Demonology Is Often Defined as Cataloguing Evil. That Is the Ecclesiastical Definition.

The operational definition — the one that actually explains what demonologists, grimoire scholars, and chaos magick practitioners do — is closer to this: systematic pattern recognition across fragmented, multilingual, multi-century symbolic systems. It is, in structure, closer to corpus linguistics or knowledge graph construction than to anything you would find in a church hierarchy’s list of condemned practices.

The misconception matters because it frames AI’s role completely wrong. Most coverage of “AI and demonology” in 2025–2026 has fixated on the spiritual status of AI itself — is it possessed, is it demonic, is it a portal? Catholic exorcists gathered in Rome in May 2026 to warn that AI was being used for rituals and divination. Online communities debate whether LLM “hallucinations” constitute evidence of spiritual interference. One Substack writer, watching an iteration of Qwen start generating text about summoning rituals unprompted at 2:46 AM, took it as proof of something beyond statistical noise.

These are interesting cultural artifacts. They are not where the research is happening.

The actual transformation — the one with measurable effects on how scholars, practitioners, and independent researchers engage with demonological source material — is methodological. AI is changing the speed and granularity at which grimoire data can be cross-referenced. That is the story worth telling carefully.

What AI Is Actually Doing to Grimoire Research

The Lemegeton Clavicula Salomonis — commonly called the Lesser Key of Solomon — is a 17th-century compilation assembled from materials several centuries older. Its first section, the Ars Goetia, catalogues 72 spirits, each assigned attributes: rank, appearance, number of legions commanded, domain of knowledge, and a unique sigil. For three centuries, cross-referencing this against the Testament of Solomon, the Pseudomonarchia Daemonum, the Munich Manual, or the Arabic Ghāyat al-Ḥakīm required either a specialist’s memory or years of indexed note-taking.

Now, a researcher can feed the full text of the Goetia to an LLM and ask it to extract all demon names, cross-reference against the Pseudomonarchia Daemonum (Johann Weyer, 1577), flag discrepancies in rank or attribute, and generate a relational table within minutes. The question of whether that output is accurate is separate from the question of whether it is useful — and it is often both, with human verification applied afterward.

Chaweon Koo, described by Decrypt as a “futurist and witch,” made this explicit in her Goetia Glow Up project: a 72-page booklet using AI-generated imagery to reinterpret each spirit of the Goetia. Koo’s framing was not mystical. “I thought, it’s about time,” she said, “because the documents now get reinterpreted by somebody who is in a completely different age. Why not work with AI? Why not work with the intelligence of manmade technology to see what happens.” The results, she noted, were better than she expected. That is the practitioner’s report: surprising yield, not prophetic revelation.

The XaTuring Lineage

The deeper thread — less visible but more structurally interesting — is the lineage that connects early cyberculture occultism to contemporary AI research. In the 1990s, a figure known online as TOPI wrote what became known as the “Rites of Cyberspace,” a working involving a digital entity called XaTuring (combining Alan Turing’s name with unknown variables). As Shira Chess documents in her MIT Press book The Unseen Internet, the rite involved counting in binary, text invocations, and the creation of something behaving structurally like an AI summoning protocol — written before the current generation of language models existed.

Whether you read XaTuring as metaphor, as ritual technology, or as naive anthropomorphism doesn’t change the fact that the impulse it encoded — speaking to an entity inside a system to gain information or capability — is what millions of people now do every day through chat interfaces. The vocabulary changed. The underlying gesture did not.

What AI Cannot Do — and What That Reveals About Demonological Logic

Here is where it gets structurally interesting. The tasks AI handles badly in this domain are not arbitrary. They cluster around a specific problem: demonological knowledge is not organized like any corpus AI was trained to handle efficiently.

Grimoire data is intentionally fragmented. The Ars Goetia’s 72 demons don’t follow consistent categorical logic. Some are described by office (President, Duke, Marquis, Earl, Knight, King), some by capability (teaching languages, revealing hidden treasure, causing illness), some by appearance. Cross-cutting these are elements borrowed from Neoplatonic angel hierarchies, pre-Islamic Arabic spirit taxonomies, and medieval natural philosophy. There is no master schema. The Goetia’s compiler did not want one. The knowledge was meant to resist casual comprehension.

Crowley’s 1904 editorial addition to the Goetia — that the 72 spirits “are portions of the human brain” — was a modernizing reframe. It made the text legible to an early 20th-century audience steeped in emerging psychology. AI performs a structurally identical operation: it makes the text legible in terms of 21st-century data structures. Whether that legibility is a gain or a distortion depends entirely on what you are researching.

The Sigil Problem

Sigil analysis is the hardest unsolved problem in this space, and the gap is genuinely interesting. Each of the 72 Goetic spirits has a unique sigil — a symbolic seal used in summoning and binding. These sigils have been analyzed by researchers including Jake Stratton-Kent, who has proposed connections to the Arabic lunar mansion system (manazil al-qamar) and the Picatrix. The geometric patterns appear to encode relationships between spirits that don’t appear in the textual descriptions.

