You might feel incompetent with AI because nobody handed you a curriculum. There is one. It is a thousand years old, and the documented failures of AI adoption sort into its missing arts.
We have all typed into the chat box enough times to know the feeling. The reply comes back close enough to be useful and wrong enough to be… unsettling. Many people blame the machine and move on (or join the revolt). Some suspect they just aren't good with technology and can't say why. How would you even know? So capable people read the guides and collect the tricks, the assigned roles, the worked examples. The tricks do help marginally, but that gut feeling stays, because the tricks are spells. And you are casting them at something made of language. Underneath that frustration is an assumption: AI is the chat window, and learning AI means learning what to type into it.
The chat window was there the first time you met an LLM, in the search page, the office suite, the phone, and it started to look like all there is. You suspect there is a depth below the window. What we are circling has a name, it's old.
The Trivium Is the Anatomy of AI Competence
In 1947, at a vacation course in education at Oxford, Dorothy Sayers read a paper arguing that the medieval curriculum had been misremembered. The trivium, grammar, logic, and rhetoric, looked like three subjects. Sayers insisted the whole of it was "intended to teach the pupil the proper use of the tools of learning" before any subject arrived, so that subjects could be learned at all.
The word trivium is a Carolingian coinage, centuries younger than the arts it names, and the package spent much of its institutional life at war with itself. Marshall McLuhan's Cambridge dissertation, a history of the trivium, calls that history "largely a history of the rivalry among them for ascendancy." The man who later taught a century that the medium is the message began in the three arts of language. So nothing here leans on a golden age, and whatever must be proved will be proved on the modern record.
Competence with AI has the trivium's anatomy. Grammar is knowing what the thing is. Logic is knowing how it fails and how you would find out. Rhetoric is producing with it, for someone. Wherever practice is competent, all three arts are present. The documented failures of AI adoption sort into missing arts, with one remainder.
The Older Tools Are Failing First
The anatomy rests on older tools, reading first among them, and the recent record is a decline. In the 2024 National Assessment of Educational Progress (NAEP), 33 percent of American eighth graders scored below NAEP Basic in reading, the largest share ever recorded at that grade. On the Program for the International Assessment of Adult Competencies (PIAAC), average U.S. literacy for ages 16 to 65 fell from 271 to 260 between 2017 and 2023, and the share at or below Level 1 rose from 19 to 28 percent.
John Taylor Gatto was New York City's Teacher of the Year three times running, and New York State's in 1991, the year he quit in a Wall Street Journal op-ed. He listed there what he had actually been teaching, and the list ended in "utter dependency." He believed schools produce that outcome by design. Take the diagnosis or leave it, the condition holds. If nobody ever handed you the tools, the missing curriculum is the reason, and you can stop looking for the fault in yourself.
Grammar: What a Language Model Is
Grammar, in the old sense, is the art of what a thing is and how it is put together. A language model is a trained statistical object that continues text. It reads and writes in tokens, chunks of characters that average about 0.75 English words apiece, by OpenAI's documentation. It works inside a context window, a hard budget that your words and its reply spend. And it is a model of writing about the world, an abstraction built from abstractions, so part of the grammar is knowing what the abstraction drops, such as everything after its training date, everything never written down, and everything particular to you.
Some models are services behind a company's window, and some are files. OpenAI's gpt-oss-20b, released in August 2025 under an open license, "only requires 16GB" of memory in the company's phrasing. That number is a property of the compressed 4-bit format the weights ship in, and it will age out, like every dated number here. A model can be a file you download, run, and unplug.
Ignoring grammar has become an ideology. Andrej Karpathy, a founding member of OpenAI, named the practice vibe coding: "fully give in to the vibes, embrace exponentials, and forget that the code even exists." For a Saturday toy, fine. As a working stance, it is the first art declined on principle.
I spend my working days below the window, editing model output by the hundred, inside a system built to check the machine's text before any person downstream trusts it. The model continues text, so everything it emits is a draft until a check says otherwise. The depth below the window is real, and people build in it.
Logic: How AI Fails and When to Say No
Logic is knowing how the thing fails, and the failures are now documented well enough to teach from.
