We May Be Automating Away the Apprenticeship While Preserving the Test

In my previous commentary around AI, I mentioned how using AI as a cognitive crutch starves the development of human expertise. Many of us, especially educators, compliance officers, and parents, share a nagging suspicion that AI is progressively taking away our capacity for critical thinking rather than boosting it. Surely there should be some scientific basis behind this gut feel?

When you discover your surgeon is your friend who consistently used ChatGPT to pass his exams.

The most glaring concern about AI is also its biggest selling point. AI prides itself in making tedious, difficult things easier. Meetings are transcribed and minutes taken without lifting a finger. Papers we have not quite understood get summarized and even analyzed. Solutions to problems are produced in seconds what might otherwise take an afternoon. AI feels amazing when the difficulty is administrative, repetitive, or beside the point. It becomes more troublesome when the difficulty is the point.

Days spent mugging* at libraries or McDonald's should already revealed that learning is hard work, and for good reasons. New knowledge becomes more durable when we interpret it, connect it to what we already know, generate examples, compare it with other ideas, and apply it in context. Cognitive psychologists call elaborative rehearsal. Our mental map of Singapore is not built by memorizing street names in isolation. We recall the MRT lines we take, the expressways we drive along, the food centres we visit, and the neighborhoods where we work, shop, or meet friends. It reflects the same active learning principle found in established learning theories such as constructivism and experiential learning.

Memory is only the start. Expertise is developed through repeated encounters with patterns, feedback, and consequences. Each attempt gives us another opportunity to notice an important cue, refine a response, and discover what we previously missed. Over time, recurring patterns become easier to recognize, routine actions demand less conscious effort, and mental capacity is freed for the harder work of interpretation, judgment, and adaptation.

Eventually as experts, we reap the power of recognition-primed decision making, the ability to swiftly detect meaningful patterns amidst subtle cues and noise, and generate a response effective enough for the complex situation, without laboriously comparing a long list of options. An experienced nurse can literally smell something is wrong in a patient, way before any single observation appears alarming.

But wait, there's more. Doing 10,000 hours of the same thing won't get you to the level of expertise you seek. You also need deliberate practice, putting in time to work just outside the edge of our capabilities. Expertise is thus developed through this cycle of stretching, falling short, adjusting, and trying again, each time shifting the boundary of competence a little further.

AI is particularly effective at deleting these repetitions and denying the opportunities to build these libraries of patterns. Companies behind AI programs have vested interest to let people create and generate as frictionless as possible through their AI. All our struggles at work would disappear, so long we instinctively turn to AI for help. It even ends off with additional ideas and suggestions, all in the spirit of maintaining user engagement.

Like Ozempic, AI tools' allure stem in part from their quick and effortless results. Source

It does not help that our metrics measure our merit. A student submits a flawless essay. An associate builds a slick deck in record time. On paper, both look like stars. But the finished product tells us absolutely nothing about whether they can defend their logic, spot a hidden glitch, or survive if the power goes out. Yet, because scorecards only praise visible output, companies are eagerly replacing entry-level headcounts with algorithms, trading long-term human competence for a brief, blinding spike in productivity.

In such a landscape, telling students to be mindful when using AI tools is like having an escalator beside a staircase but insisting that the climb is good for them. They may agree in principle, right before stepping onto the escalator. We should not be surprised when convenience wins. Nor should we automate the climb and then blame students for arriving at the top without stronger legs.

Mindfulness alone is therefore a rather flimsy safeguard. Schools and workplaces must decide which forms of friction are pointless and which are developmental. AI should certainly spare us from formatting tables, searching badly organized files, and rewriting the same administrative email for the sixth time. But AI literacy should not crowd out the essential work of forming an argument, navigating a wrong turn, defending a judgment, or retrieving knowledge without assistance. Some friction is waste. Some friction is exercise.

Where the right balance lies remains up for discussion. I've heard academic colleagues returning to basics, deploying viva voces to supplement written assignments. Facing an insidious atrophy of professional expertise, workplaces too need to intentionally redesign processes so that sheer convenience does not cannibalize the on-job apprenticeship. We are only beginning to figure out how. And, ironically, asking ChatGPT for the playbook rather proves the point.

Avenue Q offers a cautionary tale of what might happen when the general public is left to negotiate AI tools on their own.

*Unlike its original English meaning, to mug is an antiquated Singlish verb that means to study hard, and a mugger was used to describe a studious person.