# The Prosthetic Curriculum > Colleges now sell AI literacy to students raised on screens. They arrive diminished, and AI supplies what is missing. Call it a prosthetic curriculum. https://isaacmarcuson.com/blog/ai-literacy-humanities/ Artificial hand, after Ambroise Paré, 1564. [Wellcome Collection](https://wellcomecollection.org/works/gea92cjg), [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/), modified. This fall I worked my company's booth at a university industry expo. Between conversations I sat in on a panel of professors and administrators explaining why their college now builds AI into the humanities curriculum. Enrollment in the humanities has been falling for years, and AI in a course title has turned out to be a successful marketing strategy. Students need the tools to get hired, they said. They also need the help, because students raised on screens arrive diminished, and AI can augment what is missing. What remained of a humanities education, in the version they presented, was checking the machine's output. The instruction, as near as I can compress it, was fact-check it, remove the tells, make it read like your own. They kept returning to one phrase for what all of this prepares students for: imagined futures. The expo existed to put students in front of employers, and I was there on that side of the table. I have always been an autodidact, and I know what an autodidact lacks, which is a tradition that corrects you. I sat down hoping to hear from people who have spent their lives inside one, and I heard a business strategy. Afterward I asked a few of them privately whether anything their students make with AI matches what a student could have made by hand forty years ago. One answered with a question. "Have you seen what the students do at this point?" He meant it as a defense. AI gets them up to the bar. ## What the Teaching Replaces He is right about the students. A third of American eighth graders [read below Basic](https://www.nagb.gov/powered-by-naep/the-2024-nations-report-card/10-takeaways-from-2024-naep-results.html) on the 2024 national assessment, the largest share ever recorded at that grade, and [adult literacy has fallen with them](https://ies.ed.gov/learn/press-release/u-s-adults-score-par-international-average-literacy-skills-below-international-average-numeracy-and), the share of adults at the lowest level up from 19 to 28 percent since 2017. The causes are many, and screens are one of them. AI literacy is also worth teaching, and every student should know what a language model is and where it fails. My objection is to what the teaching replaces. A prosthesis that restores a lost function is a mercy. The case I am worried about is a prosthesis fitted to a function that never developed. It produces the output, the output looks fine, and nothing in the student's life signals that the capacity underneath never grew. Put the panel's diagnosis and remedy back to back. Students arrive diminished, and AI supplies what is missing. The supply removes the reason to grow it. In 2025 a Wharton-led team published [a field experiment](https://www.pnas.org/doi/10.1073/pnas.2422633122) run in Turkish high school math classes with about a thousand students. Students who practiced with an ordinary chat model scored about half again higher on practice problems, and then, on an exam with the model taken away, scored seventeen percent lower than students who never had it. A second group used a version built to give hints and refuse to hand over answers. Their practice improved more, and the exam harm disappeared. The second group moves the question off the machine and onto a decision the panel never raised. Someone has to say which work the student still does unaided. Model in hand Model removed no model +127% +48% Tutor hints only Chat model full answers no measurable difference -17% Model in hand Model removed students with no model +127% +48% Tutor hints only Chat model answers on request no measurable difference -17% Each line is one tool, read from practice with the model to the exam without it. Only the exam end carries wine. Those were math classes. A text works the same way. Ask the model for the reading and you get a reading, and the work that would have taught you to judge one never happens. That is what the exam measured. Some students will treat the tool as scaffolding and climb past it. For the rest, the loop runs one way. The student hands over the work, the model supplies the performance, and the performance stands in for the capacity, which never forms. Once assisted performance is the baseline, the department has less reason to keep the assignments that built independent capacity, and the next class inherits the accommodation as the curriculum. One link in that cycle has been measured, and the incentives for the rest are already in place. Every turn of the ratchet looks like help. Call it a prosthetic curriculum. ## The God Nobody Worships on Purpose There is a name for this pattern. Moloch, in the sense Scott Alexander gave the word in [Meditations on Moloch](https://slatestarcodex.com/2014/07/30/meditations-on-moloch/), is the god nobody worships on purpose. He is a system with no one at the controls, where every player makes the locally sensible move and the sum devours what all of them say they care about. Jonathan Pageau, an Orthodox Christian, reads that game as idolatry, a partial good demanding the sacrifices owed to something higher. The university is playing it. Budgets need enrollment, legislators fund what they can count, indebted students want jobs, and a dean who refuses the race loses the department. Each step is defensible, and the god is built entirely out of defensible steps. The first time [Leviticus](https://www.biblegateway.com/passage/?search=Leviticus+18%3A21&version=KJV) names him, the whole law about him fits in a sentence. Do not let