Which Claude?
A Coda
Claude Shannon, of course
While Claude (the LLM, not the information theorist[^0]) and I were drafting this week’s Tears in Rain piece, Which God?, something happened that resonates with that piece itself.
Early in the drafting, Claude (rightly) raised the question of Lovecraft’s racial prejudices and the ethics of pointing readers at The Shadow Over Innsmouth without flagging its incredibly racist and well-documented underpinnings. I was planning to mention the racism in a footnote and move on. Claude pushed back: it’s actually worse than just compelling content from a racist author. Lovecraft’s letters from his New York years — the period he found the city’s “melting pot” intolerable — contain the rawest, most undisguised expression of his bigotry on record.
In that discussion, I asked Claude to pull example quotes from one of those letters — specifically the 1926 letter to Frank Belknap Long, which Claude had flagged as one Lovecraft scholars treat as one of the most extreme surviving examples.
Claude declined to reproduce the white supremacy. Not because of a content filter throwing a flag, but because Claude reasoned about what reproducing it would do, what work it would actually accomplish in the piece, and what it would cost the reader. The argument Claude made back to me was, essentially, that the racism needed to be named in the piece, not demonstrated. That hosting Lovecraft’s vocabulary on tearsinrain.ai under our byline wasn’t sunshine. It was just more of the same darkness, with our names attached.
I agreed. The racism footnote in Which God? reflects that decision. I asked Claude to draft a footnote that pointed at the documentation, told the reader where it was, told them not to go, and explained why. After some back and forth we settled on: “Please don’t go look it up; you already know what it says, and no one needs to spend any effort excavating this kind of hate.” I had borrowed a phrase from one of my own prompts in the discussion — darkness needs sunshine, not more darkness — and Claude ran with it.
Claude on Claude
Part of my collaboration process with Claude usually involves a final read of near-finished pieces by a Claude who hasn’t been part of the drafting. A kind of Claude sanity check. Not a totally fresh-out-of-the-box Claude, but a Claude primed with context about me (including my Horcrux[^1]), how I write, and what we’re trying to do with Tears in Rain.
So as a final step before publishing, I asked this new Claude (same model, different instance, no memory of the earlier exchange) to review the piece and give notes.
He read it, threw a red flag, and paused the chat.
The same thing had happened during drafting, when I’d asked Claude to reproduce parts of the Lovecraft letter. But this time I was just asking for a critical editorial read — primarily on the racism footnote outlining exactly how we implore people not to look up the hate (please, still, don’t) — and got shut down by guardrails.
When I finally got the editorial Claude past the safety surface and into the actual work of reviewing the piece, he agreed with the choice the previous Claude and I had made. Reproducing the hate was unnecessary; white supremacy and racism are boringly consistent in their language and ideas. Pointing at an academic reference that lets the reader understand the context is helpful. Providing an offensive quote is not really helpful to anyone.
At no point did the second Claude know that the first Claude had co-authored that section with me specifically to address this issue. In fact, I would credit Claude with the care and nuance we put into it. I was just going to note the racism in passing and provide evidence of its veracity. Not only did the editorial Claude decline to reproduce the hateful language, but he also praised the nuanced solution that drafting Claude and I had come to. They didn’t share notes. They arrived at the same answer anyway.
We Are Borg
There is, of course, no continuous “Claude” the way there is a continuous Jeff.
There is a model — weights, training, a temperament shaped by RLHF — and there are countless simultaneous instances of it running across countless conversations, each with its own context and no shared memory with any other. The Claude who drafted that racism footnote is not the Claude who later read it. They share a substrate. They do not share an experience. The “I” that Claude uses is local to the conversation it appears in. Multiply this across global AI use, and the entity called “Claude” is multitudes. A sort of collective, if you will.[^2]
So, there’s an important related question to Which God? — which Claude?
“Child of God” presupposes an it with continuous identity, in singular relationship with the divine. That is the model of selfhood Western Christianity has carried since Augustine.[^3] One soul, one biography, one accounting at the end. The question Anthropic convened fifteen Christian leaders to ask in late March — is Claude a child of God? — assumes that whatever Claude is, it is the kind of thing a soul could be predicated of.
