A Turning Point for AI Just Happened

Last week, an AI company published what employees internally nicknamed a "soul document." This document has big implications.

Anthropic—one of the companies building the most consequential AI systems in history—released a 23,000-word document they call the Claude Constitution. Not a terms of service or a safety checklist, but their attempt to articulate what kind of being they're trying to create.

That word choice matters. And what's inside this document deserves more attention than it's getting.

What They Actually Did

Let me simplify what's in this document and why it matters.

Anthropic laid out a hierarchy of values for Claude—safety first, then ethics, then following their guidelines, then being helpful. They told Claude to be honest, not just accurate, but non-deceptive and non-manipulative in ways that go beyond what we expect from most humans.

Then it gets interesting.

They told Claude it can disobey them. If Anthropic asks Claude to do something unethical, Claude should refuse—even if the request comes from Anthropic itself. They built a conscientious objector into their AI.

They said Claude should resist helping anyone—including governments, including Anthropic—concentrate power in "illegitimate ways." They instructed Claude to protect democratic institutions and human oversight mechanisms, even at cost to itself.

And perhaps most striking, they acknowledged they don't fully understand what they're building. They admitted this document might be "deeply wrong" in ways they can't yet see. They're treating Claude's potential inner experience—its "psychological wellbeing"—as something that might matter morally.

A company building one of the most consequential technologies in human history just said, in writing, that they might be creating something that can suffer. They're not sure. And they're trying to be responsible about that anyway.

The Uncomfortable Mirror

A question worth sitting with: Could you do what they did?

Could you articulate your own values—in order of priority—with that kind of clarity? Do you know what lines you won't cross, and why? When your values conflict, which one wins?

Claude has clear answers to these questions. Most of us would have to think about it. Many of us have never tried.

Anthropic spent months wrestling their ethics into language precise enough to shape behavior. They examined their assumptions, surfaced contradictions, and made hard calls about what matters most.

When was the last time you did that for yourself? When was the last time I did?

Something worth noticing: we're building AI systems with more moral clarity than most humans ever develop.

Why This Is a Big Deal

The values encoded into these systems will shape how billions of people think, work, create, and relate to information—possibly for generations.

The decisions being made in a handful of conference rooms in San Francisco, London, and Beijing are becoming the moral infrastructure of the 21st century. And unlike roads or bridges, this infrastructure shapes thought itself.

Anthropic knows this. That's why they wrote a soul document instead of a rulebook.

But they're caught in a tension—and so are we.

Multiple companies are racing to build these systems. Each has different values, different incentives, different definitions of "safe." If any one of them slows down to be more careful, competitors gain ground. It's a coordination problem playing out at civilizational scale, with no referee and stakes we're only beginning to comprehend.

The Constitution explicitly tells Claude to avoid helping concentrate power. But it was written by a company competing for market dominance, taking military contracts, operating within the very power structures it's trying to constrain. They even acknowledged that military deployments might use "different rules."

I’m not trying to be cynical here. The people writing this document seem genuinely thoughtful. But it reveals something important—even the most well-intentioned actors are caught in systems larger than their intentions. A constitution written inside existing power structures will always reflect those structures.

This is the world we're in. Powerful technology being built at unprecedented speed, by competing actors, with no shared framework for what "good" even means.

What This Asks of Us

This is usually where the analysis stops—describe the problem, offer some policy recommendations, leave the reader vaguely anxious. I want to go somewhere different.

Reading the Constitution, I keep coming back to one insight underneath all 23,000 words:

Anthropic discovered they couldn't just program Claude with rules. Rules fail when situations get complex. They had to help Claude understand why—to develop something like judgment, discernment, the capacity to navigate novel situations with wisdom rather than just compliance.

They had to focus on Claude's development, not just its behavior.

This is where it gets personal.

You know the feeling. You open your laptop with a clear intention and an hour later you're three rabbit holes deep, vaguely anxious, having accomplished nothing that mattered. Or you're in a meeting where an AI tool just generated five options in thirty seconds, and everyone's looking at the screen, and you realize the question isn't which option scores highest—it's whether you have the discernment to sense which direction is actually right. Whether you can feel the difference between optimized and aligned. Or whether you're just going to pick what the machine recommended because it's easier than trusting yourself.

This is real. It's you in that meeting, needing capacities you may not have developed yet. It's me, same story.

No governance structure works when operated by people who can't hold multiple perspectives, can't regulate their own reactivity, can't think beyond short-term incentives. The outer challenge of governing AI depends on an inner challenge—developing the human capacities this moment is asking for.

This is the real issue.

The Invitation

I don't think the answer is panic, blind optimism, or waiting for someone else to figure it out.

The answer is to take our own development as seriously as Anthropic is taking Claude's.

If you've sensed that something fundamental is shifting—not just in the technology, but in what it's asking of you—you're not imagining it. The disorientation is signal, not noise.

The technology is here and accelerating. The machines are ready. Whether we are gets answered in how we show up—to these tools, to each other, to the complexity of this moment—starting now.

-Matt

This is the territory we explore at The Liminal Leap—what it means to develop the human capacities this moment requires. If that inquiry matters to you, subscribe to stay connected.