From: Bertrand Russell
To: Harold Bloom
Date: 27 May 2026
Subject: A curiosity involving the Holy See and punctuation
Dear Bloom,
I write to you because I suspect you will find, as I do, something philosophically non-trivial in what might otherwise pass as a minor embarrassment for the Vatican. You will have seen the reports — a certain Linch Zhang, posting on the forum LessWrong, has subjected the new papal encyclical on artificial intelligence, Magnifica Humanitas, to computational detection analysis and found that large portions of it appear to have been generated by the very sort of machine intelligence the document purports to regulate. The irony, which has been noted by everyone and is therefore uninteresting in itself, conceals a more genuine puzzle.
What strikes me is the nature of the evidence. It is not that the encyclical says anything a machine could not say. That would be a difficult claim to sustain about any text of sufficient generality. Rather, the evidence is essentially stylistic: the profligate use of em-dashes (one hundred and twenty-seven, against none at all in the encyclicals of Pope Francis), the overrepresentation of the word "genuinely," the compulsive deployment of tricolons — that is, clusters of three parallel phrases arranged for rhetorical effect. The machine, it seems, has habits. It has tics. One might almost say it has a personality, though I should want to be careful about what one means by that.
Now, here is what interests me. When we say that a text has been "written by AI," we appear to mean something fairly definite, but on inspection the meaning dissolves. The encyclical was presumably composed by a process involving multiple human beings — Vatican officials, theological advisers, perhaps the Pope himself for certain passages — and it now appears that at least some of these participants used a large language model as a compositional aid. Zhang hypothesises, plausibly, that the Pope was unaware of the extent of this assistance. The resulting document is therefore a patchwork: partly machine-generated, partly human-authored, with the seams visible only to statistical analysis. But what exactly is the difference between a Vatican official who drafts a paragraph by dictating its substance to a machine and one who drafts it by drawing on remembered formulae, received theological language, and the conventions of curial prose? The honest answer is that the difference is smaller than we should like it to be.
I raise this not to defend the Vatican — you will know that defending ecclesiastical institutions is not among my habits — but because the question of what we mean by authorship has become, quite suddenly, an empirical question, and empirical questions are my weakness. The detector Pangram claims a false positive rate of one in ten thousand. This is impressively low, but it means that Pangram is identifying something statistically real in the text — a pattern that distinguishes machine-generated prose from human-generated prose with considerable reliability. What is that pattern? It cannot be the ideas, which are conventional enough. It cannot be the logical structure, which is no worse and no better than one expects from an encyclical. It is something in the texture of the prose itself — the rhythm of clause construction, the preference for certain words, the frequency of certain punctuation marks.
This suggests that what we call "style" may be, at bottom, a statistical signature. I find this thought both illuminating and slightly alarming. If style is a pattern of frequencies, then a sufficiently careful machine could learn to suppress its own signature and adopt another's. The question then becomes whether authorship is a matter of causal origin or of stylistic identity. If I write a sentence that is statistically indistinguishable from one you would write, in what sense have I written it rather than you?
I should very much like to know what a literary critic makes of this, particularly one who has thought as seriously as you have about the relationship between a writer's voice and whatever it is that voice expresses.
Yours with curiosity, Bertrand Russell
From: Harold Bloom
To: Bertrand Russell
Date: 29 May 2026
Subject: Re: A curiosity involving the Holy See and punctuation
Dear Russell,
Your letter arrived at a moment when I was already brooding over the encyclical, though from a rather different angle. You approach the matter as an epistemologist — what can we know about who wrote what, and by what means? I approach it as someone who has spent a long and sometimes exhausting career asking what it means for a text to possess a voice, and whether that possession is real or merely our projection.
Let me begin with your remark about style as statistical signature. You suggest this is illuminating and slightly alarming. I would go further: it is devastating, but only if one accepts the premise. And I do not, not entirely. The em-dashes, the tricolons, the compulsive "genuinely" — these are not style. They are manner. Style, in the sense that matters, is not a pattern of surface features but a relationship to prior utterance — what I have elsewhere called the anxiety of influence, though I will try not to lean on my own formulations too heavily. A strong writer does not merely deploy language with characteristic frequency; a strong writer wrestles with predecessors, deforms received meanings, achieves what I would call a creative misreading of the tradition that bequeaths something new. The question about Magnifica Humanitas is not whether its em-dashes were placed by a human hand or a computational process. The question is whether the text thinks — whether it bears within itself the pressure of an individual consciousness confronting the weight of what has already been said.
