Coursera Was Supposed to Replace the University. It Only Replaced Some Coursework.

Every generation of learning technology promises to remove the human from the loop. The generation that actually worked kept putting more humans in it.

A prediction that didn't land

In 2012, a wave of venture-backed platforms set out to unbundle the university. The pitch was simple: put the world's best lectures online, let anyone in the world watch them, and the credential-granting, tuition-charging, campus-owning institution becomes optional. Coursera was the most visible of these bets, and by most measures it succeeded at the distribution problem. Millions of people have taken a Coursera course.

It did not unbundle the university. Coursera listed on the New York Stock Exchange at 33 dollars a share in March 2021 (Coursera investor relations, 2021). As of this month, it trades at roughly a sixth of that (NYSE trading data, September 2026). People still apply to universities in record numbers. They still, overwhelmingly, describe their most valuable learning experience as a person, not a platform: the professor who pushed back on a bad argument, the mentor who noticed they were stuck before they said so.

What the data still says

None of this means the technology failed on its own terms. Course completion, content delivery, and credentialing at scale all work better now than they did fifteen years ago. What failed was the specific claim that distribution plus content equals learning. Learning, it turns out, has never been primarily a data transfer problem. It is a social and emotional process, built on attention from another person who can tell the difference between a learner who is stuck and a learner who is coasting, and who adjusts accordingly.

That distinction matters more, not less, as a new wave of AI tools makes the same promise the MOOC platforms made a decade ago: that the human in the loop is now optional. Some of that promise is real. A well-built AI tutor can answer a question at two in the morning, generate infinite practice problems, and never get impatient. None of that is the same skill as noticing, in real time, that a learner has understood the mechanics of a problem but not the judgment behind it.

The part AI actually replaces

Every training programme Skills Union runs still puts a working practitioner in the room, even in domains where the content itself could technically be delivered without one. That is not a nostalgic choice. It reflects a specific and repeatable observation from running these programmes since 2020: the content can be automated. The judgment about whether a specific person, at a specific moment, is actually building capability rather than just completing an assignment, cannot be, at least not yet.

What AI is proving useful for is instrumentation, not instruction. It can tell an instructor, in aggregate, where a cohort of forty learners is collectively stuck, faster than a weekly check-in ever could. It can surface the three learners who are quietly falling behind before they say anything. It cannot decide what an instructor should do with that information, because that decision depends on context the tool does not have: what this particular learner responds to, what they are actually struggling with underneath the surface error, what would land as encouragement versus what would land as pressure.

Where the leverage really sits

This has a direct implication for anyone currently designing AI-fluency curricula, which by 2026 means nearly every training provider and university continuing-education arm. The instinct is to treat AI capability as a technical skill to be taught alongside the existing syllabus: here is how the tool works, here is the prompt structure, here is the output format. That is necessary. It is not sufficient.

The more durable skill, the one that determines whether a graduate can actually use AI to compound their own capability over a career rather than plateau after the novelty wears off, is the judgment layer sitting on top of the tool: knowing when the output is good enough, when it is confidently wrong, and when the task in front of you needed a person's attention in the first place rather than a generated answer. That is a mentoring and coaching skill before it is a technical one, and it does not scale the way content delivery scales. It requires more human attention per learner, not less, at exactly the moment budgets are being redirected toward more tooling.

The implication for institutions

Universities, government training agencies, and enterprise L&D teams building their next generation of AI curricula would do well to remember what actually happened the last time distribution technology promised to remove the human from learning. It didn't. It just changed what the remaining humans needed to be good at. The institutions that treat this AI wave the same way, as a reason to invest more heavily in the judgment and mentoring layer rather than less, are the ones likely to produce graduates who can use these tools for a career, not just a semester.