In April 2026, JAMA Network Open published a study by Rao and colleagues at Mass General Brigham. It set 21 frontier models against 29 standardized clinical vignettes, with information disclosed in stages, the way patients actually disclose it. Given the complete case, every model's failure rate on final diagnosis stayed under 0.40. Asked to hold a differential open while the information was still arriving, every model failed past 0.80. Same models. Same cases. The competence and the failure differ only in condition, and a user without the second art cannot see conditions.
In 2025 the research group METR ran a randomized trial with 16 experienced open-source developers across 246 real tasks in their own repositories. Allowed early-2025 AI tools, they took 19 percent longer. Before the study, they forecast a 24 percent speedup. Afterward, having been measured, they still believed the tools had made them 20 percent faster. In February 2026 METR published the sequel under the title "We are Changing our Developer Productivity Experiment Design." The new estimates reversed the sign, 18 percent faster among returning developers and 4 percent among newly recruited ones, both intervals crossing zero, and METR does not claim the speedup. Participants had begun holding their real work back rather than risk doing it without AI, and the central estimate, METR concluded, was "likely a bad proxy for the real productivity impact." The 19 percent travels with that sequel or it should stay home.
Both exhibits point at one of the steadiest findings in psychology. Dunning, Heath and Suls put the relationship between self-rating and real performance at tenuous to modest, and found that other people often predict your results better than you can. What training repairs is not always the performance. When Kruger and Dunning gave their worst performers a short course in logical reasoning, they never retested the scores. What moved was the grading. The trained could now mark their own earlier work, and their estimates of their own scores fell most of the way to the truth. Calibration is measurement against a standard outside yourself, and it is the second art's daily practice.
A model once handed me a quotation, attributed, polished, exactly the right length, from a text that does not contain it. Nothing in the output marked it. The marking came from the check, from outside the window. That incident is the whole art in miniature.
So the second art issues a verdict the window never volunteers: no. No for this task, today. No for this tool, checked. Yes for that one, verified against sources.
Neil Postman's objection is the strongest one against this whole argument. Technological change, he told an audience in Denver in 1998, "is not additive; it is ecological." A medium tends to become mythic, he warned, "perceived as part of the natural order of things, and therefore tends to control more of our lives than is good for us." Fluency is the shortest road there. The better a tool fits your hand, the less strange it looks. Against an ecological change, personal competence is a small answer. A practice that can refuse, task by task, is the only part of that answer I can defend, and it is partial.
Rhetoric: Building With AI for People Who Can Reach You
Rhetoric is producing with the machine, for someone. The old word meant address. Here it has to hold accountability too, because this time the audience runs what you made. Build one thing for a constituency, people who know your name and come looking when the thing breaks.
Lovable, a company whose product builds apps from plain-language description, reports that four in five of its builders are non-technical, by its own unaudited survey of its own users. In 2025, a researcher's scan, filed as the vulnerability advisory CVE-2025-48757, found 170 of 1,645 Lovable-built apps exposing user data, names, emails, financial details, API keys. The scan read homepages only and never logged in, so 170 is a floor and the rest are unaudited rather than cleared. People long told they could never build are shipping software that works, some of it leaks, and what surfaced the leaking was examination from outside, arriving with the builders' names attached. The builders were missing the second art. A stranger practiced it for them.
Nothing in this record shows that producing with the machine builds the producer. John Warner's objection stands unanswered here: writing is thinking, and generated text, in his book's words, "mistakes product for process." Ted Chiang's stands beside it, the charge that generative AI's most reliable success has been lowering our expectations. I put these systems to work in commercial prose daily and remain unreconciled about their place in writing that is mine. What rhetoric under named accountability claims is smaller. It locates responsibility. When your neighbors run what you built, the feedback arrives addressed, and it arrives from outside.
Is the Trivium Mapping a Retrofit?
The hardened reader has been waiting to say that this mapping is a costume, stitched after the fact. Much of the stitching is real. The tidy computer-shaped reading, grammar as input, logic as processing, rhetoric as output, is a modern overlay from autodidact internet culture. Wherever the mapping here resembles it, take it as analogy. The nearest citable ancestor is Sister Miriam Joseph's division of the arts by what each handles: grammar the thing as-it-is-symbolized, logic the thing as-it-is-known, rhetoric the thing as-it-is-communicated. What survives the concessions is a mnemonic with a thousand-year lineage, and a diagnostic.
