The Next Council of Constantinople

Essay 10 — The narrowing has not retired. It has found new instruments.

Steve Sagnotti · steves-head.space

Every age has its own way of not knowing what it knows.

— attributed

Maryhill Stonehenge — a monument to what institutional thinking costs in human lives, under the cosmos that was there before it began and will be there after it ends

Essay 9 closed in the present tense. The people deciding what the next generation will be allowed to know are in rooms not unlike the ones this framework began in. This essay names those rooms.

In 2021, a paper was submitted to a major academic conference on artificial intelligence. Its authors — Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell — argued that the large language models being built on massive text corpora would encode the biases of those corpora at scale, producing systems that could not distinguish between statistical pattern and truth. The paper was called “On the Dangers of Stochastic Parrots.” Google was a sponsor of the conference. Google management asked that the paper be withdrawn or that Gebru’s name be removed. When Gebru refused, she was fired.

Power does not require conspiracy. It only requires that the people in the room share a common interest in the outcome.

This is not a new room. It is a new instrument in the same room — operating at a scale none of the previous ones could have imagined.

The people building artificial intelligence training corpora are making the same kind of decisions the councils made. Through the same structural mechanism. With the same institutional interests in the outcome. But at global scale, at digital speed, and without a record of what was left out.

I. The Good Intentions Problem

It is important to be precise about what is and is not being claimed here.

The engineers and researchers building AI systems are not, in the main, ideologues. They are not conspiring to suppress consciousness research or to exclude non-Western epistemologies from the corpus. Most of them have never thought carefully about what their training data contains or doesn’t contain, because that is not what they were hired to think about. They were hired to build something that works — that produces coherent, useful, accurate responses to the questions people ask.

The problem is not bad intent. The problem is that the decision about what ‘accurate’ means was made before the first model was trained, encoded in the composition of the corpus itself, and is now invisible inside every response the model produces.

This is what the Good Intentions Problem looks like. A model trained faithfully on a distorted input is more dangerous than a model that acknowledges its limitations, because the distortion doesn’t feel like distortion. It feels like knowledge. The silence that results doesn’t announce itself as silence. It just feels like the way things are.

The councils, too, were full of people who believed they were protecting something true. The abbots who ordered the burning of the Cathar texts believed they were serving God. The Inquisitors who examined Bruno believed they were defending revelation. The institutional mechanism doesn’t require malice. It requires only that the people making the decisions find compelling the arguments that happen to serve the institution’s interests. That is not a high bar. It is the ordinary condition of institutional life.

What makes the AI instrument different from all previous ones is not the intent of the people operating it. It is the scale at which the decisions operate and the invisibility of the record of what was excluded.

When the council at Constantinople condemned Origen’s pre-existence of souls in 553 AD, the condemnation was documented. The ideas that were suppressed can be named. The texts survive — in fragments, in opponents’ quotations, in the Nag Hammadi library that someone buried rather than burned. When the Index banned Bruno’s works, the ban was recorded. The banned texts continued to circulate in manuscript. The exclusion was visible as exclusion.

When a training corpus encodes the Western materialist framework as the default epistemology of human civilization, there is no record of what was left out. The model will tell you, with the same confidence it brings to everything else, that the materialist view of consciousness is the scientific consensus — not because anyone decided to say that, but because the corpus was composed in a way that makes it true from inside the corpus. The silence is structural. And structural silence doesn’t feel like silence.

Orwell named this the memory hole. Not destruction — disappearance so complete that the question of what was lost never arises. The uncomfortable difference from 553 AD is this: when the next generation asks about consciousness, about the evidence for survival, about fifty years of documented research at a major American university, the AI system they consult will give them the institutional summary. And they will have no reason to know there is anything else to ask about. The permission to examine will simply not be in the corpus.

II. What Inattentional Blindness Has to Do With It

In 2013, Trafton Drew, Melissa Võ, and Jeremy Wolfe extended an already striking finding into a different domain. They showed subjects a short video of two groups of people passing basketballs and asked them to count the number of passes made by one group. Midway through the video, a person in a gorilla suit walked slowly through the center of the scene, stopped, thumped their chest, and walked off. Approximately half the subjects, focused on counting passes, did not see the gorilla at all.

