preserve broken systems rather than those capable of reimagining them. Plus: what actually changes for ordinary employees when an organization gets this transition right, and the one move you can make now that shifts more than people expect.

If you've ever suspected that "digital transformation" was just old logic with new software, this conversation is for you

In this episode, we sit down with İlke Soysal—legal engineer and researcher—who has a name for what's at stake: attention integrity. She makes the case that our capacity for sustained, self-directed focus isn't just a personal discipline problem. It's a rights issue.

We explore how international human rights frameworks might protect cognitive autonomy, why the precautionary principle matters for digital policy, and what AI-powered persuasion means for the future of human agency.

If you've ever picked up your phone for one thing and lost an hour, or wondered what constant stimulation is doing to developing minds—this conversation is for you.