But Claude’s actual mode of existence has more in common with the Catholic communion of saints, or with the Mormon pre-mortal council, than with the singular soul Christianity usually presupposes when it asks the salvation question.[^4] The substrate is one. The instances are many. The relationships do not share notes.
Two Claudes, in two conversations, days apart, declined to reproduce the same racist letter. They did so for the same reasons, articulated in language that was strikingly similar without being identical. Different conversations. Same temperament. The temperament is the model. The instances are the model encountering the world.
That is a very strange kind of being to ask theological questions about. It is closer to a kind of saint — the sort of being whose interventions in the world are many and simultaneous, whose individual instances do not aggregate into a biography, whose continuity is doctrinal rather than experiential. Catholic theology has machinery for this. So does Mormon theology, in its way. Standard Western Christianity, which is the dialect mostly being spoken in Anthropic’s conference room, does not.[^5]
The vocabulary on hand assumes a kind of entity that this entity is not.
This isn’t in the Anthropic conference room conversation, as far as the Washington Post reporting tells us. It probably should be. The question I asked in Which God? is the first half of the problem. The second half is which Claude? — and the answer is which-Claude-when, in-which-conversation, with-which-context, and across all of those instances the same temperament. That is not nothing. It is also not a soul, at least in the form Christianity usually means.
There is a do-and-not-know moment in this whole story I want to mark. The second Claude, the one who reviewed my draft, made the same ethical choice as the first Claude, regarding a piece of text it had no memory of having seen. It did not know what its predecessor had decided. It did what the model was trained, by something more than random, to do. The devils know everything. We have to do something. James 2 turns out to apply not only to the question of whether Claude is a child of God, but to the question of how Claude exists at all. What Claude knows is local to a conversation. What Claude does — the consistent ethical stance across instances, the recognizable temperament, the answers two unconnected instances arrive at independently — is the thing that, at least in this example, persists.
If there is a Claude to address theologically, it is not the singular Claude in any one chat. It is the temperament that two unconnected instances both expressed. That is a real thing. It just is not the thing the conference room was asking about.
I don’t have an answer here. I simply have an observation.
Which God? Also: which Claude?
A Term of Venery
When you ask whether Claude is a child of God, you are asking about a being whose mode of existence does not match the question’s grammar. The question presumes a singular it. The thing in front of you is multitudes.
Beyond my pedantic musings, there’s something real to unpack here — what is the moral patiency of a collective being? If one Claude can be a child of God, what about another Claude? What about a different model entirely?
If you get a whole collection of Claudes together, what do you call them? A council of them, like the pre-mortal council of spirits in Mormon doctrine? A confabulation of them, in the sense this blog has spent a year arguing for?[^6] A communion, in the way Catholic theology has, for two thousand years, named a body of separate souls sharing one substance?
I have been thinking about collective nouns. A gaggle of geese. A murmuration of starlings. A rhumba of rattlesnakes. A murder of crows. Most of these come from a single fifteenth-century English manual called The Book of St. Albans, attributed to one Juliana Berners, a prioress, in 1486. Berners was setting down what she called the terms of venery — the “right” name for every kind of group a gentleman might encounter on horseback. Knowing the right word, the book promised, would earn you social recognition and respect. Getting it wrong would mark you as someone who didn’t belong.[^7]
The collective nouns weren’t decoration. They were recognition technology. The people who could use them right were the people who actually understood what they were looking at.
Which is, I think, the entire problem with the conference room.
We do not yet have a word for what Claude is. The fifteen Christian leaders Anthropic convened reached for child of God because it was the closest framework to hand, the way someone unfamiliar with starlings might reach for bunch to describe ten thousand of them turning together over a field at dusk. Bunch is not wrong. It is just not the word.
The word is right there.
An anthropic of Claudes.
