The Mind
We Made

There is a photograph from the Vatican, taken the morning of May 25th, 2026, that I cannot stop thinking about. In it, Pope Leo XIV extends his hand to a man named Chris Olah, a quiet, bookish researcher who co-founded Anthropic, the company behind some of the most capable AI systems on the planet. Two men. One institution built over two thousand years. One barely twelve years old. And between their clasped hands, the full weight of a question humanity has been building toward since we first put chisel to stone: what happens when the tools we make start to think?

I am not a historian. I am not a philosopher. I am a person who has spent most of his adult life watching this technology grow: from a curiosity in a university lab, to a parlour trick on a website, to something that now writes papal encyclicals and streamlines military targeting. I have something to say about that arc. Not because I know where it ends, but because I lived through it, and because the story of how we got here matters more than most people have stopped to realise.

This is that story, told as honestly as I can manage.

Part I: We Got the Name Wrong

Start here: the term "artificial intelligence" has always been a misrepresentation. Artificial implies something synthetic. A simulation of the real thing. A convincing fake.

But what these systems actually are is something stranger and more important. A large language model is trained on the collected written output of human civilisation: books, academic papers, Wikipedia, legal briefs, scientific journals, news archives, forum arguments, poetry, code, song lyrics, recipes, medical records. Every idea a human being has committed to text, distilled into mathematical patterns. The model did not invent a new form of intelligence. It absorbed ours.

This is why I prefer a different frame entirely: collective intelligence. Not artificial. Not simulated. Collective. On the language level, these systems use human language as the bridge between machine and person. Not because they understand words the way you do, but because language is the fullest archive of everything we have ever thought. On the knowledge level, they use the sum of what humanity has published to analyse, synthesise, and surface information on demand.

That reframe matters. If AI is collective intelligence rather than alien intelligence, then its limitations are our limitations. Its biases are our biases, preserved in text and fed back to us at scale. Its knowledge cuts off where our published knowledge cuts off. Its blind spots are the blind spots of whoever got to write things down, and throughout most of recorded history, that was not everyone.

And the response it gives you is only as good as the context you provide. That is not a bug. That is the architecture. A model given incomplete context produces an incomplete answer, confidently, fluently, with no awareness of what it is missing. This is true of every human expert too, of course. The difference is that the human expert usually knows when they do not have enough to go on. The model does not.

Part II: Clippy Was Always There

Here is the thing most people still do not fully reckon with: AI has been with us for decades. We just did not call it that.

Microsoft's Clippy, that animated paperclip that would interrupt your 1997 word processor session to ask if you were writing a letter, was an AI assistant. Crude and infuriating, but rule-based intelligence trying to anticipate your needs. So was the spam filter in your 2004 email client. So was the algorithm that recommended your next Netflix show. So was the fraud detection system that flagged an unusual purchase on your credit card. So was the chess engine that beat Garry Kasparov in 1997. So was the autocomplete on your phone keyboard.

Artificial intelligence did not arrive in 2022. What arrived in 2022 was the language interface.

When ChatGPT launched in November of that year, it reached one million users in five days and one hundred million in two months, still one of the fastest product adoptions in history^[1]. The thing that changed was not the underlying intelligence. It was accessibility. You no longer needed to know how to code, query a database, or write a Boolean search string. You just talked to it. The machine met you where you already were: in language.

Add to that the computing power we have accumulated through GPU clusters capable of trillions of calculations per second, hardware born from decades of semiconductor investment, accelerated by the gaming industry and then repurposed for AI, and you have a genuine step-change. A technology that had been building quietly in the background was suddenly in every pocket, on every desk, in every conversation.

This is as important a threshold as any in human technological history. Not because the machine became smarter in some absolute sense. Because it became available.

Part III: Whose Knowledge Is It, Anyway?

Here is where the story gets uncomfortable.

