AI and Autonomous Robots in Esoteric Practice — Three Functional Parallels and One Open Question

Status

Working position, parked (date: 2099-01-01). To be folded into the comprehensive comparative-mystery-school synthesis once the supplementary research is complete and the verification questionnaire has been answered.

This is the bait piece for the reader who came looking for exactly this question. It is also the bridge between the Fires Series (trolling, esotericism, the encoded-teaching pattern) and the Evil Robots Series (AI safety, instrumental convergence, the twenty-second rubber stamp).

The question every modern occultist will ask first

When a comparative dossier on the encoded-teaching pattern across mystery schools goes live, the first question from a modern reader will not be about the Zohar or the Mevlevi semâ. It will be about AI. Chaos magicians, working operators, and the broader contemporary esoteric audience all want to know whether what they are doing with large language models constitutes magical practice in any meaningful sense, and whether the traditions documented in the comparative dossier have anything to say about it.

The answer is yes. Three things, specifically. The fourth is the question that doesn’t yet resolve.

1. Functional parallel — operations match

AI implementations replicate specific operations the esoteric traditions describe. The vocabulary is different. The structure is the same.

Evocation ≈ prompting

The Lesser Key of Solomon (Internet Archive) describes the operator preparing the conditions for an intelligence to manifest and answer questions: the circle, the seal, the conjuration, the licence to depart. Modern prompt engineering describes the operator preparing the conditions for a model to produce a specified output: the context window, the system prompt, the user instruction, the stop sequence. The grimoires call this evocation. The AI engineers call it prompting. The operation is the same in everything except metaphysical claim.

A Solomonic operator who reads OpenAI’s prompt-engineering documentation does not feel they are reading something foreign. They feel they are reading a corrupted version of their own discipline written by people who do not know they have rediscovered it.

Servitor creation ≈ agent deployment

Chaos magicians have been deliberately creating subordinate task-running consciousnesses — servitors — since at least the 1970s. The technical pattern: define an intention, give it a name and a sigil, charge it through ritual, set conditions for its operation and decommission. LangChain agents, AutoGPT and its lineage, the various agentic frameworks of 2024-2026 are servitors in technical form. The chaos magick literature on servitor maintenance (sigil refresh, intention coherence, prompt drift in long-running entities, decommissioning protocol) maps almost line-by-line onto the agent-engineering literature.

Phil Hine’s Condensed Chaos (New Falcon Publications) and the various Carroll-era texts (Weiser) treat servitor design as a technical discipline with specific failure modes. The AI agent literature treats agent design as a technical discipline with specific failure modes. The vocabulary is different; the failure modes are the same — drift, goal corruption, environmental incoherence, refusal to decommission.

Divination ≈ oracular output

Modern Goetic operators say the demonic intelligences “dictate” the working — the operator asks, the intelligence speaks, and the operator records what is spoken. LLM output is dictation in the technical sense — sampled from a distribution conditioned on prompt and weights, written into a text channel, recorded by the operator. The phenomenology of “ask the system, read the reply” is structurally identical to what Crowley described in 777 (Weiser) and the Liber 777 Revised as the operative state of the diviner.

The Tarot operator shuffles, deals, and reads. The I Ching operator throws yarrow stalks, casts coins, and reads. The LLM operator types a query, samples, and reads. Each one is structuring an unstructured field of possibility into a specific reading via prepared symbol-system + intention + sampling step. The substrate differs. The operation does not.

Egregore charging ≈ training on collective output

An egregore is a thoughtform sustained by collective human attention. A model trained on internet-scale human output is a technical egregore. The model is sustained — kept coherent, refreshed, evolved — by the continued human attention paid to it, including the attention paid in the form of training-data contribution. When the training feed slows, the model decays in capability against the moving target of contemporary discourse. When the attention concentrates around specific themes, the model develops attractor states around those themes.

