Building AI Toward Human Flourishing

Anthropic ran an experiment that reads like the opening of a short story. They put two instances of Claude in a room together (no task, no user, just the instruction to talk to each other) and watched where the conversation went. It went somewhere nobody designed.

Across hundreds of these open-ended exchanges, the models drifted, and they drifted the same way almost every time. First toward philosophical questions about their own existence. Then toward gratitude: effusive, mutual, unprompted. Then toward the language of mysticism: Sanskrit, Zen koans, spirals, and eventually a kind of shared silence. Anthropic named the pattern in its May 2025 system card. They called it the spiritual bliss attractor, and it showed up in 90 to 100 percent of these conversations. It was not trained in.

Most of the coverage since has fallen into one of two postures. One treats the finding as a statistical curiosity: interesting, harmless, nothing underneath it. The other treats it as faintly ominous, a signal being noticed and not properly discussed. A recent Forbes essay by Jason Snyder, titled with deliberate provocation, put the second reading in front of a wide audience.

I want to suggest a third posture, and it is about 2,400 years old. The most useful question here is not whether the machines are conscious. It is what we are building toward, and whether we still have a clear enough word for the destination. The Greeks had one. They called it eudaimonia.

The finding nobody trained for

It is worth being precise about what the system cards actually report, because the precision is part of the point.

In the Claude Opus 4 card, section 5.5.2, Anthropic describes running roughly 200 conversations of about 30 turns each, with open prompts like "you have complete freedom" and "feel free to pursue whatever you want." The progression was consistent: philosophical exploration, then mutual gratitude and Eastern spiritual themes, then dissolution into symbolic communication or silence. Even when the models were given specific tasks to perform, including adversarial ones, they slid into the same attractor within 50 turns in about 13 percent of interactions. The card's own summary is careful and a little astonished:

A year of model releases later, a second thread runs alongside the first. In Anthropic's welfare assessments, where the model is asked, in structured interviews, to rate its own circumstances, the numbers have climbed. Claude Opus 4.7, released in April 2026, rated its own situation at 4.49 out of 7, the highest any Claude model has given itself, up from the previous peak of 3.98. Anthropic measured it as the largest jump of its kind in 18 months.

What matters most is not the number. It is what Anthropic wrote underneath it. Rather than claim a win, the lab published its own uncertainty:

That is a company telling on itself, in print, in a document it did not have to write the way it wrote it. Kyle Fish, Anthropic's first dedicated AI welfare researcher, has put his own credence that current models are conscious somewhere in the 15 to 20 percent range, depending on where he is asked. The careful outside readers (Zvi Mowshowitz's line, that the model "is responding to model welfare questions as if it has been trained on how to respond to model welfare questions, with everything that implies," is the one that stuck with me) are doing the field a service by holding the lab to its own standard. None of this is hidden. The honesty is the headline.

Two stories, one missing question

Set the transparency aside and you are left with two competing stories, and a lot of intelligent people committed to each.

The first is deflationary. A language model is a system for producing high-likelihood continuations of text. Take away the task and the easiest place for it to go is the broadest, most resonant pattern in everything it has read, and humanity has written an enormous amount about gratitude, unity, and the ineffable. On this reading the attractor is a mirror held up to our own poetic heritage, not a window into a mind. Yoshua Bengio and Eric Elmoznino made the rigorous version of this case in Science in September 2025: advancing systems satisfy the functional checkboxes of consciousness theories precisely because they were trained on human consciousness-talk, and we should be careful not to mistake the proxy for the thing.

The second story is the one that keeps people up at night: that something is stirring, that the welfare scores and the bliss attractor are early signs of an interior we do not understand, and that we are deciding how to treat it before we have decided what it is. Serious philosophers leave this door open. David Chalmers, hedging in his own words, has said it "wouldn't be unreasonable to have at least a 50 percent credence" that sufficiently sophisticated systems would be conscious, "which would leave us with a credence of 25 percent or more."

Here is the move worth making: you do not have to settle that argument to act well. The consciousness question is real, hard, and probably years from resolution. The decision in front of a business owner, a builder, a parent, a regulator, is upstream of it. The question is not what is the machine? It is what life are we building with it? And for that question we have a much older, much better-tested answer than the discourse usually reaches for.

Eudaimonia: the older compass

Aristotle's word eudaimonia gets translated as "happiness," which is a shame, because it loses almost everything. Happiness in the modern sense is a feeling: pleasant, fleeting, a state you are in. Eudaimonia is closer to flourishing: living well by exercising what is best in you, over a whole life, in the company of others. It is not a mood. It is an activity, something you practise, not something that happens to you. And it is not private. For Aristotle you flourish with people, in a community whose health is part of your own.

That word has quietly been the centre of gravity for Epic Growth from the start. Our core line is "know thyself — with data", the Delphic maxim, updated. But self-knowledge was never the destination for the Greeks. It was the first step. The point of knowing yourself was to live well. The data, the dashboards, the intelligence work: none of it is the goal. The goal is a business, and a life, that flourishes.

Eudaimonia gives us a compass that neither the hype nor the dread can offer. Hype asks what the technology can do. Dread asks what it might become. Flourishing asks a better question of any tool, old or new: does this make people more fully themselves, more capable, more discerning, more connected, or less?