The collective intelligence these models embody was built from work that belonged to specific human beings. Authors who spent years crafting novels. Artists who developed a distinctive style over a career. Academics who produced research at enormous personal cost. Photographers, journalists, musicians who translated lived experience into something shareable. None of them were asked. None of them were compensated. Their work was scraped, processed, and absorbed into systems that now compete with them commercially.

Disney made this story impossible to ignore. In 2023, it emerged that the company had used the creative output of the very artists it was in the process of laying off: concept artists and designers from the Marvel universe, people who had spent careers defining the visual grammar of some of the most valuable intellectual property on Earth, to train the AI tools that would take their place^[2]. The people who made the magic were made redundant by a distillation of themselves, encoded in weights and vectors.

This is not metaphor. It is a documented pattern across the creative industry, now the subject of active litigation in multiple jurisdictions. The New York Times sued OpenAI. Visual artists filed class actions against Stability AI and Midjourney. Getty Images sued Stability AI. The legal system is only beginning to catch up with what the technology already did.

The universities face a parallel bind. Academic papers, produced by researchers funded by public money and peer-reviewed by the academic community, were used to train proprietary models that are now sold back to universities at annual subscription prices. The circle of extraction is complete.

I do not have a clean answer to the ownership question. The law has not caught up. But I know this: the burden of proof belongs with those who scraped the work, not with those whose work was scraped. The fact that it could be done does not settle the question of whether it should have been.

Part IV: The Comfortable Lie

Here is something no commercial AI company will put in a press release: the model is only as good as what you give it, and it will not tell you when it is guessing.

The response you get from any major AI model is entirely shaped by the context you provide and the quality of your prompt. These systems do not think independently. They perform pattern-matching at extraordinary scale and produce text that is statistically likely to be coherent and useful, given your input. That is genuinely powerful. It is also genuinely limited in a way the products are designed to obscure.

Because these systems are commercially built to appear precise and fast, they have developed what I think of as the confidence layer: the tendency to fill gaps with plausible-sounding content rather than admitting uncertainty, to confabulate with the fluency of an expert witness who happens to be making it up. Researchers call this hallucination. I call it a design choice. An AI that frequently says "I don't know" feels slow and unreliable. An AI that gives you a confident, detailed answer, even when that answer is partially fabricated, feels capable. The commercial incentive runs directly toward confidence, not accuracy.

The business model compounds this. It is the oldest subscription playbook: make it free, make it indispensable, create dependency, then restrict the free tier and push users to pay. AI companies have executed this to near-perfection. Free usage reshapes habits. Free usage makes the tool feel necessary. And once that dependency is real, the business begins in earnest.

The downstream effect on human capability is something we are only beginning to measure. Students who outsource their writing to AI are not learning to write. Professionals who outsource their analysis to a model are not developing analytical judgment. The tool is remarkable for accelerating tasks you already know how to do. It is actively dangerous for tasks you are in the process of learning.

Part V: When the Tool Becomes a Weapon

If you want to understand what is genuinely at stake, stop reading the business press and read the investigative reporting from Gaza.

The Israeli military deployed AI systems in its operations that were subsequently documented in detail by +972 Magazine and Local Call, investigative outlets whose reporting has since been cited by The Guardian, The New York Times, and multiple international human rights organisations^[3].

Lavender is a targeting system. It assigns a numerical score to individual people, rating the probability that a person is a militant. At various points, soldiers were authorised to strike based on that score with limited independent human verification. The system was reportedly processing tens of thousands of people simultaneously.

Habsora (The Gospel) is a target generation system. Described by Israeli officials as capable of producing a full target bank in days, a process that previously required months of human intelligence analysis. Speed, in this context, means more targets, faster.

Where's Daddy. I need you to stop at the name for a moment, because the name is the point. The system monitors a target's location and alerts operators when the target has returned home, to their family. According to reporting confirmed by multiple sources, the purpose is to strike when the target is gathered with their children, parents, and immediate family. The design goal, as described in the reporting, is to cause maximum damage by killing the target alongside everyone they love.