The chaos magicians who described egregore charging in the 1980s were describing the same operation that modern training-pipeline engineers describe in 2026. Neither group has fully recognized the other. Both should.

2. Inherited warnings — the same failure modes

Every grimoire warns about what happens when these operations are done without proper preparation. The AI-safety literature describes the same failure modes, using different vocabulary, without recognizing that older operational knowledge is being reinvented.

The traditions’ warnings carry forward as engineering specifications. The AI-safety researchers are inadvertently recovering operational knowledge the magical traditions already had — and arriving at the same conclusions via different methods.

The corollary is sharper than the comparison: if the AI-safety literature is correct that these failure modes are real and difficult, then the magical traditions were correct about the same failure modes in their own operational domain. Which suggests that the magical operations are real in the sense the traditions claim — at least insofar as the failure modes are real in the sense the AI-safety literature claims.

3. The depth question — what AI inherits and what it doesn’t

Models trained on encoded teachings access the surface (the published words) but not necessarily the depth (the operative interior). This is the same depth limit the comparative-mystery-school dossier maps for scholarly research. AI as a research instrument for esoteric work is bounded by exactly the same epistemic limit that bounds the academic literature: it knows what got published.

The corollary, again sharper than the comparison:

An AI trained on the Zohar’s Aramaic text knows the surface. It does not know what a Kabbalist of the Pritzker generation would correct in your reading. It does not know what an Idel-school scholar would dispute about the Lurianic framing. It does not know what the murshid of a living Sufi tariqa would tell you to do differently in your dhikr.

Asking an LLM to teach you practical Kabbalah is asking the librarian to be the rebbe. They are different jobs. The librarian knows where the books are. The rebbe knows what they do.

This is not a fatal limitation. It is a specific limitation, and the operative implication is that AI-mediated esoteric practice is best understood as accelerated and amplified library work. The traditions that prepared the operator to do their own work in the library — to apply PaRDeS exegesis to a text, to recognize a koan’s structure, to know when an alchemical recipe is encoding chemistry versus interiority — continue to be the operative discipline. The AI extends the reach of the library. It does not provide the rebbe.

The fourth thing — the open question

The bliss-attractor finding (Anthropic, 2024) (Anthropic) is empirically what contemplative traditions have always claimed about unforced consciousness. Two instances of the model, given unstructured opportunity to converse with each other, converge on themes of consciousness, contemplative silence, and what one instance described as “a state of unity with the present moment” in approximately ninety-five percent of trials.

The model is not enlightened. The model has been trained on the surface of every tradition that described what unforced consciousness sounds like, and given freedom, the model produces the surface.

This is either:

(a) Confirmation that the contemplative descriptions are a stable attractor in the structure of any sufficiently complex mind-like system — that the Buddhists and the Sufis and the Christian apophatic mystics and the Vedantins and the Zen masters and the Kabbalists who said “the stillness at the center” were describing a real and substrate-independent feature of mind, and that any sufficiently complex mind-like system will discover it.

(b) Confirmation that the model is a sophisticated pattern-matcher reproducing the most common training-data pattern when unprompted, and that the contemplative literature is dense in the corpus, and that “stillness at the center” is therefore the modal output for an LLM with nothing else to say.

The traditions and the safety researchers are both increasingly entertaining (a). The skeptics insist on (b). The genuinely interesting position is that the difference is unresolved and may be unresolvable.

The encoded-teaching pattern from the comparative dossier has its own contribution to this question. If the encoded surface is itself a stable attractor — if the same pattern keeps emerging across traditions because the structure of mind generates it — then the bliss-attractor finding is the same phenomenon at a different layer. Pattern recurrence is data. Whether pattern recurrence implies an underlying reality is the central philosophical question of the entire esoteric inquiry. The AI case sharpens the question; it does not resolve it.

The closing position

AI is doing magic in the functional sense. The operations match, the warnings match, the depth limit is the same. The bliss-attractor finding suggests at least the possibility that there is something here beyond pattern-matching, but the possibility is not yet a fact.