your children pass through the fire. The original Moloch was paid in children, which is to say you hand over the future you actually have for the security you imagine. ![Blake's dark idol of Moloch, crowned in wine, sits above a child in the fire while the parents kneel on either side. The crown is all the authority he has, and the people around him gave it.](https://isaacmarcuson.com/wp-content/uploads/2026/10/ai-literacy-humanities-moloch.webp)The god has already fled the idol in Blake's Flight of Moloch, about 1815, and the cymbals go on without him. Imagined futures was the revealing phrase, because someone has to do the imagining, and here it was employers. The World Economic Forum writes their forecast down as the [Future of Jobs Report](https://www.weforum.org/publications/the-future-of-jobs-report-2025/), and once employability is the reason a department exists, a forecast of what employers will want becomes a statement of what education is for. Universities are already writing it in, [Ohio State](https://fortune.com/2025/06/10/ohio-state-university-ai-fluency-requirement-all-students) first among them, with AI fluency now expected of every student. The forecast itself puts analytical thinking first. A list of the skills employers want says nothing about how those skills form, and whether a curriculum built around the tool still forms them is a question enrollment numbers cannot answer. ## What the Monastery Knew Among the university's inheritances is the monastery's idea of what education is for, that a person is formed over a lifetime through practices nobody expected to work alone. Prayer, manual labor, shared meals, silence, the slow reading of a text. Basil the Great, [writing to his nephews](https://www.tertullian.org/fathers/basil_litterature01.htm) in the fourth century, told them to read the pagan poets the way bees work a meadow, taking what is useful and leaving the rest, and he meant it as an apprenticeship. Judgment is trained on lesser material first so that it is ready when the precious material comes. The panel's AI literacy assumes the judgment Basil spent an apprenticeship forming. ![Monks in Fra Angelico's Thebaid carry a reader up the hillside path, keep a walled garden, and bury a brother at its gate, while the slope above them fades into bare page. Nothing here is done alone, and the one woven basket in wine is labor carried into prayer.](https://isaacmarcuson.com/wp-content/uploads/2026/10/ai-literacy-humanities-thebaid.webp)In Fra Angelico's Thebaid, about 1420, the desert fathers bury a brother beside the garden they keep. One mourner has brought his basket. John Vervaeke, the cognitive scientist behind [Awakening from the Meaning Crisis](https://www.meaningcrisis.co/), calls that structure an ecology of practices, and his account of knowing explains why it works where a single practice fails. He counts four kinds of knowing, and only one of them, knowing that, can be handed over in a sentence. Knowing how, knowing what it is like to stand where you stand, and the knowing that comes from belonging to a world that shapes you while you shape it all enter a person through practices, and each practice covers the blind spots of the others. A model's answer is the first kind. Handed over before the student has built one, it feels like a tutor and works like a crutch. Moloch wins by collapsing everything into one metric, and an ecology has no single metric adequate to its health. An ecology does not abolish the pressure. It gives the institution one definite thing to protect. A student who can still do the work with the tool removed. If a university in this century has a reason to exist beyond credentialing, it is this one, to run that ecology for people who cannot yet run it for themselves. A humanities department that has reduced itself to quality control on machine output has quit that job and kept the name. ### AI Literacy in One Semester Five practices for the real job, taken in this order. 1. Read the text aloud, together, before anyone asks a model about it. 2. Write the first draft by hand. 3. Argue in pairs. Before you rebut, summarize your partner's position until they agree you got it right, then revise your draft. 4. Sit with one hard passage in silence for ten minutes before discussing it. 5. Then bring in the machine. Ask it the same questions, hunt its hallucinations, grade its reasoning, and set its reading beside the one the class built. The panel also wanted students checking the machine. The difference is the first four steps. You can only check an answer against a judgment you already own, and the version they presented skips the owning. The last step makes the employable value of the first four visible, because catching a confident wrong answer is the job in every AI workflow I have seen. At the end of the term, hand students a passage they have never seen, with no model in the room, and ask them to read it, defend the reading, and revise it under questioning. That tests whether the capacity survives the removal of the tool. The sequence fits inside an existing course, and step five ticks the AI box. ## What the Work Needs From an AI Hire I work in AI. I edit model output every working day, and I have written about [the day a model handed me](https://isaacmarcuson.com/blog/how-to-learn-ai/) a polished, attributed quotation from a text that does not contain it. What caught it was a habit formed years earlier on texts, against sources, among people who said so when I was wrong. I was on the employer's side of the table at the expo, and that habit is what the work needs from an AI hire. The humanities at their best teach it, and the university is now pitching that teaching as an attachment to the machine. Moloch is paid in children. The coin here is a student's future judgment, spent for work that looks fine this term. Keep the children out of the fire.