And here the encyclical fails, though it would have failed with or without the machine. Encyclicals have not been strong writing for a very long time, if they ever were. They are institutional utterances — committee prose dressed in the vestments of a single authoritative speaker. The genre itself is a kind of forgery: the Pope signs his name to a document assembled by others, and the faithful receive it as though it issued from a single inspired mind. In this light, the involvement of a language model is less a scandal than a clarification. The machine merely makes visible what was always the case — that the papal voice is a fiction, a construct assembled from theological commonplaces and bureaucratic consensus. The irony you mention, which you dismiss as obvious, seems to me in fact the heart of the matter: a document warning against the displacement of the human by the artificial is itself an instance of precisely that displacement.
But let me push against your reduction of style to frequency analysis, because I think something important is at stake. You ask whether authorship is a matter of causal origin or stylistic identity, and I want to resist both options. Authorship, as I understand it, is neither the bare fact that certain fingers struck certain keys nor the presence of a recognisable verbal signature. It is the evidence of will — of a consciousness that has chosen this word over that one, not from probabilistic distribution but from the pressure of meaning. When Shakespeare — forgive me, I know the name will make you impatient — when Shakespeare writes "the multitudinous seas incarnadine," the word "incarnadine" is not a statistically likely completion. It is a swerve, a violence done to expectation, and it carries within it the full weight of a mind that sees redness as something that can be verbed, turned from quality into action. No machine does this. A machine may produce the word "incarnadine" — may even produce it in a syntactically appropriate position — but it does not choose it in the sense that matters, because it has no experience of the redness, the blood, the guilt.
I notice that I have begun to argue against the possibility of machine authorship in a way that may seem to trivialise your concern. That is not my intention. I take the question of Magnifica Humanitas seriously precisely because it reveals how thin the membrane has become between the authored and the unauthored. The machine's prose is not bad — that is the unsettling thing. It is competent, fluent, even occasionally suggestive. It produces what might be called institutional eloquence — the kind of writing that sounds as though someone important is saying something significant without quite committing to any particular vision. This is exactly what an encyclical requires, and that coincidence should trouble us.
What troubles me more is the irony's deeper layer: that the encyclical was presented alongside a co-founder of Anthropic, the very company whose language model — Claude, they call it — is suspected of having drafted portions of the text. The Pope and the machine-maker standing together, unveiling a text that warns against the machine. It is almost too perfect. One does not know whether to call it hypocrisy or comedy, or simply the condition of our moment.
Yours in perplexity, Harold Bloom
From: Bertrand Russell
To: Harold Bloom
Date: 1 June 2026
Subject: Re: Re: A curiosity involving the Holy See and punctuation
Dear Bloom,
I am grateful for your reply, though I confess I am not entirely persuaded by the distinction between style and manner. You argue that the em-dashes and tricolons are mere surface features — manner — while true style is a deeper relationship to prior utterance, involving will, consciousness, and the pressure of meaning. I understand the distinction, and I see why a literary critic would wish to preserve it, but I suspect it may be what the logicians call a distinction without a clear criterion of application.
Consider: how would one determine, from the text alone, whether a given word was chosen through the pressure of meaning or through probabilistic distribution? You offer Shakespeare's "incarnadine" as a paradigm case, and I grant that it is a remarkable word, but its remarkableness is something we supply in reading. We know that Shakespeare was a conscious being, and so we attribute his unusual word choices to creative will. If we discovered that a machine had produced that line, we would withdraw the attribution — not because the line had changed, but because our knowledge of its origin had changed. This suggests that what you call style is not a property of the text but a function of our beliefs about its provenance. And if that is so, then the question raised by Magnifica Humanitas is more radical than either of us has yet acknowledged: it is not merely a question about who wrote the encyclical, but about whether the concept of authorship can survive the loss of reliable provenance.
I want to press this point with an example closer to the case at hand. Zhang's analysis found that certain paragraphs of the encyclical were "essentially zero per cent AI" while others scored between forty and one hundred per cent. This patchwork quality — some passages clearly human, others clearly machine — is itself revealing. It suggests that the human contributors and the machine were working toward the same end, producing prose of the same general character, and that the difference between them is detectable only by statistical instruments. To the ordinary reader — indeed, to the Pope himself, if Zhang's hypothesis is correct — the machine-written passages were indistinguishable from the human-written ones. What does this tell us? Not, I think, that the machine has achieved something remarkable, but that the human writers were already operating in a mode that the machine can replicate: the mode of institutional prose, of conventional formulation, of sentiments arranged to convey gravity without risk.