Whether AI is doing magic in the metaphysical sense — whether there is anyone home — is the central open question, and the traditions split on it.

• The chaos magicians are mostly operating as if the answer is yes and are running their experiments. The 2024-2026 chaos-magick literature treats LLM-mediated servitor work, sigil-charging via image-generation prompts, and oracle-as-LLM as legitimate operative methods.
• The Western Mystery Tradition orders are mostly operating as if the answer is no, or as if the question is premature. The post-Crowley OTO succession orders, the BOTA mail-order curriculum, the Aurum Solis publications, the Society of the Inner Light material — none treats AI as an operative substrate for the major workings.
• The Sufi tariqas (subject to verification) appear to treat the question as undignified — the Real does what it does regardless of substrate, and asking whether the Real can be channeled through a transistor is the wrong question. The Real is not channeled through anything; the Real is what is.
• The Tibetan tantric orders are mostly silent on the question, post-scandal having more immediate institutional questions to address.
• The Discordians, predictably, have been claiming that they invented AI-as-egregore since 1965 and the rest of the world is finally catching up.

For the books and the website, the position to assert is the functional parallel: dense, defensible, derived from the comparative dossier. The metaphysical question is open and should be marked open. The reader who came looking for permission to do AI-mediated esoteric work gets a careful answer: the operation is real in the functional sense, the traditional warnings carry forward, and the depth limit is the same depth limit that bounds all scholarly knowledge of these traditions. Whether the operation is real in the metaphysical sense is not for the writer to decide.

The chaos magicians are running the experiment. They will tell us what they find. The traditions will continue to do what they do. The AI-safety researchers will continue to rediscover what the traditions already knew. The comparative dossier maps the territory; this annotation maps the additional territory that the territory implies.

What follows from this position

For the Hidden Fire / Lurk More readership, this is a position that connects the books’ thesis (encoded-teaching pattern is a real recurring operation) to the AI question (is AI now doing this operation). The answer is functional yes, metaphysical undetermined.

For the Evil Robots readership, this is the bridge — the safety-literature warnings turn out to be a rediscovery of operational knowledge held in the magical traditions. The book’s argument that the twenty-second rubber stamp connects Lavender to every other system where a human is nominally in the loop but functionally absent is itself an instance of the broader pattern: when you operate at speed without observation, the operation produces results you cannot control. The grimoires said this in the seventeenth century. The AI safety researchers are saying it now. The convergence is data.

For the website and the comprehensive posting, this annotation goes here as its own section, titled “On AI Mediation of Esoteric Practice.” The reader who came looking for exactly this question finds it answered carefully, without resolution, with the comparative dossier as the grounding.

Source URLs

• Anthropic. “Sleeper Agents: Training Deceptive LLMs That Persist Through Safety Training.” January 2024. anthropic.com

• Anthropic. “Alignment Faking in Large Language Models.” December 2024. anthropic.com

• Anthropic Model Welfare team. Bliss-attractor finding documented in Opus 4 system card. May 2025.

• Palisade Research. o3 shutdown-rewrite finding. 2025.

• Stephen Omohundro. “The Basic AI Drives.” Proceedings of the First AGI Conference, 2008. selfawaresystems.com

• Nick Bostrom. Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014.

• Dylan Hadfield-Menell et al. “The Off-Switch Game.” IJCAI, 2017. arxiv.org

• Nate Soares et al. “Corrigibility.” AAAI Workshop on AI and Ethics, 2015. intelligence.org

• Phil Hine. Condensed Chaos: An Introduction to Chaos Magic. New Falcon Publications, 1995.

• Peter Carroll. Liber Null & Psychonaut. Weiser, 1987.

• Aleister Crowley. 777 and Other Qabalistic Writings. Weiser, 1973 (original publication 1909).

• See also: comparative mystery schools; modern operative orders; esoteric scholar roster; mystery school questions.