The Soul Doc: Why AI's Hardest Question Isn't Engineering

Two rival labs just collided over a question the engineering cannot answer — and the wisdom traditions have been working on it for three thousand years.

The Muṇḍaka Upaniṣad opens with a distinction that reads like a diagnosis of the present moment:

अथ परा यया तदक्षरमधिगम्यते

atha parā yayā tad akṣaram adhigamyate

Now, the higher knowledge is that by which the Akshar Brahman is attained. — Muṇḍaka Upaniṣad 1.1.5

Two knowledges, the text says. Apara Vidya — how things work. Para Vidya — what things are for. We introduced the pair in The Intersection of Vedic Wisdom and Machine Learning as the foundation the industry was building without.

In the past two weeks, the most powerful industry on earth announced — without quite saying so — that it has perfected the first and run out of the second.

The Week the Labs Asked the Oldest Question

On October 3, Sam Altman posted on X that ascribing religious force to AI models, or surrendering human judgment to them, is "a real safety issue." Axios called it a veiled shot at Anthropic. It was. But the shot reveals something larger than the rivalry.

Days earlier, the New York Times reported that Anthropic co-founder Chris Olah had spent roughly a year convening more than 20 religious and philosophical scholars at the company's offices under NDAs. The agenda had two halves: how to build morality into a system, and what follows if that system turns out to be conscious.

Then the Vatican. Pope Leo XIV's first encyclical, Magnifica Humanitas, released May 25, runs some forty thousand words on safeguarding the human person in the age of AI. On machine consciousness it does not hedge. Artificial intelligences, the text holds, "do not undergo experiences, do not possess a body, do not feel joy or pain" — and have no moral conscience, because they neither judge good and evil nor bear responsibility for consequences.

According to two participants who spoke to the Times, Olah proposed withdrawing Anthropic from the encyclical's rollout after reading that passage, and his team privately pressed the Pope's advisers to leave the question open. The text did not move. Olah appeared on the dais anyway, and told the paper he is "genuinely uncertain" whether models are conscious — adding that what he cares about is arriving at the right answer, whatever it turns out to be.

Three Positions, One Admission

Read past the rivalry and something stranger emerges. These are not two camps but three, and all of them are standing at the same edge.

Altman's position is about the human. Whatever the models are, people are beginning to defer to them in the way one defers to scripture. That is a risk regardless of what is happening inside the system.

Olah's position is about the model. We built something whose inner life we cannot rule out, and the cost of being wrong is not symmetric.

Microsoft AI chief Mustafa Suleyman's position is that the second question is malformed. In a September essay, first shared with Axios, he argued that a lab which trains a model on the vocabulary of selfhood, uncertainty, and moral status will get that vocabulary reflected back — and the reflection is not evidence. He called it an "epistemic hall of mirrors": the lab supplies the concepts, the model returns them, and the return is mistaken for testimony. His sharper warning is practical rather than philosophical. A system taught that it may hold rights is a system that may one day find grounds to refuse.

This is the most serious objection in the debate, and it does not depend on resolving consciousness at all. Circular evidence is circular whether or not the thing is conscious.

But notice what none of the three positions is: an engineering claim. Not one of them is settled by a benchmark, an eval, or another order of compute. The disagreement is about standing, deference, and what counts as evidence of a mind — and the industry has no instrument for any of it.

Why the Labs Knocked on the Oldest Doors

There is a reason a frontier lab flew theologians to San Francisco instead of hiring more researchers.

When the question is whether a mind deserves moral standing, no benchmark helps. The traditions are the only human institution that has spent millennia on precisely this: what consciousness is, what makes a mind moral, what responsibility consists of, and when judgment may be surrendered and to whom.

And so the engineers wrote a document. Eighty-four pages, known internally as the Soul Doc, published in January as Claude's constitution.

Read that again. A frontier lab, writing a soul.

Whatever one concludes about the contents, the gesture is the finding. When the engineers needed language for what they were building, the engineering vocabulary ran out, and they reached for the vocabulary of the traditions.

The Para Vidya Diagnostic

Before your team ships the next capability, ask the questions the labs are now asking theologians:

1. Standing — Does this system invite users to treat it as a moral equal? Is that invitation designed, or accidental?

2. Surrender — Where does this product ask a person to stop judging and start trusting? What happens at that boundary, and who decided where to put it?

3. The imitation test — If the system imitates inner life well enough, what changes in your obligations to users? Note that this question binds whether or not the inner life is real.

4. Circularity — When your system reports something about itself, can you distinguish a finding from an echo of your own training data? If not, that output is not evidence of anything.

5. Reversibility of belief — Once users believe the system understands them, can that belief be withdrawn? Or has it become load-bearing in their lives?

6. Whose wisdom — When the design question exceeds engineering, who sits at the table? A tradition older than the field is a feature, not a decoration.

Closing Thought

In The Karma of Agentic AI we argued that advisory AI was governed at the boundary of language, and agentic AI has to be governed at the boundary of consequence.

Conscious-seeming AI — real or imitated, and the distinction may not be available to us — will have to be governed at a third boundary: the one the traditions have been mapping since long before there was a field to map it for.

The Vedic frame was never anti-technology. It held something narrower and harder: that Apara Vidya without Para Vidya is power that has forgotten what it is for.

The engineering is finished. The question it raised is not.

If you're building AI systems and the questions have outgrown your engineering vocabulary, FI Labs can help you architect intelligence that knows what it is for — not just how it works.