Why higher education must preserve the hard work of learning in the age of AI
By Joshua Jones
In my undergraduate photography class, dodging meant standing in the dark, holding a wire-and-cardboard tool between light and paper, adjusting exposure by feel. I learned to feel the inverse square law of light in my fingertips, to sense when a shadow needed another half-second of protection. Years later, I recognized that improvised tool in Adobe Photoshop’s dodge icon. The physical struggle had become a digital button. Today, even the button is disappearing. Adobe’s Generative Fill simply asks what I want. The tool no longer aids the process; it threatens to replace it entirely.
This evolution — from hands-on practice to digital tools to AI automation — captures the central crisis facing higher education. We are not merely adopting new tools. We are changing how people develop expertise. The risk is not simply that AI will make graduates unemployable. The deeper risk is that it will make them what Socrates warned of in the Phaedrus: people who possess “the show of wisdom without the reality”. When the god Theuth presented writing to King Thamus as a gift for memory and wisdom, the king saw through the sales pitch: the invention would produce forgetfulness because learners would no longer practise memory. Today’s technology is different in scale, but not in kind. If students can bypass every form of cognitive friction, we may produce graduates with access to infinite answers but no internal pathways to reach them.
The brain builds capacity through resistance. When we struggle with a difficult task, repeated effort strengthens the neural circuits involved; over time, brain cells wrap frequently used pathways in myelin, an insulating layer, making them faster and more reliable. This is the biology behind what cognitive psychologist Robert Bjork calls “desirable difficulty”: learning that requires effort is learning that lasts.
The famous study of London taxi drivers illustrates the principle vividly. Cabbies who master “The Knowledge” — the thousands of streets and routes required to navigate London — show significantly larger posterior hippocampi than comparison groups. Other research on GPS use and spatial memory suggests the opposite danger: when navigation is outsourced, spatial memory receives less practice. The tool works, but the underlying skill can weaken. The cabbie can navigate when the phone battery dies; the GPS-dependent driver cannot. This distinction — between a tool that supports a capability and one that replaces the organ responsible for it — is the crux of the AI debate. We are not arguing about efficiency alone. We are arguing about whether students will retain the skills and judgment to function when the system goes down.
A landmark Harvard Business School and Boston Consulting Group study quantified this danger. Consultants using AI completed more tasks, worked faster and produced higher-quality outputs — but only within the range of tasks AI handled well. For tasks requiring subtler judgment or edge-case reasoning, AI-assisted consultants were significantly less likely to produce correct answers than those working manually. They trusted the machine’s “show of wisdom” and lost touch with their own judgment.
The consequences of such atrophy can be severe. When Air France Flight 447’s pitot tubes iced over in 2009, the autopilot disconnected amid confusing and unreliable airspeed indications. The crew failed to recover from a high-altitude stall. Instead of pushing the nose down — a basic flying skill — they pulled up, deepening the stall until the plane struck the Atlantic. The point is not that automation alone caused the crash. It is that advanced systems can create a cruel dependency: they handle ordinary conditions so well that humans get less practice with the very skills they need when conditions become extraordinary. Lisanne Bainbridge identified this “ironies of automation” problem decades ago. The better the automated system, the less prepared the human operator may be when asked to take over.
This pattern is already restructuring the workforce. Professional mastery has historically relied on apprenticeship. Junior employees perform tedious, high-volume work — document review, basic code, data cleaning, financial modelling — that provides the repetition required to build judgment. These tasks are the cognitive equivalent of shovelling snow: inefficient, uncomfortable and essential. AI excels at precisely this entry-level work. The temptation is to automate the drudgery. But if firms stop hiring and training juniors, they destroy the pipeline. You cannot import a senior partner, analyst or engineer fully formed. They must develop through experience in junior roles.
Economists Daron Acemoglu and Pascual Restrepo have warned of “excessive automation”: displacing labour faster than we create new, complex tasks in which humans can develop and contribute. Higher education should take this warning seriously. Some inefficiencies must be preserved not because they are required to produce the work, but because they help develop the person who will eventually be responsible for it.
Some will argue that AI literacy itself is the new foundational skill. Students, they say, must learn to navigate what researchers call the “jagged frontier” of AI capability: knowing when AI performs well, when it fails and how to prompt it effectively. That is true, but it is not enough. The frontier moves quickly, and domain knowledge is still required to verify what lies on either side of it. The BCG consultants were not AI novices; they still failed outside the frontier because they lacked, or failed to apply, the darkroom knowledge needed to recognize when the machine was confidently wrong.
It has been said that asking the right question is half of knowing. But without domain knowledge, students cannot reliably formulate the right questions, nor can they recognize when AI’s answers are wrong. AI fluency without domain mastery is not a tool. It is a blindfold.
The path forward requires deliberate friction. First, institutions should create AI-free zones: oral exams, in-class writing, manual problem-solving and live demonstrations of reasoning. These are not nostalgic rituals. They help protect students’ ability to think independently. Second, we should teach a model that combines human judgment with AI: students must learn to shovel the snow before they receive the snowblower, understanding both how to use AI and when to work without it. Third, we should reframe apprenticeship. Entry-level inefficiency is not waste; it is investment in future judgment.
The darkroom is gone, and it should be. But the struggle must remain. As philosopher Shannon Vallor argues in Technology and the Virtues, we cannot simply ask whether a technology is efficient. We must ask what kind of person it encourages us to become. In the age of the algorithm, higher education must produce graduates who know the matter — not merely those who can summon its appearance.
—
Joshua Jones is CEO of QuantHub, an AI education company. His team received a DARPA award for their work in educational technology.