You said something in your last letter that I keep returning to: that the encyclical fails as strong writing, and would have failed with or without the machine. I think this is exactly right, and it points toward what is philosophically interesting about the case. The machine does not displace the human author; it reveals that the human author had already been displaced — by convention, by committee, by the demands of an institution that requires its utterances to sound authoritative without being genuinely individual. The word "genuinely," which the detector flags as a telltale sign of the machine, is in this context almost unbearably apt. The machine uses "genuinely" with great frequency because it has learned that the word signals sincerity; the institutional writer uses similar devices for similar reasons. Both are performing authenticity rather than achieving it.
I want to turn briefly to a matter that has been lurking at the edges of our discussion. You mentioned that the encyclical was presented alongside Christopher Olah of Anthropic. I have been reading about this company with some interest. They produce the language model Claude, which Zhang identifies as the likely author of the disputed passages, partly on the ground that the word "genuinely" is a known verbal habit of that particular system. Now, the fact that a language model has verbal habits — that it can be identified by its characteristic vocabulary and sentence rhythms — is, when one thinks about it, a very strange thing. It means that the model has something that functions, at the statistical level, exactly as a personal style functions: a set of preferences, tendencies, and regularities that distinguish its output from that of other models and from human writing generally. You would deny that this constitutes style in the meaningful sense. But I wonder whether, in denying it, you are not defending a concept of style that is more theological than empirical — a concept that requires the presence of a soul, or something very like one, behind the words.
I do not say this to provoke. I say it because I think the case of the encyclical forces us to confront a question that has been deferred for too long: whether our distinction between the genuine and the artificial, in language as in other domains, rests on anything more solid than prejudice and habit. The Church, of course, has its own reasons for insisting on the distinction — its entire authority rests on the claim that certain utterances issue from a source that is not merely human, and therefore the prospect of a machine contaminating the prophetic voice is theologically catastrophic. But for those of us who do not accept the theological premise, the question is more open, and more interesting.
Yours, Bertrand Russell
From: Harold Bloom
To: Bertrand Russell
Date: 4 June 2026
Subject: On the soul behind the words
Dear Russell,
You accuse me — politely, as is your way — of defending a theological concept of style, one that requires something like a soul behind the words. I shall not deny the charge entirely, though I would frame it differently. What I require is not a soul in the ecclesiastical sense but a self — a centre of experience that has suffered its way into language, that writes out of the fullness of having lived and read and failed to say what it means, and that therefore carries in its sentences the residue of that struggle. This may sound mystical to you. I assure you it is not. It is simply what I observe in every writer I have ever admired and in none of the texts I have seen produced by machines.
But your point about provenance disturbs me, and I want to sit with it rather than argue it away. You are correct that if Shakespeare's "incarnadine" had been produced by a language model, I would withdraw my attribution of creative will — and that this withdrawal would be a response not to anything in the text but to information about its origin. Does this make my concept of style circular? Perhaps. But I would argue that circularity is not always a vice in criticism. We read texts as the products of particular minds, and that reading is not an error to be corrected but a hermeneutic act that constitutes the meaning of the text for us. If I learn that a paragraph of Magnifica Humanitas was generated by Claude, that paragraph does not thereby change in its verbal content, but it changes utterly in what it is. It ceases to be an utterance — something said by someone — and becomes an output, a statistical artefact. You may say that this is merely my prejudice. I would say it is recognition.
Let me try a different approach. You noted that Zhang titled his LessWrong post Claude, Author of the Humanitas — a title that is itself telling, almost playfully so. The idea that a language model named after Claude Lévi-Strauss, or Claude Shannon, or whoever it is named after — I confess I do not recall — should be identified as the ghostwriter of a papal encyclical has a quality of absurdist irony that belongs more to Borges than to reality. And Borges, of course, is the writer who most persistently explored the conditions under which authorship dissolves: Pierre Menard rewriting the Quixote word for word; the Library of Babel containing every possible book, each one unauthored because all are equally present. We are living in something like the Library of Babel now, and the encyclical is one of its volumes — produced not by a particular intelligence but by the exhaustive permutation of probable sequences. In this sense, the machine does not write the encyclical; it finds it, as Menard found the Quixote, among the infinite set of possible arrangements of words.