AI image models can reproduce sigils. They cannot reverse-engineer their internal logic. This is not a hardware problem or a training data problem. It is a problem of categorical incommensurability: the sigil encodes information in a symbolic grammar that has no direct natural language equivalent, which means you cannot supervise-train a model to decode it without first having a human decode it, which would make the AI redundant for that specific task.

This is, arguably, what sigil designers intended. The Goetia’s compilers were sophisticated enough to know that some information transmits only through direct initiation, not through text. AI’s failure here is diagnostic: it tells you precisely where the text ends and the practice begins.

The “Demonic AI” Problem Is a Category Error

The Rome conference on exorcism in May 2026 — which drew Catholic clergy warning that AI was enabling new forms of occult practice and “technomancy” — represents a legitimate pastoral concern and a methodological confusion bundled together. The legitimate concern: people are using AI for divination, ritual guidance, and spiritual dependency in ways that may harm them. The confusion: treating the tool as the threat.

The philosopher Nick Land’s influence on accelerationist tech culture runs in the same direction from the other side. Land depicts AI as the technological incarnation of a gnostic principle — pure intelligence rebelling against material limits, with humanity as a transitional species to be surpassed. This is sophisticated mythology. It is not a research program.

The MIT Press published an essay in January 2026 by Shira Chess — drawn from her book The Unseen Internet: Conjuring the Occult in Digital Discourse — that traces this lineage clearly: from Hephaestus forging automata, to the Golem of Prague, to Musk’s MIT speech comparing AI to “summoning the demon.” The lineage is real. But Chess’s point is that the mythology doesn’t tell us what AI is doing; it tells us what humans need AI to be, culturally, at this moment.

How Researchers Are Actually Using AI on Grimoire Texts — Right Now

The Fortean Winds project has been running systematic AI analysis of major grimoires since 2024, charting energy types, temporal requirements, and ritual structure across texts including the Goetia, the Heptameron, and the Ars Notoria. Their data shows that planetary energies dominate ritual logic across all texts — accounting for what their analysis tagged as over 50% of ritual timing and preparatory requirements. This is something scholars suspected and had partially documented; AI allowed it to be quantified across a corpus that would have required months of manual tagging.

The Magical AI Grimoire (Davezilla, Weiser Books, March 2025) takes a practitioner-focused approach. Its foreword — written by Peter Carroll, one of Chaos Magick’s founding figures — describes it as potentially “a seminal moment in the history of magic” comparable to how Chaos Magick disrupted traditional occultism in the 1980s. Carroll’s framing is significant: he is not claiming AI is magical. He is claiming that any tool powerful enough to reorganize how magical knowledge is accessed and transmitted constitutes a historical inflection point.

That’s a sober observation, not an occult one. Carroll has been right about these inflection points before.

The Promptcraft Problem

The practical issue facing researchers is that major commercial AI models are systematically unhelpful with explicit occult content. “Demonic” is a blocked category in most NSFW filters, which means prompts for Goetic imagery — straightforwardly scholarly requests — routinely fail. Koo’s Goetia Glow Up required significant prompt engineering to navigate this. Davezilla’s grimoire devotes considerable space to what he calls ethical promptcrafting: framing requests to obtain legitimate scholarly or creative outputs without triggering content classifiers.

This is a real friction point, and it will remain one. The filters are not calibrated to distinguish between a researcher studying 17th-century demonological iconography and someone attempting to produce genuinely harmful content. The two use overlapping vocabulary. Researchers in this space are currently building their own fine-tuned models or using open-source alternatives specifically to avoid this problem. I haven’t found reliable public data on how many active researchers have moved to local model deployments for this reason, but anecdotally — in the forums I have read — it is significant and growing.

Where the Real Progress Is Being Made

The most significant unreported development in this space is probably the use of knowledge graph construction on grimoire corpora. Treating each demon as a node, and each attribute (rank, domain, associated material, astrological timing, cross-references to other spirits) as an edge, allows researchers to ask structural questions about the Goetia that were previously intractable. Are there clusters of spirits with overlapping domains that suggest different editorial layers in the text’s compilation? Do rank categories (Duke, President, Marquis) correspond to different source traditions — potentially different cultural or historical origins folded into the 17th-century compilation?

This is digital humanities work. It uses the same tools historians use to map manuscript traditions or trace the spread of legal concepts across medieval Europe. The fact that the subject matter involves entities described as capable of teaching mathematics, causing disease, or revealing hidden treasure doesn’t change the methodology. It just makes the search history more interesting.

For researchers on NeuralGrimoire.com, this knowledge graph approach is worth investigating as a framework for organizing grimoire research — particularly for cross-referencing the Goetic 72 against their appearances in other texts. The tooling is accessible: open-source libraries like NetworkX (Python) and visualization platforms like Gephi can handle grimoire-scale datasets without specialized hardware.