How the AI content was created

More than a chatbot

If you have met AI mainly through the news, you have probably met it as a chatbot: something you type questions at, which answers fluently and is sometimes wrong. That picture is true as far as it goes, but it misses what these systems have become. Given the right tools, a modern AI model can work: it can search the web, open and read documents, compare one source against another, keep notes, follow a rulebook, and write up what it found — hour after hour, without a person steering every step. The technical world calls this an agent; a fairer everyday name might be an automated researcher.

That matters for a project like this one, because the Church's memory of her saints is enormous and scattered: monthly service books, mediaeval collections, the calendars of half a dozen national churches, modern parish websites — often disagreeing with each other about dates, names and identities. Organising that into one consistent, structured, searchable reference is precisely the kind of work an automated researcher is suited to, and precisely the kind no one person could do alone. Eikon is, in part, an experiment in whether it can be done faithfully.

The scale of it

The work began on the morning of 17 July 2026. The last of the 365 days of the church year was finished on the evening of 26 July, and the movable cycle — Pascha and everything that swings with it — two hours after that. Ten days, and on six of them work was actually done.

What was built in that time: nearly 1,700 saints and feasts, covering every day of the calendar, with a full biography for each, hymns for most of them, and over 950 icon images — each image with its copyright history and its identification traced and recorded. Just over three million words, which is roughly four times the length of the Bible.

For one person working alone this is not a large task but an impossible one. At an honest scholarly pace — reading the sources, weighing them, writing a careful life — a few entries a week would be good going; seventeen hundred is years of full-time work, and a working hagiographer might not manage it in a career. No person could check it line by line afterwards either: I have not read all 1,700 entries, and at ordinary reading speed that alone is months of work. This is the scale at which automated research changes what one person can undertake at all. So the question isn't whether I trusted the machine. It's what I put in its way.

The tool: Fable 5

A language model is a computer program that has read an enormous amount of human writing and can draft new text on request. It is not a mind and it does not know things the way a person does; it is a remarkably capable research and writing assistant that works from the sources and instructions it is given, and it can make mistakes.

The model that did the work here is Claude Fable 5, released by Anthropic in June 2026 — the first of their fifth-generation models, in a tier Anthropic calls Mythos-class, which sits above their earlier Opus models. At the time of writing it is the most capable model Anthropic makes generally available, and it ships with additional safety measures for dual-use capabilities; a variant without those measures is restricted to approved organisations. Anthropic's own announcement describes the family.

Its power has not gone unremarked: days after its release, the United States government ordered access to it suspended under export-control law while a security concern was investigated, and Anthropic took it offline worldwide in response. The order was lifted at the end of that month and Fable returned to service with strengthened safeguards. I mention this not for drama, but because being open about the tools means being open about their history too.

The vintage of the model matters, so it is worth saying what changed. The language models available up to mid-2025 were already impressive writers, but they were assistants rather than researchers: they could draft a page from material they were handed, they lost the thread over long work, and left unsupervised they invented things freely. A project on this scale — millions of words of sourced reference prose, researched across the open web against a fixed order of authorities, kept consistent across seventeen hundred subjects, in ten days — was not a possibility with them at any price of effort.

It became a possibility with the generation Fable belongs to: models that can carry a long research task from start to finish, consult and weigh sources as they write, follow a rulebook faithfully across thousands of pages, and — crucially — record their own uncertainty instead of papering over it. That power cuts both ways, and it is the reason the rules described below exist: a more capable machine makes rarer mistakes, but it makes them in better prose, and they are correspondingly harder to catch.

Who wrote what you'll read

Every saint's life and feast article on this site was researched and written by the AI, working with access to the web. I did not write the biographies. I set the rules, fixed the order of sources, ruled on the questions the machine could not settle, and reviewed its work. It did the reading and the writing.

I am saying that first because it is the thing most likely to make you distrust what you are reading, and you would be right to start there. Language models invent things — plausible names, plausible dates, plausible feast days — and they do it in the same confident tone they use when they are correct.

Just as important is what AI did not produce. The calendar, the feasts, the scripture readings, the fasting rules — the devotional heart of this site — are not AI-generated. They follow the Church's received lectionary, unchanged.

The biographies are published under ongoing editorial review: a human editor reads, corrects and improves them continually, rather than every page having been checked before it first appeared. I would rather be honest about that than pretend otherwise — and it means corrections from readers are genuinely welcome.