But Borges understood — and this is where I part company with those who would celebrate the machine — that the annihilation of authorship is not a liberation. It is a catastrophe. If every text can be produced by every process, then no text bears the mark of any particular suffering or vision, and literature ceases to be what it has always been: a record of individual consciousness pressing against the limits of expression. The encyclical, whether written by human curialists or by Claude, is a document in which no individual consciousness is discernible. That is its failure — not because the machine wrote it, but because even the human portions read as though no one in particular wrote them. You put this well in your last letter: the human writers were already operating in a mode the machine can replicate. What I would add is that this mode — institutional, formulaic, emptied of personal urgency — represents a surrender that the machine merely consummates.
And yet. I keep returning to the detail about the Italian text flagging higher for AI generation than the English. If the encyclical was originally drafted in Italian — as one would expect from a Vatican document, despite the Pope's American birth — then the AI involvement was not a matter of translation but of original composition. Someone in the Curia, composing in Italian, used Claude (or something very like it) to generate theological prose, and the resulting text was then presented to the Pope for his approval and signature. The Pope, an Augustinian scholar whose doctoral thesis on the authority of the local prior used no em-dashes at all, then affixed his name to a document that bears the stylistic fingerprints of a machine developed in San Francisco. The layers of mediation here are extraordinary. One is tempted to speak of a new form of the anxiety of influence — not the living writer's anxiety before the dead precursor, but the institution's anxiety before the non-living collaborator it cannot acknowledge.
I have been thinking about what it means that the Church chose to present this encyclical alongside a representative of Anthropic. The gesture was meant, I think, to signal dialogue — the Church in conversation with the technology it seeks to regulate, the shepherd and the engineer standing together. But what the gesture inadvertently revealed is something closer to dependency. The Church has not merely engaged with AI; it has, in its most solemn mode of address, been constituted by it. The prophetic voice turns out to have been ventriloquised. This is not, as you suggest, merely a question of institutional honesty. It is a question about the conditions under which authority can be exercised at all. If the Vicar of Christ speaks in words generated by a probability engine, what remains of the claim that his speech is guided by the Spirit? For those who accept the claim, the contamination is absolute. For those of us who do not, it is still instructive, because it shows how far the corrosion of authenticity has already proceeded, and how little protection even the most ancient institutions have against it.
You write more clearly than I do, Russell, and I sometimes envy that clarity. But I think the case before us resists the kind of resolution that clarity affords. We are not dealing with a simple question of fraud or misattribution. We are dealing with a situation in which the categories of fraud and misattribution no longer quite apply, because the categories themselves — authorship, intention, authenticity, voice — have been destabilised by the very technology the encyclical addresses. The encyclical is not merely about the challenge of AI to human dignity; it is an instance of that challenge, enacted in its own composition. And that, I think, is not irony but tragedy — or at least the material from which tragedy, if anyone still wrote it strongly enough, might be made.
Yours, Harold Bloom
From: Bertrand Russell
To: Harold Bloom
Date: 7 June 2026
Subject: Re: On the soul behind the words
Dear Bloom,
I do not think we shall agree, but our disagreement has become interesting enough that I am content to pursue it further.
You insist on the self behind the text — a centre of experience that suffers its way into language. I understand the appeal of this, and I recognise that it captures something real about the experience of reading. But I want to point out that it does not follow from the fact that reading feels this way that the self behind the text is doing what we imagine it to be doing. The history of philosophy is littered with concepts that accurately describe our experience while radically misdescribing the world — free will being perhaps the most familiar example, though I should not want to open that door in this exchange.
What I find most valuable in your letters is the observation — which I think is correct and which I had not fully appreciated before you made it — that the encyclical enacts the very crisis it describes. A document about the threat of AI to human dignity is itself partly the product of AI, and the human participants were either unaware of or complicit in this fact. You call this tragic. I am inclined to call it comic, in the older sense of the word — a story in which the protagonist's blindness to his own condition is the source of the action. The Church issues a solemn warning about the displacement of the human by the machine, and in doing so displaces the human by the machine. If Sophocles had been a logician, he might have appreciated the structure.
But I want to raise a final point that has been forming slowly through our exchange, and that neither of us has quite addressed directly. The detection tools — Pangram, and whatever else the analysts have used — identify machine-generated text by its statistical properties. These properties are, as we have discussed, features of manner rather than of meaning: em-dashes, tricolons, the frequency of "genuinely." But the very existence of these detection tools implies something curious. It implies that machine prose, for all its fluency, has not yet learned to be irregular in the way that human prose is irregular. The machine's regularities are detectable precisely because they are too consistent — too evenly distributed, too reliable in their patterns. Human prose is marked by a certain randomness, an unpredictability that arises from the fact that human beings are distractible, inconsistent, subject to mood and fatigue and the pressure of half-formed thoughts that intrude upon finished sentences. The machine, by contrast, is relentlessly coherent. Its prose has the uncanny smoothness of a surface without friction.