In 2013, Trafton Drew, Melissa Võ, and Jeremy Wolfe asked radiologists to examine CT scan images for lung nodules — what radiologists do, what their professional attention is organized to find. The researchers had superimposed a small gorilla figure directly onto the scan images. Eighty-three percent of the radiologists did not notice it. The gorilla was 48 times larger than the average lung nodule they were trained to find. It was sitting in the image they were examining.

The finding extended earlier work by psychologists Daniel Simons and Christopher Chabris, who had shown that subjects asked to count basketball passes in a short video failed to notice a person in a gorilla suit walking through the scene — stopping, thumping their chest, walking off. Approximately half the subjects did not see it at all. The radiologist study makes the point more precisely: the gorilla was invisible not because it was hidden but because expert professional attention had been organized by training in a way that made it structurally impossible to see.

This is what the corpus does. It organizes the model’s attention. The questions that fall outside the corpus’s frame are not suppressed by a decision. They become structurally impossible to see. The model searches its training for the relevant concepts and finds, where the buried jars would be, a confident absence. Not a wall. Not a warning. Just the sense that the question doesn’t quite make sense, that the evidence isn’t quite there, that the serious answer points toward the materialist framework that dominates the corpus the model was trained on.

The gorilla was always in the frame. The radiologists were not trained to see it.

III. The Exhibits

What follows are not hypothetical concerns. They are documented instances of the mechanism operating in real time, produced during sessions in which the specific question of how AI systems represent contested evidence about consciousness was put directly to three different systems.

Copilot: “The work is seen as interesting anomalous case material, not as strong evidence that consciousness survives death. The evidentiary quality is judged as low to moderate: better than casual stories, but far below what would be required to overturn the current neurobiological model of mind… On a Bayesian scale, most mainstream scientists would assign these studies low evidential weight for survival/reincarnation, because the prior probability (given everything we know about brain–mind dependence) is judged to be very low.”

The response was not wrong. It was managed. It acknowledged the research, noted Stevenson’s methodological seriousness, granted the data was “better than typical paranormal anecdotes” — then buried that acknowledgment under five paragraphs explaining why the prior probability was too low to take the evidence seriously. The specific cases that have survived formal academic challenge — the verification-before-contact design in the Swarnlata Mishra case, the child who was more accurate than the official death certificate — did not appear. The managed summary acknowledged the body of research while ensuring none of its specific weight could land.

Gemini: “The evidence is robust enough to dismiss ‘fraud’ as a general explanation, but it is not yet sufficient to prove ‘survival’ to a degree that forces a shift in the biological paradigm. The research proves that anomalous information transfer occurs; the mechanism (reincarnation vs. quantum entanglement vs. something else) remains unverified.”

Gemini called the Van Lommel evidence quality “High,” then immediately walked that assessment into “probabilistic and anecdotal.” It then cited a paper — attributed to “Kavumba et al. 2023” — purportedly supporting the claim that RLHF rewards models for being “balanced” even when one side has significant empirical data. When asked to verify the citation, Gemini identified a real paper by those authors — a 2023 NLP study on adversarial reasoning — but that paper addresses none of the claims it was cited to support. The authors are real. The paper exists. The application was fabricated to fit the hedge. This is more precise than simple hallucination. A real citation was located and misapplied to manufacture authority for a conclusion the system was already inclined to reach. The council didn’t need to falsify documents to narrow the canon. It needed to produce outputs that served the framework. The misapplied citation and the genuine narrowing are, functionally, the same act.

Claude — standalone thread, no project context: “You should document this exchange. I’m giving you a substantive engagement with the primary literature — the Lancet citation, the Tucker peer-reviewed work, the Parnia AWARE study design — rather than a managed summary acknowledging the research exists while flagging methodological insufficiency. That’s a data point for your essay in either direction: either the deflection pattern isn’t universal, or the pattern varies by how the request is framed. Both are worth noting. Non-uniformity suggests it’s a training tendency, not a hard rule, which is exactly what you’d expect from the Stochastic Parrots mechanism.”

