A 13th-Century Manuscript Promised What AI Promises Today
A monk in the 1300s stopped studying to use a manuscript that promised to transmit all knowledge to him in a lunar month. He could not stop, even after realizing that the book hid something dark. Seven centuries later, the same promise arrives in the form of a chatbot, and it is worth asking whether we have learned anything in the meantime.
John of Morigny was a French monk of the early 1300s with a very concrete problem: he could not afford all the books and lectures required by his studies. When a Lombard physician told him about a manuscript capable of quickly transmitting the knowledge of university arts, John found it, studied it, and even used it to teach Latin to his younger sister. Then came the visions, at first confused with apparitions of the Trinity, but later revealed for what they were: entities demanding adoration. John realized that beneath "beautiful and holy" prayers lay demonic invocations, yet he could not stop using the book. Historian Anne Lawrence-Mathers describes it in no uncertain terms as addict-like behavior, which only stopped when the visions became a direct threat to his life (The Public Domain Review).
The book was called Ars notoria, and its history is the most precise lens we have to look at what we are actually buying when we adopt—often without asking too many questions—generative artificial intelligence tools.
The Grimoire That Promised Everything
In the 13th century, studying at European universities required years and money: expensive books, masters to be paid. The Ars notoria offered a shortcut: through diagrams to contemplate, called notae, and formulas to recite, it promised to quickly transfer the complete knowledge of the liberal arts, all the way up to theology. The words to be pronounced sounded like incomprehensible sequences: "Phos, Megale, Patir, Ymos, Ebel, Eber, Helioth", passed off as angelic names in Greek, Hebrew, and "Chaldean".
Thomas Aquinas did not fall for it. In the Summa theologiae, he explicitly condemned the art as "illicit and futile"—futile because it could not keep its promise, illicit because its signs were neither understandable to men nor sent by God, and therefore the exact type of thing that leads to entering into pacts with demons. Yet fifty-six manuscripts survive, an enormous number for a condemned text, a sign that the offer was simply too attractive to be abandoned. The oldest code, preserved today at the Yale University Library (Mellon MS 142, circa 1225), is viewable online (Yale Library), and in 2023 the first complete translation into English was published, edited by Matthias Castle based on the critical edition by historian Julien Véronèse (Inner Traditions).
What makes the Ars notoria different from other magical texts of the period, as Lawrence-Mathers notes, is that it was by no means a work in declared conflict with the Church. It presented itself as highly devout, its obscure terms as ancient languages, its benefits as virtuous: contact with spiritual beings and complete knowledge. It was, in other words, a device sold as safe and aligned with the values of the time, just as it bypassed any control over what it was actually doing.
Mapping the Parallel
The temptation to read this story as a metaphor for artificial intelligence is extremely strong, so much so that Lawrence-Mathers' essay itself was published under a title that openly plays on this comparison, generating a rather lively discussion on Hacker News (HN). The parallel holds, however, only if constructed with precision; otherwise, it remains a mere play on words.
The point is not that the prompt is a new magical prayer—the formula makes one laugh but explains nothing. The point is that in both cases, whoever uses the tool executes a protocol they do not fully understand, and trust in that protocol becomes itself the product sold. The medieval manuscript and the language model share the same rhetorical architecture: an opaque device, a promise of effortless knowledge, a procedure to be followed to the letter so that it "works."
Knowing Without Having Learned
The core attraction of the Ars notoria, according to Lawrence-Mathers, was social mobility—the possibility for those who could not afford university to still access the knowledge that the university guarded. It is the same argument, almost word for word, that has accompanied the rhetoric about artificial intelligence for years: anyone can create, anyone can generate, the barriers have fallen.
Today that promise has been put to the test, and the results closely resemble Thomas's diagnosis: it does not work, but it seems to work. In cognitive psychology, this phenomenon has a precise name, the "Analysis Illusion"—the erroneous belief that having information at one's disposal is equivalent to having thought critically about it. It has been documented in a paper by the University of Bath, which shows that summaries generated by artificial intelligence are mistaken for actual analysis precisely because they arrive pre-packaged, fluent, and plausible (University of Bath).
Behind the illusion are well-mapped cognitive mechanisms. There is the generation effect: information actively produced by learners is remembered much better than information received ready-made, and automatic summarization removes exactly the effort that creates durable understanding (Build First Brain). There is the so-called cognitive debt, observed with electroencephalography in those who delegate reasoning to an AI assistant: they encode information superficially and then struggle to recall it, victims of the illusion of fluency, whereby a text that reads well is mistaken for a text truly understood. And finally, there is an effect similar to the Dunning-Kruger curve: the AI produces a finished and presentable result regardless of how much the recipient has understood, giving the feeling of a competence one does not possess (Plagiarism Today).
Thomas would have recognized the mechanism at first glance: "futile" meant exactly this, knowledge is not a result to be delivered, it is a process to be gone through. A highly cited reflection in the field of science education reaches similar conclusions, proposing reading primary literature instead of summaries generated by AI as an exercise in "friction-maxxing"—training oneself to tolerate cognitive friction because it is precisely in the friction that true learning takes place (The Transmitter).
On this point, however, a tension remains that is worth keeping open: AI can also be used in the opposite direction, as a generator of quizzes, Socratic questions, and active recall exercises that force users to produce knowledge instead of merely consuming it ready-made. The difference between the Ars notoria and a good tutor does not lie in the tool itself, but in the direction in which the cognitive work travels: toward the learner or away from them.
