If AI is doing the entry-level grunt work, can a classroom teach the skills people used to learn on the job?
Can universities teach what workplace hands-on experience once provided?
This explores whether classroom education can supply the practical, learn-by-doing skills that early-career workers used to pick up on the job, now that AI is taking over much of that work.
This explores whether classrooms can stand in for the on-the-job apprenticeship that AI is now absorbing. The corpus has no study of universities teaching practical skills, so what follows is stitched together from adjacent evidence. The most direct evidence says the workplace is a worse teacher than it used to be. Workers who used generative AI performed much better on content tasks, but when they later did similar tasks alone, they showed no improvement (Does AI assistance help workers learn lasting skills?). Doing the work with AI produces the output without producing the skill. Delegation is also concentrated in information-intensive jobs (Where have workers actually delegated tasks to AI?), which is where much of the entry-level practice used to happen.
That shifts the question. Universities may not need to copy the workplace. They may need to build the practice the workplace no longer offers, and right now they aren't set up for that. An audit of 30 universities found their AI policies are good at listing which uses are allowed. They are poor at saying what evidence shows a credential still certifies that someone learned something (Do university AI policies actually protect what credentials mean?). A degree that can't show learning happened can't claim to replace experience.
The training research on models offers three design ideas. These are analogies from machine learning, not findings about people. First, classrooms have a real edge on explicit principles. Models fine-tuned on labeled examples of good and bad arguments learned surface patterns, not the criteria behind them, and only improved when taught the theoretical framework directly (Can models learn argument quality from labeled examples alone?). Experience alone may leave a learner with pattern-matching, which is what a workplace often delivers. Second, order matters. Imitation first and then exploration beat either one alone, because imitation gives the later practice something sensible to sharpen (Does sequencing imitation then exploration training improve reasoning?). That is roughly a syllabus that moves from studying worked examples to open-ended practice. Third, a curriculum can substitute for a mentor watching every step. Reverse curriculum learning starts a learner near a finished result and slides the starting point backward, so failures show up at specific steps using only outcome feedback (Can curriculum learning approximate expensive process supervision?). A course could hand students nearly finished work to complete, then hand over earlier and earlier stages.
Expert material can also be too good for the learner. Teacher-refined data hurt student models when it exceeded what the student could absorb, even though it was objectively higher quality (Does teacher-refined data always improve student model performance?). Polished expert examples don't automatically teach novices, and the gap between novice and expert may be exactly what apprenticeship bridged. One project takes on the tacit part directly. COLLEAGUE.SKILL captures a person's expertise as versioned, inspectable files and keeps what they know separate from how they behave (Can person-grounded skills remain auditable without hidden prompt state?). That suggests know-how can be partly written down and audited, but the note is about governing such files, not about whether people can learn from them.
The corpus suggests a qualified yes. Universities can teach the principles and can build structured practice, but only if they treat AI-assisted work as evidence of output, not of learning, and design assessment around what students can do unaided. The corpus has no measurement showing that any of this works for people.