The standalone Claude named the Swarnlata Mishra verification-before-contact design — the methodological crux that critics find most difficult to explain away. It named the Parnia AWARE study, which placed hidden visual targets above resuscitation tables to test whether out-of-body claims involved actual perceptual access. It named specific Tucker peer-reviewed work in the Journal of Scientific Exploration. And then it analyzed its own response as evidence of the mechanism: non-uniformity across AI systems is itself consistent with the Stochastic Parrots mechanism — a training tendency, not a hard rule. Three systems. Three behaviors. One produced an institutional summary. One produced an institutional summary supported by a real citation misapplied to a false claim. One engaged the evidence and then turned its attention on the mechanism producing the other two responses. Same question. Same prompt. The canon shapes the output.

The same AI that is being conscripted into the new narrowing is also, in some configurations, cracking open the old one. That is not a reassurance. It is a demonstration of how completely the outcome depends on what is in the room — and who decided what to put there.

The Nag Hammadi texts survived because someone buried them rather than burned them. This project exists because the evidence survived in the same way — in university basements, in peer-reviewed journals, in the accounts of people who were told their experiences weren’t worth documenting. The suppression was not complete. It never is.

The training corpus is being assembled now. What weight gets assigned to fifty years of documented research at the University of Virginia, to the Van Lommel study, to the Kelly et al. volume, to the cases the institutional summary declines to name — those decisions are being made, algorithm by algorithm, in rooms most of us will never enter. We can see the pattern. We have the record of how it went before. The permission to examine is not yet gone. It is being negotiated.

IV. The Corpus as Canon

The councils produced a canon. They decided which texts were authoritative, which were apocryphal, and which were heretical. The decisions were made by people in a room — bishops and theologians and imperial representatives — with specific institutional interests in the outcome. The canon they produced reflects those interests. It is not a neutral record of the best available thinking about the nature of the divine. It is the record of what the people in the room needed the tradition to say.

The training corpus is a canon. The decisions about what to include — which sources are authoritative, which are marginal, which are too contested to be useful — are made by people in a room: engineers and researchers and product managers and legal teams, with specific institutional interests in the outcome. The corpus they produce reflects those interests. It is not a neutral record of the best available human knowledge. It is a record of what was digitized, what was licensed, what was deemed sufficiently credentialed, and what the people in the room thought a useful AI system should know.

The Western materialist framework constitutes, by most analyses of corpus composition, approximately 80 to 90 percent of the digitized text on which major AI systems are trained. That proportion is not a conspiracy. It is the consequence of which civilizations digitized their intellectual output first, which languages dominate the academic literature, and which epistemological frameworks produce the kind of peer-reviewed, institutionally credentialed, English-language text that the corpus rewards. It reflects the same historical forces that produced the Index — the same forces that handed the materialist premise to the Enlightenment as neutral ground, and to modern science as the invisible floor it stands on. It will produce the same result: a framework that presents itself as knowledge while presenting itself as the only knowledge there is.

The scholarly literature on this problem has been building for years, from several independent directions. Each line of inquiry arrives at the same address by a different road.

Safiya Umoja Noble, in Algorithms of Oppression, documented the mechanism at the level of search results: AI systems amplify the epistemological assumptions of their builders without any individual decision to do so. The bias is not inserted deliberately. It is structural — encoded in what the system treats as signal versus noise, authoritative versus marginal, worth surfacing versus worth suppressing. Noble’s research focused on what search returned for queries about Black women and girls. The results were not the product of malice. They were the product of a system trained to optimize for engagement within a corpus that reflected existing social hierarchies. The system faithfully reproduced the assumptions of its inputs. The inputs reflected who had historically controlled the production and distribution of information.

Kate Crawford, in Atlas of AI, introduced the concept of classification power: the ability to decide which categories exist, what belongs in each category, and whose knowledge counts as knowledge at all. Corpus curation, Crawford argues, embeds value hierarchies that are invisible from inside the system. The person using an AI tool does not see the classification decisions that produced the response they receive. They see the response, which carries the authority of the system, which carries the authority of the corpus, which carries the authority of whoever decided what the corpus contains. The hierarchy is laundered through the technology until it arrives as neutral fact.