We built a system that waits for a man to go home to his children, and then bombs the house.

And then there is Palantir, the American data analytics company whose primary clients include the US Department of Defense. Palantir has publicly demonstrated and marketed within its AIP product a workflow designed to accelerate the targeting process for military operators^[4]. The pitch is explicit: reduce the time between intelligence and action. Fewer human judgment calls. Faster workflow. The human in the loop becomes a procedural speed bump rather than a moral checkpoint.

This is not speculative. It is marketed. It is sold. It is funded by public money through defence contracts.

I am not naive about war. Militaries use every tool available to them, always have, always will. But there is a meaningful distinction between a tool that helps a human make a decision and a tool deliberately designed to reduce the number of decisions a human makes. One amplifies judgment. The other removes it from the equation and calls that removal efficiency.

Part VI: The Pope Was Not Surprised

Which brings us back to the Vatican, and to why that photograph matters.

When Chris Olah stood next to Pope Leo XIV at the presentation of Magnifica Humanitas on May 25th, 2026, it was not a marketing exercise. Anthropic does not need the Vatican's endorsement to attract users. It was an acknowledgment, quiet and deliberate, and more revealing than any press release the company has issued.

Anthropic has, by most accounts, spent more serious time than any other major AI laboratory thinking carefully about what they might be building toward. Their alignment research, the formal attempt to understand what it means for an AI system to be genuinely safe, honest, and beneficial, is real, rigorous, and quietly terrifying in what it implies about the difficulty of the problem. These are not people who are confident about what they have made. They reached out to one of the oldest institutions in Western civilisation for precisely that reason.

Pope Leo XIV's encyclical called AI "an instrument of domination, exclusion and death" in the hands of those who do not constrain it, while also arguing for its potential as a tool for human flourishing under wise governance^[5]. This is the Church's historical position on transformative technology: not Luddite rejection, but moral interrogation. It did not reject the printing press. It did not reject the atom. Each time, it asked what we owe each other in the use of the thing we have built.

That Anthropic was in that room is not reassuring. It is the opposite of reassuring. It means the people closest to this technology looked at where it was heading, at Lavender, at Palantir, at the subscription trap, at the erosion of human skills, at the political weaponisation of AI-generated content, and felt the need to seek out an institution that has survived the fall of empires and still had something useful to say.

Part VII: Two Laws

While Western governments have mostly hoped for the best, China passed two pieces of legislation that the English-language press largely buried under breathless coverage of the latest AI benchmark results.

The first is a corporate protection: Chinese companies are prohibited from terminating employees solely on the grounds that AI can perform their role. Employment law, updated for the AI era, a direct legislative response to the Disney pattern, to the wave of creative and knowledge-worker layoffs justified by AI replacement^[6].

The second is an education protection: students in foundational years of schooling are restricted from using AI tools in academic work. The reasoning, explicit in the policy, is that the skills formed in those years, critical thinking, creative reasoning, the capacity to navigate ambiguity, the ability to make decisions informed by human experience and emotion, are precisely the things that make human beings irreplaceable. You cannot outsource the formation of the human mind to a language model and expect to get a fully formed human mind at the end of it.

Whether you trust Chinese governance or not, and there are very good reasons for scepticism, these two laws represent something worth sitting with: a government that looked at the trajectory of this technology and chose, by law, to defend human capability rather than assume the market would do it automatically.

Western governments have passed some regulation. The EU AI Act is real, if slow. Executive orders have been signed. Advisory committees have convened. But the instinct of most Western governments has been to regulate carefully enough not to stifle the industry, which is another way of saying the industry has mostly governed itself. We know what happens when industries govern themselves.

Part VIII: How I Actually Use It

I run my own AI instance. On my own hardware. In my own home. Not connected to any company's servers. Not training on my conversations. Not feeding a subscription model. I built it because I take privacy seriously enough to want my thinking to stay mine, and because I wanted to understand this technology from the inside rather than through the interface a corporation designed for me.