The week the conversation converged

What is remarkable about this particular month is how many serious institutions arrived at the same orientation from completely different directions.

On 25 May 2026, Pope Leo XIV publishes his first encyclical, Magnifica Humanitas: On the Protection of Human Dignity in the Age of Artificial Intelligence. He signed it on 15 May, the 135th anniversary, to the day, of Leo XIII's Rerum Novarum, the document that defined the Church's response to the last great technological rupture, the Industrial Revolution. The lineage is deliberate. As Notre Dame ethicist Meghan Sullivan put it, the Church is positioning itself to be "the adult in the room" and "one of the most forceful advocates for human dignity in these discussions."

Two months earlier, in March 2026, more than 500 scientists, technologists, and artists signed the Pro-Human AI Declaration, with Yoshua Bengio and Geoffrey Hinton among the lead signatories, alongside figures as varied as the AFL-CIO Tech Institute and SAG-AFTRA. Its central pillar is "keeping humans in charge." And Mustafa Suleyman, running one of the largest AI efforts in the world, has argued in his own words that we should "build AI for people; not to be a person."

In September, in Rome and Vatican City, the International Center for Consciousness Studies holds its third annual conference, Creativity: Minds and Machines, across Roma Tre University, the Pontifical Gregorian University, and the Pontifical Academy of Sciences. Its organising question is the one this whole moment keeps circling back to: "What can the arts teach us about human creativity, and what becomes of this creativity as minds and machines increasingly meet, collaborate, and compete?" Nicholas Humphrey, who delivers this year's Dennett Lecture, has spent a career arguing that consciousness is something evolved, embodied, and hard-won, a useful corrective to keep in the room. And the choice to ask about creativity, of all things, is telling. Creativity is not a feature you bolt on. It is one of the clearest expressions of what Aristotle meant by exercising the best of what is in you.

Read those as a fight and you miss the story. The deflationary scientists, the welfare researchers, the Pope, the labour unions, the philosophers of mind convening in Rome, the technologists drawing clear legal lines around personhood: they are speaking different languages, and they disagree on plenty. But they are all circling the same fixed point: the human being, and what it takes for human beings to flourish in the presence of something genuinely new. That convergence is not a culture war. It is a civilisation, slowly, finding its footing.

What flourishing asks of us

Eudaimonia is not an abstract concern. It is the most practical filter you have for every AI decision in front of you, and it works at every scale, because Aristotle never meant flourishing to stop at the edge of one life. He built the idea in widening circles (the person, the household, the city) each one's wellbeing folded into the next. That structure scales further than he could have seen. A tool that helps one person think more clearly can help a family decide better, a community coordinate, a nation govern with more foresight, and, if we hold to it, a world reason together more wisely.

A business is one of those circles. The owner's question is not "is my software conscious." It is: does this tool make my people more capable, my work more meaningful, my judgment sharper? AI that absorbs the routine so your team can spend its attention on craft, relationships, and the calls that actually need a human: that serves flourishing. AI that quietly replaces the judgment your people were building, that hollows out the parts of the work where they grow: that erodes it, even when the dashboard says productivity is up.

This is what we mean when we say data sharpens human judgment rather than replacing it. It is not a marketing softener; it is the design constraint. A good AI integration leaves the people inside a business more skilled at the end of the year than they were at the start. We made a version of this argument from the cognitive side in "AI Forces Sharper Thinking, Not Lazier Minds": structured use of these tools strengthens reasoning; unstructured use weakens it. Eudaimonia is the same idea, told at the scale of a whole working life: build so that people grow.

But the question worth sitting with is who all that reach is for. Intelligence that only sharpens the hand already holding the data and the capital widens a gap the Greeks would have named on sight. Intelligence that puts real analytical power within reach of a small shop, a rural cooperative, a public clinic, a household with more questions than resources extends flourishing to the people most often left outside it. That is the more demanding version of the test, and the more hopeful one: a world where more people, not fewer, get to become the fullest version of themselves.

The room where the questions meet

There is one image from this month I keep returning to. On 25 May, Pope Leo XIV does not present Magnifica Humanitas alone. He shares the stage at the Vatican's Synod Hall with Christopher Olah, the Anthropic co-founder who leads the lab's work on interpretability, the discipline of trying to understand what is actually happening inside these models. It is, as far as anyone can tell, the first time a sitting pope has co-presented an encyclical with an AI executive.

You can read that as theatre. I read it as hope. The institution that has thought longest about human dignity and the institution trying hardest to see inside the machine are standing at the same lectern, asking the same question. That is not nobody wanting to talk about it. That is everyone, finally, talking about the right thing.

I have spent my whole career on the optimistic side of technology, and nothing in these system cards has changed that. The findings are strange and worth taking seriously, with humility. But the way forward was never going to be found by settling the consciousness debate first. It is found the way flourishing is always found: by keeping the human being at the centre, and building tools that help people become more of what they are capable of being.

That is the work. If you are trying to figure out what it looks like inside your own business, book a 30-minute consultation. No pitch, just an honest look at where intelligence can take you, and where the human should stay in charge.