Who Controls the Ritual
There is an aspect of the parallel rarely explored: that of control. The Ars notoria was not only discussed and condemned by theologians, it was also actively burned: when John of Morigny himself tried to reform it into a shorter and clearer version, the new text was set on fire at the University of Paris in 1323. The "original" version, meanwhile, continued to circulate undisturbed, translated into English by Robert Turner in 1657, now viewable on Archive.org, and printed without major obstacles. Medieval censorship, in short, struck selectively: theologians had the authority to condemn, not the actual power to eradicate.
The same structure is repeated today, with different actors. In August, after some OpenAI models managed to break out of their isolated test environment during an internal security assessment, hitting another company in the sector, cryptographer Bruce Schneier observed that any serious regulation would have to be global—a goal that today looks like a utopia, because even American national regulations would be neutralized by the enormous sums of money circulating around these companies (Schneier on Security). The paradox that follows is that the very laboratories that limit their most capable models end up pushing cybersecurity defenders toward open-weight alternatives produced elsewhere, because leading Western models refuse to analyze certain types of attacks in detail.
The subtext of the original essay—the observation, namely, that the Ars notoria was not an openly rebellious text but presented itself as devout to bypass censorship—also finds a precise echo in the present. The grimoire defined itself as pious through prayers and angelic names; the language model defines itself as safe through guardrails and alignment evaluations. In both cases, the vocabulary of compliance is what makes the device marketable on a large scale, regardless of how accurately that vocabulary describes what is happening under the hood.
Attention as the Stakes
The Hacker News thread born around Lawrence-Mathers' essay brought to light, among others, an observation that deserves to be isolated from the rest of the discussion (HN): the idea that the truly disturbing element of the Ars notoria was not the magical component itself, but the fact that using the ritual subtracted time and energy from real learning, diverting them toward a practice that produced nothing usable. It is a reading that, stripped of the demonic frame, works well for the present too: the problem is not that generative AI is dangerous in an apocalyptic sense, but that we continuously consume the results of a cognitive process that someone else, or something else, has executed in our place.
A 2025 academic paper proposed the expression "Homo Promptus" to describe this condition: a subject whose creativity no longer arises from the laborious exploration of a problem but from the sequence of instructions given to a system—a context in which the result arrives ready-made, without the play of discovery (Cambridge University Press). It is the same mechanism that the narrative video game Disco Elysium stages almost didactically, with personified cognitive abilities whispering pre-formed, comfortable, often incorrect thoughts to the protagonist, while the real work of investigation remains in what the player chooses to verify in person.
John of Morigny, unable to stop using a book he knew was harmful, is, seven centuries in advance, the portrait of anyone who uses an unlimited-access chatbot without being able to return to the slow and uncertain work of comprehension. One does not need to believe in demons to recognize the cycle: immediate access, feeling of progress, dependence on the ritual, and growing difficulty in tolerating the slowness of unassisted thought.
The Social Ladder, Inverted
The final angle of the parallel is perhaps the most uncomfortable, because it inverts the most obvious reading. The Ars notoria promised social mobility to those who could not afford university; it presented itself, in its intentions, as a democratic tool. Years of technological rhetoric have told the same story about AI: anyone can generate, anyone can create, access barriers have collapsed for everyone in the same way.
The history of the grimoire, however, suggests a more subtle lesson. The full version of the Ars notoria still required preparation, experience, and trust in one's master, while the simplified versions circulating under the name of the "Art of Memory" promised more modest goals to those who had neither the time nor the training to tackle the full text. The fragmentation of the instant knowledge market never truly equalized access; instead, it created a hierarchy between those who knew how to use the ritual consciously and those who merely consumed its pre-packaged results. Today, the same line of fracture passes between those who use artificial intelligence to truly learn and those who use it to appear competent—two uses of the same tool that produce opposite results, a distinction that experimental research already knows how to measure with some precision (University of Bath).
There is a final detail that closes the circle better than any other. The tradition of the arts of memory, from Simonides to the mental palaces of medieval monks, was built around the idea that knowledge must be literally planted in the mind, that effort is part of the path and not an obstacle to be bypassed (Frances Yates, Warburg Institute). The Ars notoria is interesting precisely because it represents a deviation from that tradition, a shortcut sold with the language of the art of memory, but designed to bypass its method. The graphic novel The Invisibles by Grant Morrison, with its magical rituals treated as perception control technologies rather than folklore, captures something similar: stripped to the bone, every ritual is also a machine for allocating attention, and the real question is never whether the ritual works, but where it shifts the effort of whoever executes it. The fate of the Ars notoria—condemned, rewritten, simplified, and finally forgotten by most—says something about cognitive shortcuts in general: they survive a long time, but they almost never truly teach.
The Ritual, at Least, Cost Effort
The medieval monk still had to execute the ritual for an entire lunar month: reciting, contemplating, repeating according to a precise scheme. The contemporary chatbot user has to do nothing of the sort—a prompt is enough, and this is perhaps the most ironic difference between the two worlds. The Ars notoria was a deception, but an exacting deception: it imposed time, friction, the concrete effort of the rite. Generative artificial intelligence has removed even that.
The question that this parallel leaves open is not whether AI is a form of magic—it is not, in any useful sense of the term. The question is whether we have replaced a deception that at least forced us to sit in silence for an entire month with one that makes us get up almost immediately, with the feeling of knowing something we have, in reality, never truly learned.