Agency, Assessment, and the Absolute Theoretical Minimum
Right now, we are still preparing many students with a 2022-style curriculum: help them pick a major or domain, get them into that lane, and help them get a job. While we run that increasingly outdated curriculum, some students study in passenger mode. They may coast through class, turn in assignments, and use AI for “cognitive offloading.” Rebecca Winthrop says that offloading can gradually become “cognitive stunting.”[1] Many professors are not yet trained to teach with AI, and they often lack an adequate support system when they confront AI-generated work. They rightfully so worry about both false negatives and false positives. So far, Rebecca Winthrop says the institution’s main response has been to teach students to use AI responsibly. That is almost like asking a two-year-old to show responsibility by not eating the cake on the dinner table. I am not blaming anyone. I am pointing to a basic fact of human behavior: frictionless convenience wins every time.
We want to prepare students for 2028 and beyond. Andrew Ng, drawing on task-level job analyses by Erik Brynjolfsson and Andrew McAfee, estimates that AI may cover 30 to 40 percent of many jobs, which leaves about 60 to 70 percent where the human role can still be large.[2] That remaining work demands high agency: decide what to do next while working with machines and people, then finish it with minimal supervision. To carry that load, graduates need more than a single major. They need an agentic, collaborative, end-to-end systems-engineering hat on top of multiple domain hats. Domain knowledge still matters, but it is not enough on its own. They have to build, not only think. Software production has become cheap, and by Jevons’ paradox, cheaper tools invite more use, not less.[3] Just as every accountant once had to learn the computer, a growing number of students across disciplines now need to build with software for real workflows. They do not all need to become software engineers. They do need software development literacy appropriate to their field.
The road from here to there starts with strong guardrails. Teaching responsible AI use is necessary, but not sufficient. True sufficiency requires AI-aware or AI-resistant pedagogy and a practical way to deliver it. The most effective methods remain those that require real-time thinking with short oral and written checks that put students on their feet. While students often complain about oral exams, many later admit they are the only assessments that truly force them to prepare and engage.
A tool like mymeritguide.com can possibly serve as the logistics layer: it schedules, delivers, and records oral or in-person written exams, collects peer reviews, and streamlines grading and scheduling. This removes the administrative burden that usually prevents frequent, authentic checks at scale. Class size and workload still present challenges, but the product does not replace faculty; it enables frequent, meaningful human checks. Institutional culture and leadership must adapt to support sustainable innovation. We must welcome constructive critique, adjust as real outcomes emerge, pair technology with non-tech supports, and monitor for equity gaps, privacy, and student well-being.
We want students who can read, think, write, speak, and build. Our AI-aware or AI-resistant pedagogy requires them to read, think, write, and speak, which builds problem-solving muscle, creativity, and ethical reasoning. How do we teach so that muscle forms? This is where faculty should shine. We should not try to teach everything. We should teach the absolute theoretical minimum until it becomes part of the student’s identity.[4] The absolute theoretical minimum is the essential core of concepts and methods that one must truly understand to master a subject, as defined by Landau’s standard for foundational knowledge.[5]
In my experience, in microeconomics, much of the course rests on the demand curve, the supply curve, total revenue, and game theory. In statistics, much of the course rests on the sample mean, the sample standard deviation, the t-statistic, and simple regression. In personal finance, a large share rests on simple growth, compound growth, amortization, annuities, and perpetuity. We have to re-explore these absolute theoretical minimums for each syllabus, and generative AI is useful for that work. AI has also enabled me, as a professor, to teach advanced topics such as time series forecasting, Monte Carlo simulation, Nobel Prize-winning models like Harry Markowitz’s mean-variance portfolio optimization, Black-Scholes methods, and advanced probability concepts like volatility drift and the Kelly criterion.
Alongside that domain knowledge, we must teach computer literacy. Many students still struggle with basic file and folder systems. Next comes AI fluency: how large language models work, how agentic AI works, and how to control complexity while building through an agentic software development life cycle.[6] Third, send students into internships as early as the pipeline allows. That is where domain knowledge, computer and AI skills, and human skills get tested and sharpened through real experience. Finally, curricula and educational tools must continually adapt as AI and technology evolve.
So our 2028 graduate should not be the student who picked a lane and waited for the boss. These students of ours should be able to decide the next move, collaborate along with machines and people, get the task or job done with little supervision, wear an engineering hat across domains, and build real workflows. To get there, we stop centering responsible-use of AI as the main fix. We stop grading take-home work a chatbot can finish alone. We stop teaching long topic lists. We focus on teaching the absolute theoretical minimums. We start with orals and in-class writing that force reading and speech, minimums that stick, literacy and AI fluency, and early internships.