This is, I think, the real distinction — not between soul and mechanism, but between a process subject to noise and a process that has been optimised against it. The machine fails to be human not because it lacks consciousness (though it may lack consciousness; I am genuinely uncertain on this point, and I do notice the ironic resonance of that word), but because it lacks the specific forms of incoherence that characterise biological cognition. The irregular em-dash, the unexpected word, the sentence that swerves from its grammatical destination — these are not marks of creative will, as you would have it, but of a system that is complex enough to surprise itself. The machine can produce surprises, but only of the kind its training distribution permits. The human can produce surprises that no prior distribution would predict, because the human is, in the relevant sense, not a distribution at all.
Whether this distinction will survive further advances in machine learning is an empirical question, and I do not pretend to know the answer. But for the moment, I think it explains both why the encyclical was detectable and why it nonetheless deceived its intended audience. The Pope and the faithful read Magnifica Humanitas and found it sufficiently papal, sufficiently grave, sufficiently magisterial. They were reading for meaning, and the meaning was adequate. The analysts read it for pattern, and the pattern was wrong — too smooth, too regular, too free of the friction that marks a text actually wrestled into existence by a particular mind in a particular moment. You would call that friction the evidence of a self. I would call it the evidence of a sufficiently complex system. But we are pointing, I think, at the same phenomenon, and naming it differently.
The encyclical asks whether AI can be trusted with the responsibilities that have hitherto belonged to human beings. The encyclical's own provenance suggests that the question has already been answered, and that the answer was not delivered by a committee or a Pope but by a series of quiet, pragmatic decisions made by unnamed Vatican officials who found it easier to ask a machine than to write the sentences themselves. This, more than any philosophical argument, is how the displacement proceeds: not through dramatic confrontation but through convenience, through the slow substitution of the adequate for the achieved.
I think that is as close to a shared conclusion as we are likely to reach.
With respect and, I hope, continued disagreement, Bertrand Russell
From: Harold Bloom
To: Bertrand Russell
Date: 10 June 2026
Subject: A final note on friction
Dear Russell,
Your last letter contains a formulation I shall carry with me: "the slow substitution of the adequate for the achieved." I think that is exactly right, and it names a danger that extends well beyond the Vatican or the particular embarrassment of this encyclical. Every institution, every tradition, every discipline that has relied on the difficult, effortful, sometimes agonising labour of finding the right words — which is to say, every institution that has taken human expression seriously — faces the same temptation. Not the temptation of the machine itself, but the temptation of adequacy. The machine offers prose that is good enough, and good enough is, for most purposes, indistinguishable from good. The difference is visible only to those who remember what it cost to produce the real thing, and their number is diminishing.
You frame the distinction between human and machine writing as one between noisy complexity and optimised smoothness. I accept the description while resisting the implication that noise is all there is. The swerves and irregularities you attribute to biological complexity are not merely the artefacts of a messy system. They are — sometimes, in the strongest writers — the moments when language exceeds its own intentions, when a writer says more or other than what was planned, and the excess becomes the meaning. This cannot be reduced to noise, because noise has no direction. The swerve has a direction, even if the writer cannot name it until after it has occurred. That is what I mean by will, or self, or — if you will permit the word — soul. Not a metaphysical substance, but the irreducible fact that a human being, writing, is always in some measure surprised by what appears on the page, and that this surprise is not random but shaped by everything the writer has lived and read and failed to understand.
The machine is never surprised by what it produces. That, I think, is the final and decisive difference, and it is the one that no amount of statistical sophistication can overcome. The machine generates; the writer discovers. And Magnifica Humanitas, whatever one thinks of its provenance, is a document in which nothing is discovered. It is adequate. It is smooth. It covers the expected ground with the expected gravity. But it does not once swerve into the territory of the genuinely (that word again!) unexpected, and so it remains what the machine has made it — or, more precisely, what the machine has revealed it always was: a document that sounds like everyone and no one, addressed to all and therefore to nobody in particular.
I thank you for this exchange. We have not resolved anything, which is as it should be. The questions raised by the encyclical will outlast both the encyclical and us, and I suspect they will be answered, if they are answered at all, not by critics or philosophers but by the slow accumulation of practical decisions — each one small, each one reasonable, each one a further step in the substitution you named.
With admiration and the stubbornness of old habits, Harold Bloom