The human work: designing the machine's job

None of this ran itself. Before the machine wrote a word, two jobs had to be done by a person, and they are where most of my own time went.

The first was writing the rulebook — a written constitution the machine had to follow. It fixes which authorities count as sources and in what order they must be consulted; what may be quoted and what copyright forbids; how a saint's name and spelling are chosen when traditions differ; a fixed vocabulary for classifying saints, so a "martyr" or a "fool for Christ" means the same thing on every page; the search that must be done before any new saint is created, so the same person is never written up twice under two names; what to do when sources disagree; and which decisions the machine was forbidden to make alone and had to bring back to me. Every rule described further down this page began as a sentence in that document — and when the work surfaced a case the rules didn't cover, the rule was written first and the work redone under it, not the other way round.

The second was designing how the work could be automated at all. No single AI session can hold a year of the Church's calendar in its head, so the task had to be broken into pieces a machine can do well: the calendar was divided day by day, and a separate AI agent was assigned to each day's saints — many of them working side by side, each confined to its own day so that no two could ever write over each other's work. Above them sat a coordinating layer that merged their results, kept the shared indexes consistent, ran the automatic checks, and recorded exactly where the work stood so it could stop at any moment and resume without loss. Deciding those boundaries — what each agent may touch, what it must hand back, how its output is checked before it becomes part of the whole — is design work no model does for you. It is the difference between an assistant you supervise sentence by sentence and a workforce you govern by rules.

What I put in its way

A fixed order of sources. Every life had to be built from named authorities, consulted in a set order: the Antiochian Archdiocese first, then the other jurisdictions (the Orthodox Church in America, the Greek and Russian churches), then the classical collections the Church herself compiled — the mediaeval Synaxarion of Constantinople, the Greek service books, the saints'-lives collections of Nikodemos the Hagiorite. Where a saint left writings, or a contemporary wrote about them, the machine was sent to the oldest witness we have: the entry for Ignatius of Antioch quotes his own seven letters; the fourth-century lives lean on Eusebius, Theodoret, Sozomen and Palladius, who knew some of these people. Wikipedia was allowed as a lead, never as the only support for a fact.

Every life closes with a numbered list of what it was actually built from — that is the reader's audit trail, and you will find it at the foot of each saint's page. Across the whole collection those lists cite the Menaion (the Church's monthly service books) about 1,500 times and the Synaxarion of Constantinople over 800 times, alongside the living calendars of the jurisdictions, checked on the day the life was written.

Nothing copied. All the prose is written fresh from the sources, not reproduced from them. Hymns are used only where they are out of copyright or freshly translated; where a modern service text is still in copyright, you'll find only its opening line and its tone, and no text.

Disagreement recorded, not resolved. The churches genuinely differ about dates and identities. Where they do, the life says so and names both usages rather than quietly picking one. Where the evidence simply ran out, the life says that too.

A duplicate check before any new saint. Saints share names relentlessly — four Barlaams, three Nestors on one day. Nothing new could be created without first searching for everyone it might already be.

An automatic checker that had to pass before every save. It catches broken cross-references, repeated identifiers, invalid classifications, calendar arithmetic errors and missing images. It is currently clean.

The best evidence the rules were needed

Four times the AI reported that an existing life had the wrong feast day. Three of those four reports were wrong — confidently, specifically wrong, with reasoning that read perfectly well. Because the rule was to check the claim against the source before acting on it, three correct entries survived that would otherwise have been quietly corrupted. A record is kept of every mistake the machine has made, alongside the open questions it could not settle.

What is not verified

• Readings. Fewer than one saint in ten has their Epistle and Gospel confirmed from a lectionary. The rest fall back to the standard readings for their type of saint, which is how the services actually work — but it is a fallback, not a finding.
• The Antiochian check. The tradition this is written from is the one I can least verify: the Archdiocese publishes its calendar one day at a time, so the day-by-day confirmation is complete for one day out of 365.
• Icons. Hundreds of saints still have no image, because I would rather show nothing than something I can't prove the rights to or the identity of.
• Coverage. Hundreds of minor and local saints are listed on their day but not yet written up.

How to read it

Take the source list at the foot of a saint's page as the actual claim. Where a life hedges, records two dates, or says a tradition is unverified, that is deliberate and it is the most trustworthy part of the page. Where it states something flatly, it is telling you the sources agreed. And if you find an error, tell me — corrections are part of the design, not an embarrassment to it.