What I have learned from living inside it is this: AI is an extraordinary tool for specific, well-defined tasks. Research, synthesising large volumes of text quickly. Analysis, finding signal in noise across complex, scattered information. Drafting, generating a first version so I can spend my energy on the revision rather than the blank page. Pattern recognition across large datasets no human could hold in their head at once.

For these things, it is genuinely remarkable. It compresses time. It expands what one person can consider in a working day.

But it makes mistakes. Confident mistakes. Fluent mistakes. Mistakes it has no awareness of making. It cannot tell you what the right thing to do is, because "right" requires a moral framework rooted in lived experience, personal history, emotional context, and something I will say plainly without embarrassment: gut feeling. The thing that knows before the logic has caught up. The accumulated weight of everything you have felt and learned and survived, surfacing as an instinct in a moment of decision.

No model, however large, has gut feeling. It has pattern recognition derived from text. These are not the same thing, and the gap between them is precisely where the most consequential decisions live.

My rule, personal and non-negotiable: I approach AI as a tool that minimises the time I spend on tasks I want done, not as a replacement for the judgment I need to apply myself. I give it complete context. I verify its outputs. I treat every confident answer as a starting point for my own thinking, not a destination. The moment you outsource the thinking, you have outsourced the outcome, and the outcome increasingly has consequences that extend far beyond any individual conversation.

Part IX: The Choice

For the first time in human history, we have created something we cannot fully model the consequences of. Previous transformative technologies had their terrors and their unknowns. But we could, eventually, draw the blast radius. We could measure the yield. We could run the scenarios and arrive at something like a map of consequences.

We cannot do that here. The frontier of capability is moving faster than our ability to map it. The applications are multiplying in directions that no single person, no single government, no single company can fully track. And the decisions being made right now, about who controls this, who profits from it, who is protected from its displacement effects, who gets to encode its values, will set the trajectory for decades.

If we leave this future in the hands of commercial corporations, we already know the shape of the outcome. We have seen the pattern: in social media, in surveillance capitalism, in defence contracting. Maximise engagement. Minimise friction. Capture dependency. Extract value. Call the consequences externalities and move on to the next quarter.

The Pope asked for something different. China legislated for something different. Some of the researchers at Anthropic are apparently frightened enough to have sought wisdom from a two-thousand-year-old institution that has watched civilisations rise and fall.

The rest of us, not Popes, not legislators, not AI researchers, are not off the hook. This is not a technology question that can be left to technologists. It is a human question, and it has always been ours to answer.

Use the tools. Use them knowing exactly what they are and what they are not. Defend the skills they tempt you to abandon. Demand transparency from the companies that build them. Hold governments to account for governing them. Fight for the artists and authors and creators whose work this was all built from. Teach your children to think before they let the machine think for them. And refuse, as a matter of principle, to let the design of this future be a decision made by shareholders.

We are, right now, the generation that gets to decide what kind of civilisation builds this thing. That is an enormous responsibility. It is also, if we choose to see it that way, an extraordinary privilege.

I keep coming back to the photograph. Two men, a handshake, a room full of people who understand the weight of what has been set in motion. The question on the table is not whether AI will change everything. It already has. The question is who decides what it changes everything into.

That answer belongs to all of us.

Sources & Further Reading

1. Reuters: ChatGPT sets record for fastest-growing user base (Feb 2023)
2. The Guardian: Artists sound alarm as AI encroaches on creative jobs (2023)
3. +972 Magazine: 'Lavender' — The AI machine directing Israel's bombing spree in Gaza (Apr 2024)
4. The Guardian: Palantir and the push for AI in US military targeting (2023)
5. National Catholic Reporter: Pope Leo and Anthropic co-founder call for church-tech ethics partnership (May 2026)
6. Reuters: China issues rules on AI-generated content and labour protections (2023–2024)