Generative Learning Fixes What Is Wrong With Digital Learning

Digital learning did not fail because it was digital. It failed because we copied the lecture and left the lecturer behind.

Walk into a good classroom and watch the person teaching, not the slides. They are not performing a script. They are reading the room: a pause that lasts too long; three faces that go blank at the same moment; a confident wrong answer that exposes a shared misconception. The next sentence changes because of what they just saw.

That loop is the most valuable thing teaching does. For twenty years, much digital course design threw it away.

The missing half of the course

We put the content online and called it a programme: videos, PDFs, a discussion board, a quiz at the end of the week. Students could complete the module without anyone noticing that they had not learned it. Ghost attendance is the digital version of the back row: present in the system, absent in the mind.

What we called interactivity was often just clicking. A multiple-choice question that everyone passes on the second try is not observation. It is a gate. A lecturer watching a class is doing something else. They are continuously sampling understanding, then changing the next move because of what they find.

Courses that continuously assess and adapt to what each person knows can restore that loop. They can do it one student at a time, which a lecturer in a hall of eighty cannot. That is not a small upgrade. It is the missing half of digital learning.

Two meanings of generative, one idea

Merlin Wittrock named the first meaning decades ago. Learning is not receiving a message. It is generating meaning: selecting what matters, organising it, and connecting it to what you already know. Logan Fiorella and Richard Mayer later set out the activities that make that happen: explaining, self-testing, teaching, mapping, drawing, imagining and enacting. A live lecturer prompts those moves when they see they are needed. A static course cannot.

Generative now has a second meaning, and the two belong together. A generative system can produce the next question, the next example or the next explanation from what this student just did. Used badly, that is a chatbot that writes the assignment. Used well, it is a course that will not let the student stay passive. It keeps asking them to generate the next piece of understanding, and it checks whether they did. It then adapts what comes next and requires the learner to complete the relevant learning path before it declares competence.

The lecturer's glance across the room is a crude, collective assessment. Continuous individual assessment is the same act, made more precise. After every short stretch of learning, the system asks: can you do this yet? Not, did you watch the video. Can you do this?

If you can, you move. If you cannot, you get another path: a simpler example, a different representation, or a gap in a prerequisite that nobody spotted in week one. That is what a good lecturer does when they abandon the plan for ten minutes. The difference is scale. One lecturer can catch the class. They cannot catch every person at every fundamental.

Fundamentals are where this can beat the old method

This matters most for fundamental skills: the algebra under the science, the sentence under the argument, the procedure under professional judgement. Traditional teaching is often weakest here. The lecture moves at one pace. The student who is lost stays lost. The student who already knows it is bored.

Benjamin Bloom's 1984 two-sigma problem put a name to the challenge: one-to-one tutoring and mastery learning produced results that ordinary group instruction rarely matched. We could not staff that model at scale. We can now design towards some of its features – timely feedback, self-pacing and corrective practice – without pretending that software is a tutor by default.

A 2025 randomised trial at Harvard offers a useful, bounded signal. In two introductory physics lessons, a carefully designed AI tutor was tested against an active-learning classroom. Students using the tutor learned more in less time and reported greater engagement and motivation. The result does not settle the question for every discipline or every form of higher education. It does show what is possible when the system is designed around a pedagogy rather than dropped into a course as a chatbot.

The authors, Greg Kestin, Kelly Miller and colleagues, built scaffolding, timely feedback and self-pacing into the tutor. They also warned, in effect, against the opposite design: unguided AI can complete the work so the student does not have to think.

That distinction is the whole argument. Adaptive learning is not "AI in the course." It is a course that keeps looking at the learner and changes what happens next. For building fundamentals, that can be an alternative to the traditional method. Under the right conditions, the evidence says it can improve on it.

What it will not replace

It will not do this for every kind of learning. Synthesis, judgement, studio critique, clinical presence and the social work of becoming a professional still need people in a room. Virtual internships can simulate parts of the environment; they do not make presence, supervision or accountability irrelevant. As generative systems become more convincing, the important distinction is not whether an interaction feels human. It is who remains accountable for judgement, standards and care.

The better use of an AI tutor may be to get everyone to the starting line, then spend class time on the harder work. That is a better use of a lecturer than covering chapter four again.

Design still decides everything. A course that quizzes for completion will still produce ghost attendance. A system that marks a right answer wrong, or a wrong answer right, will teach the error with confidence. Continuous assessment only fixes digital learning if it is assessment of learning, not of clicking – and if a human still owns the curriculum, the standard and the moments that matter.

Put the conversation back

What was wrong with digital learning was never the screen. It was the silence after the content: no one watching, no one adjusting, no one asking the next real question of this student.

Generative learning, in both senses of the word, puts that conversation back. A course that continuously assesses the individual does what a lecturer does when they look up from the notes. Adaptive AI, used to build the skills underneath everything else, may do it more often, more patiently and for every student at once.

If that is true, the question for higher education is not whether digital learning can replace the classroom. It is whether we will finally build digital courses that behave like teachers.

Sources

• M. C. Wittrock, "Learning as a generative process" (1974)
• M. C. Wittrock, "Generative Processes of Comprehension" (1989)
• Logan Fiorella and Richard E. Mayer, Learning as a Generative Activity (2015)
• Logan Fiorella and Richard E. Mayer, "Eight Ways to Promote Generative Learning" (2016)
• Benjamin S. Bloom, "The 2 Sigma Problem" (1984)
• Greg Kestin et al., "AI tutoring outperforms in-class active learning" (2025)