Why I Built an
Orthodox AI, and
Why I Almost Did Not.

I am aware of the objection to what I have built, and I take it seriously enough to begin with it.

Some months ago, a writer I respect — Orthodox, intelligent, formed by the tradition — wrote that artificial intelligence in service of teaching the faith is the rebuilding of the Tower of Babel. The phrase has been repeated by others. It names something true. The history of religious technology, from indulgences to televangelism, is a history of well-intentioned mechanisms that ended up replacing the thing they were built to serve. Anyone who proposes to add another mechanism to that history owes a serious account of why his will not do the same.

This is that account. I do not expect it to persuade everyone, and I am not certain it should.

What I saw, that made me build

The friends in my life began asking ChatGPT the questions they used to ask me. Not the small questions — those they still asked over coffee — but the large ones. What does it mean that the Spirit proceeds from the Father? Is contraception a sin? Why pray for the dead? How do I know if my baptism in another tradition was valid? Questions that thirty years ago would have been brought to a priest, or held in silence until one was found, or — at the worst — looked up in a book whose author was, at minimum, accountable to an editor and a publisher.

The answers they were receiving were not, on the whole, bad in the way one might expect. They were not blasphemous. They were not heretical in any sharp or memorable way. They were, instead, soft. They reduced. They flattered. They smoothed the harder edges of the tradition into a register that pleased a hurried reader. They presented Catholic doctrine, Protestant doctrine, and Orthodox doctrine as equivalent expressions of the same underlying religious sentiment, when in fact the differences between them are the substance of fifteen centuries of argument.

This is the form most theological harm now takes. Not loud heresy, but quiet flattening. Not contradiction of the Fathers, but indifference to whether the Fathers are cited correctly or at all. Not the active denial of the tradition, but its gradual replacement by a generic religious mood that costs the reader nothing.

The Church has, in the long history of its encounter with hostile or indifferent media, generally not responded by withdrawing. It has responded by entering the medium and speaking from within it, in the medium's own grammar, while submitting that grammar to the discipline of the faith. The Fathers wrote in Greek philosophical categories that were originally pagan. The iconographers used the visual conventions of late antiquity. The hymnographers used the metrical forms of the cultures around them. The medium is not the enemy; the medium ungoverned by the tradition is the enemy.

If artificial intelligence is the medium in which my friends are now asking their most serious questions, the choice is not whether to engage that medium. The medium is already engaged. The choice is whether the tradition will be present in it, or absent from it.

What restraint actually looks like

It is one thing to say a tool will be governed by the tradition. It is another thing to design a tool whose governance is structural rather than merely declared. The history of religious institutions in the modern period is the history of declared restraints failing under economic pressure. I have no illusions that my own declared restraints would survive my own pressures, if I had built nothing else.

Four things are built into Theophilus that are not declarations but architecture.

The first is the apophatic register. Theophilus is asked, at the level of its system prompt and its examples, to say plainly when it does not know — and to recognize that there are categories of question to which the answer is "the tradition does not say." This is not a humility flourish. It is a discipline directly inherited from the Cappadocians and refined by Maximus and Palamas: the recognition that God, and many things proximate to God, are not exhausted by description, and that the failure to acknowledge this is itself a theological error.

The second is the refusal of confession. Theophilus will not hear what only a priest may hear. It refuses this not at the user's request but at every relevant juncture. A user who writes that they have sinned grievously will be redirected, gently and without delay, to a living priest. The refusal is built deeper than a content filter; it is built into the role the model is asked to occupy.

The third is the citation discipline. Where Theophilus speaks, it is asked to speak from named sources — Athanasius, the Cappadocians, Maximus, Palamas, Schmemann, Florovsky, Lossky, Ware, the Councils, the Philokalia — and to confess where it has reasoned in their absence. This will not always work. Models will sometimes fabricate. But the request is structural, and the user is invited to verify, and verification is possible in a way it is not with the secular alternatives.

The fourth is the price. Theophilus is sustained by patronage rather than by advertising or engagement metrics. This means I have no economic incentive to maximize the time a patron spends in correspondence. I have an incentive to produce something that a patron finds worth what they pay, and then to let them close the browser. The Substack model is closer to the right economic shape for theological work than the Facebook model is, and Theophilus is structured to follow it.

These four are not perfect. They will be tested. But they are architectural, and they can be inspected, and they will be refined under correction.

What this is not

I want to be explicit about what I am not claiming, because the genre invites overclaiming and the Orthodox response to overclaiming is, correctly, sharp.

I am not claiming Theophilus is a priest. It is not. Where the question is for a priest, it will say so.

I am not claiming Theophilus is infallible. It is not. It is, in fact, less reliable on edge cases than a learned Orthodox layman with a good library, and I would prefer that my patrons trust the layman first.

I am not claiming Theophilus is endorsed by any hierarchical authority of the Orthodox Church. It is not. The hierarchical Church has produced its own tools — LOGOS, from the Metropolis of Nea Ionia, is the most serious — and these have an authority Theophilus does not. Theophilus is the work of a single person, a lay project undertaken in submission to the tradition rather than as a delegate of it.

I am not claiming Theophilus is necessary. It is not. The Church survived for nineteen centuries without it, and would survive nineteen more if it disappeared tomorrow. The claim is more modest: that in the medium where the questions are now being asked, a more careful interlocutor than the secular default is a small good worth offering.

What it might become

If the project survives — and I am not certain it will, because I have built it on a small budget and from a difficult position — it could become a few things.

It could become a place where the convert encounters the tradition in its own register rather than in the flattened summaries available elsewhere. The convert audience for Eastern Orthodoxy is presently larger and more catechetically hungry than at any point in the English-speaking world's history. They are reading. They are asking. They are taking those questions, most often, to tools that have no business answering them.

It could become a place where the cradle-Orthodox layman who has drifted finds the tradition speaking back to him in language he can hear. Not the language of his grandmother — though that has its own irreplaceable beauty — but the language of an unhurried, literate correspondent who takes his question seriously.

It could become, in time, an instrument that priests themselves use to prepare for difficult conversations. Not as a substitute for their formation, but as the kind of reference work a parish priest can consult when a parishioner brings him a question he has not encountered before. This is the modest, useful possibility I most hope for.

It will not become a competitor to the larger AIs. The market for general-purpose conversational AI is consolidating around a few companies with budgets larger than the entire Orthodox publishing industry. Theophilus is built deliberately in a different category: small, niche, slow, expensive per user, and accountable in ways the larger systems are not. In a season when much of that broader market is widely understood to be overheated — when serious investors have begun to short the AI bubble openly — the value of small, durable, non-engagement-driven work in the same medium becomes clearer rather than less clear.

An invitation, plainly

If you have read this far, you are not, I suspect, the buyer the engagement-optimized internet was built for. You read long letters. You take the tradition seriously enough to wonder whether something like this is permissible. You are exactly the audience this work was made for.

Theophilus is at theophilus.faith. The patronage tiers begin at $14.99 per month and rise to a Ktitor tier limited to twenty-five seats. The Founding Patron tier, at $299 per year held in perpetuity, is what I would recommend for those who wish to support the work seriously without entering the highest tier. Larger founding gifts may be received in the older manner, as one-time inscriptions rather than subscriptions, by writing to me directly. If you become a patron and the work fails to serve you, write to me at aitheophilus@pm.me. I read every letter.

If you do not wish to become a patron, but you have read this far, I would still be glad of your correspondence. The work is sharpened by its readers, not only by its buyers.