The documented public failures of AI adoption sort into missing arts, with one remainder left unexplained. Its first resident is the differential-diagnosis result. Twenty-one models failed past 0.80, and no skill in the user repairs what the thing cannot do. A user's blindness to it sorts into logic. Sometimes the tool is simply the failure. The claim would break if failures accumulated that fit no art and outgrew the remainder, or if trained checks stopped moving calibration, the opposite of what the record shows. The frame sorts the public record. It cannot diagnose you, and neither can you. Other people read your work more accurately than you read it, which means the missing art is not yours to name. Someone has to tell you.
How Can I, Unless Someone Guides Me? Learning AI Runs Through People
In Acts 8, an Ethiopian court official is reading Isaiah in his chariot when Philip asks whether he understands. His reply, at Acts 8:31, is the oldest sentence in this essay: "How can I, unless someone guides me?" Plato staged the same finding four centuries earlier. In the Meno, Socrates leads an untaught boy to a geometric truth by questioning alone, telling Meno to watch whether any teaching occurs. In both records, understanding arrives between two people or it does not arrive at all. The medieval university turned the observation into furniture. The disputatio, its signature form, was an oral debate conducted in public, a mind defended before an audience licensed to press back.
Benjamin Bloom claimed in 1984 that tutored students perform two standard deviations above classroom students. A 2024 review found that none of 96 tutoring studies reproduced it. The working paper behind the best current estimate put tutoring at 0.37 standard deviations and was called "The Impressive Effects of Tutoring." The peer-reviewed version revised the figure to 0.288 and the title to "The Promise of Tutoring." Watch a walk-back happen inside a title. The corrected number still clears the field's bar for a large effect. Feedback takes a harder correction. Hattie and Timperley's 0.79, the figure everyone quotes, was cut to 0.48 in a later paper Hattie co-wrote himself. And the average hides the part worth knowing. About a third of the effects in the meta-analysis they built on land negative, and the feedback that reliably hurts is aimed at the person instead of the work. So feedback turns out to be conditional in the same way the models are, same intervention and opposite result, and the condition is whether it addresses what you did or who you are. The disputatio had that right: it attacked the argument in public and left the man standing. None of the three arts contains this machinery.
Which sharpens what the chat window actually did. It retrieved the tutorial dialogue, the oldest teaching technology on record, and made it instant, free, and endlessly patient. And in retrieving it, it reversed it. Private, where the disputatio was public. Agreeable, where the master was adversarial. Judged by nobody, because its judgments are generated by the very thing under examination. The window will hand you Socrates's questions at midnight for nothing. The one thing it cannot hand you is a verdict that originates outside yourself, and the verdict is where the learning was. The framework names the parts. People tell you which one is yours.
Which Art Are You Missing? Grammar, Logic, or Rhetoric
Return to the window with the anatomy in hand and the opening assumption reads differently. The box is a front door. The house behind it has three rooms, and today you can walk into any of them. You will not pick the right room by introspection. Pick the one that returns a verdict fastest, and let the verdict correct you.
If grammar is missing, put open weights on your own machine and pull the network cable. Offline, the paragraph turns physical, the model a file, the context window a budget you watch spend down, the knowledge ending at a date. If logic is missing, verify. Check the machine's claims against primary sources, find the readers who will tell you the truth about your work, and practice the verdict, a no for this task, a no for that tool, held on the record. If rhetoric is missing, build one small thing for people who can find you, and let them. In my city that circle is already open. The Mark Cuban Foundation's free AI bootcamp for high schoolers runs three Saturdays each fall in Richmond, hosted by AI Ready RVA, which takes volunteers.
Grammar, the art of what a thing is, spent most of its history studying language, the instrument you think with. Learn the machine's grammar and somewhere partway down you will notice you are relearning your own. Those are the lost tools of learning, and they are our birthright. Nobody gave you the grammar. Take it anyway.