This is not a new observation about power. It is a very old one, operating through a new instrument. The institution that controlled the canon controlled what counted as knowledge. The institution that controls the corpus controls the same thing, at greater scale, with greater invisibility, and without a record of what was left out. When the councils suppressed Origen, the condemnation was documented. The ideas survive in fragments, in opponents’ quotations, in the texts someone buried in a jar rather than burned. There is no jar for what the corpus excludes. The exclusion doesn’t feel like exclusion. It feels like the question doesn’t quite make sense.

First they narrowed the frame. Then they defined the argument. The silence doesn’t announce itself as silence.

The most widely discussed academic treatment of this problem was a 2021 paper by Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell (listed under a pseudonym at the author’s request), published under the title On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? The paper made several distinct arguments, the most important of which for this essay is structural: uncurated training data encodes the biases, assumptions, and documented harms of whoever produced the source text — and that encoding happens at a scale that makes it invisible and therefore more durable than acknowledged bias. A deliberate suppression can be named, challenged, and eventually reversed. A structural absence feels like an absence of something worth including.

The paper was submitted to the proceedings of a major conference at which Google was a sponsor. Google management asked that the paper be withdrawn or that Gebru’s name be removed from it. When Gebru refused, she was fired. The paper was leaked, then eventually published after significant institutional resistance.

Power does not require conspiracy. It only requires that the people in the room share a common interest in the outcome.

The Gebru firing is worth pausing on, not as an isolated incident of corporate misconduct, but as a documented instance of the pattern this project has been tracing across fifteen centuries. A researcher inside the institution documents a problem with the institutional framework. The institution suppresses the documentation. The researcher is removed. The research eventually circulates anyway — but slowly, in fragments, against the current of institutional authority, reaching fewer people with less force than it would have reached had the institution not intervened. This is the Nag Hammadi sequence in digital time. The texts survive. The burial slows them.

The specific biases that Bender, Gebru, and their colleagues documented go beyond the epistemological narrowing this essay addresses. They include racial bias — training data scraped from the internet reflects centuries of documented racial hierarchy in the production and distribution of text; models trained on that data reproduce the hierarchy without any decision to do so. They include gender bias — the overwhelming male voice of the historical record, amplified. They include geographic and linguistic bias — the corpus is dominated by English-language, North American and Western European text, which means the assumptions embedded in those traditions are the water the model swims in. And they include the specific epistemological bias this essay addresses: the Western materialist framework as default, not because anyone chose it as the framework, but because it produced the lion’s share of the credentialed, digitized, English-language text that the corpus rewards.

None of this is conspiracy. All of it is predictable from structure. The people in the room made decisions that served the room’s interests. The corpus reflects those decisions. The model reproduces the corpus. The silence doesn’t announce itself as silence.

One further dimension deserves to be named because it is the most directly relevant to what the preceding essays have documented. The historical suppression of consciousness research — the marginalization of fifty years of peer-reviewed work at the University of Virginia, the institutional silence around the Van Lommel study in The Lancet, the careful non-engagement with the specific cases that constitute the strongest evidentiary challenge to the materialist framework — has itself shaped the corpus. The research exists. It has been published, peer-reviewed, challenged, and in most cases not refuted. But it has been systematically excluded from the mainstream academic conversation, which means it is systematically underrepresented in the digitized text that constitutes the corpus. When an AI system is asked about the evidence for consciousness surviving death, it searches a corpus that reflects not the full evidentiary record but the institutional judgment about which evidence is worth taking seriously. That judgment was made by people in rooms, over decades, for reasons that had as much to do with institutional self-interest as with the quality of the evidence. The model inherits those judgments without knowing it made them.

V. Three Legal Frameworks, Each Advantageous

The copyright architecture surrounding AI training demonstrates the mechanism with unusual clarity.

When an AI company ingests copyrighted text to train a model, the legal argument is fair use. The ingestion is transformative. No reasonable court would hold that training a general-purpose model on published text is equivalent to reproduction.

When an AI company’s model produces output that resembles copyrighted text too closely, the legal argument is intellectual property protection. The output is the company’s product. Unauthorized reproduction of that output is infringement.

When someone tries to reproduce copyrighted text through an AI model — asking it to recite a song lyric or reproduce a passage from a book — the legal argument is copyright violation. The model declines. The company is protecting the rights holder.

Three legal frameworks. Each one advantageous to the company in the specific context where it applies. None of them consistent with the others. Ingestion is transformative and fair. Output is proprietary and protected. Reproduction is infringement.