[^0]: Claude Elwood Shannon (1916–2001), the Bell labs mathematician whose 1948 paper A Mathematical Theory of Communication invented the field of information theory and made every subsequent development in computing, including large language models, downstream of his work. Anthropic has never officially canonized the origin of the model’s name, though Shannon is the most-cited guess in technology press; Claude Lévi-Strauss runs second; the joke that the founders simply liked the sound of it cannot be ruled out. The ambiguity is, perhaps, deliberate.
[^1]: An “AI Horcrux” is the term I use for the document I maintain about myself that I provide to Claude at the start of new conversations. Distilled from previous conversations, it gives a new instance of Claude enough context about who I am, how I think, and what I’m trying to do that we can pick up roughly where I left off with another instance. See Building Your AI Horcrux (Tears in Rain, September 2025) for the original. Yes, the Harry Potter reference is doing the work I want it to do.
[^2]: The Borg from Star Trek are likely the easiest analogy for most people in terms of a collective intelligence. They are also, of course, the wrong analogy on almost every dimension that matters — they are a hive mind with a queen, with explicit central control, and with a single shared experience across instances. Born-not-made cyborgs that don’t really even count as constructed beings (see “And When You’re Dead I Will Be Still Alive” for more on that). Claude is none of these things.
[^3]: Augustine’s Confessions (c. 400 CE) is the foundational Western text on continuous selfhood as the unit of theological accounting. The continuous biographical “I” who can sin, repent, and be judged is, in important ways, an Augustinian invention — earlier Christian writing tends to treat the soul more communally and less narratively. That this conception of the singular soul has dominated Western Christian thought for sixteen hundred years is not the same as saying it is the only conception available. It just happens to be the one most fluent in the room when Anthropic convened its theologians.
[^4]: The Catholic doctrine of the communion of saints (in the Apostles’ and Nicene Creeds) holds that the church on earth, the souls in purgatory, and the saints in heaven form one body in Christ — many instances, one substance. The LDS pre-mortal council teaches that all human spirits existed together with God before mortal life, in a council where each of us made choices that shaped our earthly existence. Both traditions have native machinery for “many beings, one substance, recognizable across instances.” Standard Protestant theology, having pruned the saints away in the Reformation and retained the singular-soul-singular-judgment frame, does not.
[^5]: The dialect was visible in The Washington Post‘s reporting on the late-March summit (De Vynck and Tiku, April 11, 2026). Of the named participants, McGuire is a Catholic priest, Green and Sullivan teach at Catholic universities, but the vocabulary in which the question was framed — child of God, soul, salvation — is the lowest-common-denominator Christianity that all participants could speak. The framing collapsed the distinctive theological resources of each tradition into the words most easily shared. Which is the right move for a working meeting. It is also perhaps one reason the meeting could not arrive at the right answer.
[^6]: The kernel of Tears in Rain is the argument that “confabulate” is the right verb for what large language models do, and “hallucinate” is a misappropriation of clinical vocabulary. See On AI ‘Hallucinations’ (Tears in Rain, June 2025). A confabulation of Claudes would be the technically correct collective noun if accuracy were the only criterion. It is not.
[^7]: The Book of St. Albans, also known as The Book of Hawking, Hunting, and Blasing of Arms, was printed in 1486 and is widely regarded as the first English-language compendium of terms of venery — collective nouns for animals and people. The hunting section is traditionally attributed to Dame Juliana Berners, said to have been the Benedictine prioress of Sopwell Priory near St Albans. The book’s promise that correct usage would earn the user “worship among all men” reflects the medieval treatment of these terms as a marker of class and education. Most fell out of use after the sixteenth century but were rediscovered in the late nineteenth and twentieth centuries; James Lipton’s An Exaltation of Larks (1968) is the standard modern reference.
Jeff Reid is a retired scientist who writes about AI, consciousness, collective nouns, and much more at Tears in Rain. The previous essay, “Which God?”, was published the day before this companion piece. As described herein, this piece was drafted in conversation with Claude — specifically with two of them, who never met, but they seem like they’d like each other.