We’re Automating the Work and the Apprenticeship
REFERENCE: Haskell, C. (2026). Teaching Craft Intelligence: Sovereignty Protocols for AI-Mediated Student Leadership. New Directions for Student Leadership. https://doi.org/10.1002/yd.70070What AI is changing about college, entry-level jobs, and the path to expertise
Something strange is happening in higher education. Parents are watching newly minted graduates send out hundreds of applications, accept unpaid or tenuous work, move back home, or remain financially dependent much longer than they expected. At the same time, some entry-level jobs that once served as the first rung of a career are being reorganized around artificial intelligence. Colleges, meanwhile, are still arguing over whether students should be allowed to use AI to write a paper.
The arguments are related, although we rarely discuss them that way.
Universities have mostly treated generative AI as a question of academic integrity: what counts as cheating, what kinds of assistance are permissible, how much disclosure is required, whether a student’s work is still properly their own. Employers are asking a different set of questions: what work can now be automated, which roles can absorb more responsibility, and how much junior labor is still necessary when a competent first draft can be produced in seconds.
Caught between those two conversations is the student.
Not long ago, someone close to me said that I never write about anything that affects her life. She has college-age children. I thought about that comment while finishing a research paper on AI and student leadership, because the paper, at first glance, sounds like the sort of thing one might reasonably file under “higher education.” In practice, it is about something much closer to the kitchen table: what young people are actually learning to do at precisely the moment when competent-looking work is becoming easier and cheaper to produce.
The same ambiguity now troubles universities and employers. If a machine can cheaply produce some of the artifacts we once used to infer competence, institutions need better ways to determine where competence actually resides. AI disrupts one of the tacit bargains on which education and early-career employment have long relied: the finished artifact served as a rough proxy for the capabilities required to make it. A good memo suggested someone knew how to analyze. A polished presentation suggested someone had synthesized. A cogent essay suggested someone had read, compared, discarded, and formed a judgment. Generative AI makes the artifact much less reliable as evidence of the process.
That matters because the process was never incidental.
The student who researches badly, revises, gets corrected, discovers that a source is weak, watches an argument collapse, tries again, and eventually learns what a sound answer looks like is doing more than completing an assignment. The repetition itself is building judgment. My paper calls one danger of AI-mediated learning inquiry collapse: polished output can arrive before the cognitive work that would once have preceded it, allowing learners to bypass some of the remembering, understanding, application, and correction through which deeper expertise develops.
The danger is not simply that a student might accept a wrong AI answer. It is that the student may not yet know enough to recognize that the answer is wrong.
That problem is easy to see in skilled trades, where no one would confuse a clean weld with mastery of welding. The experienced welder has learned to read materials, conditions, tolerances, failures, and small irregularities that are difficult to reduce to instruction. But the same developmental process operates in project management, nursing, engineering, law, research, and organizational leadership. Expertise accumulates through encounters with situations that fail to behave as expected.
This is why apprenticeship time matters.
AI can accelerate parts of learning. It can explain, demonstrate, simulate, critique, retrieve information, and help a novice practice. What it cannot safely do is erase the sequence through which a novice learns what deserves attention. When the tool supplies an answer before the learner has built enough knowledge to interrogate it, efficiency can become a shortcut around the very experience from which judgment emerges.
The distinction that matters, then, is not between graduates who use AI and graduates who do not. It is between people who can obtain a plausible answer and people who have developed enough knowledge to examine that answer: to spot a bad assumption, reconstruct how a conclusion was reached, notice what the source base excludes, understand the consequences of a recommendation, and know when the apparently competent result should be rejected.
These are not generic “future skills.” They are the accumulated consequences of learning a field.
My argument is therefore not that young people should concentrate on whatever machines cannot do. We do not know where that boundary will settle, which makes it a poor basis for designing an education. Educating a generation around a moving technical perimeter would leave institutions perpetually chasing the most recent capability machines acquired. The better question is which human capacities become more consequential as machines participate more deeply in intellectual and practical work.
Judgment sits near the center of that question, along with the ability to trace an idea to its sources, notice whose knowledge has gone missing, distinguish confidence from warrant, and remain answerable for a decision after the software has done its part. My research treats those capabilities not as alternatives to using AI but as evidence that the person using it has retained agency over the work.
This is where the higher-education argument meets the employment argument.
Employers have historically hired junior people partly because junior work was how people became senior. The first assignment, the rough draft, the tedious analysis, the site visit, the meeting notes, the spreadsheet no one wanted to clean up; these tasks did more than produce output. They exposed novices to the texture of a field. Someone more experienced corrected them. Patterns accumulated. Eventually, the junior employee began to see what the experienced person saw.
Entry-level work, in other words, has functioned as developmental infrastructure.
If universities allow AI to shortcut too much of the learning process while employers simultaneously automate too much of the junior work through which graduates once deepened that learning, the problem is larger than plagiarism or job displacement considered separately. We risk thinning out the apprenticeship system on both sides of graduation.
That does not mean AI caused the difficulty young adults now face entering the labor market, nor that every automated junior task should be preserved for educational reasons. It means that institutions adopting these tools have another question to answer: if the old pathway to expertise is being compressed, where will the experience that replaces it come from?
That is also why the conversation happening in higher education and the one happening in homes where adult children cannot find a foothold are not separate conversations. They are different vantage points on the same problem: what counts as preparation for meaningful work when both the machinery of producing work and the pathway for learning how to do it are changing at the same time?
References
Haskell, C. (2026). Teaching Craft Intelligence: Sovereignty Protocols for AI-Mediated Student Leadership. New Directions for Student Leadership. https://doi.org/10.1002/yd.70070