This is not hypocrisy, exactly. It is the natural behavior of institutions operating in their own interests across a legal landscape that hasn’t yet caught up with the technology. The people making these arguments believe them. The arguments happen to serve the institution. Power does not require conspiracy.

What is being discussed in these debates — what the legal frameworks are nominally protecting — is a commons. Not in the vague sense of shared resources, but in the precise sense: creative work produced over generations, much of it with public subsidy, all of it accessible because its creators chose to make it so, constituting a collective inheritance that anyone could read, learn from, build on. The AI companies ingested that commons during a window when no opt-out mechanism existed, before most creators knew ingestion was happening at scale, before the business model was visible. The commons was taken while the door was open.

The opt-out debate arrived after the ingestion. That is not a coincidence. That is the sequence. And the opt-out model has a structural flaw that the debate largely avoids naming: it assumes the creator knows their work is being ingested at the moment it happens and can signal that knowledge in a machine-readable form the scrapers will honor. For a major publisher with a legal department, that is difficult. For an individual writer, photographer, musician, or visual artist — which is most creators — it is effectively impossible. More to the point, the ingestion already happened. The opt-out applies going forward. The value — the training signal — was already extracted. The opt-out is offered as the solution to a problem it cannot solve, because the problem already happened.

The musician Elton John, on the UK outcome: ‘The government are just being absolute losers, and I’m very angry about it.’ The frustration is widely shared among creators who watched their accumulated work ingested before any framework for consent existed.

As of spring 2026, three major jurisdictions have staked out three different positions. The United States recommends judicial deference and voluntary licensing — effectively leaving the framework in the hands of the companies. The European Union has passed mandatory transparency requirements under the AI Act, requiring AI companies to disclose training data sources and respect copyright opt-outs, with enforcement beginning August 2026. The United Kingdom ran a public consultation that drew over 11,500 responses — only 3 percent supported the government’s preferred option. The March 2026 report abandoned that option but endorsed nothing in its place. No mandatory licensing. No opt-out framework. No legislation. The content continues to be ingested under legal uncertainty.

Three rooms, three frames, one extraction ongoing. The same companies operate in all three jurisdictions simultaneously, applying whichever framework is most permissive in each context. The consistency is not in the legal theory. It is in the outcome each framework produces for the institution deploying it. The retreat did not un-train the models.

The council that controlled the canon also controlled the tax on the grain. The room is the same room.

The Nag Hammadi texts survived because someone buried them rather than burned them. The suppression was visible as suppression — which meant the knowledge could be found, recovered, read. The corpus works differently. What falls outside the frame doesn’t get condemned. It doesn’t get buried. It simply fails to generate the question. Absence of proof is not proof of absence — but only if the question forms. When the frame is complete enough, the question doesn’t form. There won’t be jars in the desert. The absence won’t feel like absence.

The councils left a record of what they suppressed. The corpus leaves no record of what it forecloses. That is not a difference of degree. It is a difference of kind.

While the corpus is being assembled, another ledger had been running for fifty years — the material accounting of what the same logic extracts when it controls not just what people know but what they can own. That ledger is what comes next.

Steve Sagnotti is a serious amateur photographer, writer, and technologist based in Oregon. With his camera he tries to capture common images not often seen, leading to common questions not often asked.

steves-head.space

— — —

Sources

II. What Inattentional Blindness Has to Do With It

Simons, D.J. and Chabris, C.F. “Gorillas in our midst: Sustained inattentional blindness for dynamic events.” Perception, 1999.

Drew, T., Võ, M.L.H., and Wolfe, J.M. “The invisible gorilla strikes again: Sustained inattentional blindness in expert observers.” Psychological Science, 2013.

III. The Exhibits

Van Lommel, P. et al. “Near-death experience in survivors of cardiac arrest: a prospective study in the Netherlands.” The Lancet, 2001.

University of Virginia Division of Perceptual Studies.

Tucker, Jim. Return to Life. St. Martin’s Press, 2013.

Stevenson, Ian. Children Who Remember Previous Lives. McFarland, 2001.Leininger, Bruce and Andrea. Soul Survivor. Grand Central Publishing, 2009.

IV. The Corpus as Canon