Generating a complete course with AI: why it is tempting, and why it is often a bad idea

For the past year or two, one promise has circulated through every instructor forum: describe your course to an AI, and it generates a complete learning pathway in minutes. Titles, objectives, activities, quizzes: everything neatly packaged and ready to import into your LMS. Some tools on the market have made this their main selling point.

The temptation is real. Most of the instructors we work with spend hours facing a blank page, structuring, rephrasing and calibrating. If AI can absorb part of this work, why deny yourself it?

The honest answer is that AI can indeed help with specific tasks, at the right level of granularity. Generating an entire course in one go, however, is almost always a false lead. Not on principle, but because what AI produces in that case is not really training.

What AI can genuinely do: concrete gains

Let us start with what works, because the debate is not between "AI is magic" and "AI is useless".

AI is very effective at:

  • Overcoming the blank page. Drafting an initial outline or proposing a five-part structure from a vague learning objective. Not necessarily right, but enough to react to.
  • Rephrasing and simplifying. Turning dense technical text into an accessible explanation for a non-expert audience. This is one task current LLMs are genuinely good at.
  • Generating quiz questions. From existing content, producing variations of multiple-choice questions, gap-fill questions and practical cases. The result almost always needs reviewing, but the time saving is real.
  • Suggesting examples and scenarios. Providing a starting point for a problem situation, a use case or a simulation.

The strongest evidence on this point comes from a study listed by Stanford's SCALE centre in its review of causal research on AI in education. Teachers given access to ChatGPT alongside a user guide spent around 30% less time preparing their lessons and resources, or around twenty minutes per week. The decisive point is quality control: experts assessed the lessons produced blindly, without knowing which had benefited from AI, and found no difference.

This protocol is worth more than a self-reported survey. Teachers are not asked whether they feel they saved time; time is measured and the result is assessed blindly.

This figure deserves to be read carefully. It does not say that AI designs training. It says that specific, repetitive tasks with low pedagogical added value can be accelerated, provided you know which tasks to delegate.

What AI cannot do: where pedagogy remains human

A course generator produces generic content. Pedagogy is not generic content.

Three things no LLM can replace in learning design:

The organisation's tacit knowledge. What experts in your sector know but have never written down. The edge cases that make the difference between a capable professional and a dangerous novice. The common errors newcomers make. AI trained on general data does not know your context, internal processes, clients or specific regulatory constraints.

Examples that resonate with your real audience. An effective learning example speaks to a particular person, in their sector, with their references and at their level. AI produces plausible examples, not examples relevant to your learners.

Progression suited to actual levels. Calibrating a learning sequence means knowing what learners already master, what will block them and how quickly to introduce complexity. An LLM has no access to this reality. It produces logical progression, not learner-centred progression.

The problem of "AI slop": recognising a course generated without critical review

Instructors who attended training-AI sessions in 2025 and 2026 describe the same phenomenon: content that looks clean, is well structured and has compelling titles, yet on inspection says nothing specific. Generalities expressed confidently. Examples that could apply to any industry. Learning objectives formulated in vague, unverifiable terms.

This type of content has a name in the English-speaking e-learning community: "AI slop". Plausible filler. Volume without substance.

A few warning signs for instructors receiving or assessing AI-generated content: examples are interchangeable from one sector to another; learning objectives all begin with "understand" or "know" without ever specifying what to do with that understanding; the level of difficulty does not vary across the pathway; the content contains no opinion, position or nuance that would reveal genuine expertise.

Content well designed by an experienced instructor bears traces of its author. Mass-generated content does not.

The difference between generating a course and generating an activity

This is where granularity changes everything.

Generating an entire course means asking AI to decide in your place: which objectives, outline, progression, examples, level of complexity and activities. You receive a deliverable you did not design and for which you must now assume pedagogical responsibility.

Generating a specific activity from a learning intention you have formulated is different. You remain the author of the course: you have decided what to teach, in what order and for which audience. AI produces a first version of a quiz, explanatory page or practical case, which you review, adjust and approve before integrating it into your course.

The short loop between intention, proposal and instructor decision is what turns time saving into time genuinely recovered. When the loop is too long, when the instructor receives thirty generated pages at once, the work of reviewing and correcting ultimately costs more than starting from a blank page.

This last claim is not a practitioner's intuition. A study presented at the CHI conference in 2023 compared two ways for teachers to produce quiz questions: editing automatically generated questions or working with a system that kept them in control at every stage. Participants clearly preferred the second approach, judged questions mass-produced by the first as lower quality, and above all found editing automatically generated content more difficult than writing from a blank page.

One teacher interviewed summarised the mechanism better than any sales argument: when something is presented to them, they find themselves thinking only from that starting point, which puts a ceiling on them. Starting from zero, they can aim higher.

The authors conclude with two conditions for success: keep control with the user, and immediately show them a preview of what AI produces.

The real gain comes from selectivity, not volume

One result from the study cited above deserves attention because it contradicts the dominant commercial intuition.

Over time, the observed teachers used AI less and less: it was involved in 39% of their lessons at first, then 29%. Yet their time saving increased, from 27% over the first five weeks to 31% over the next five.

Less AI, more gain. The authors' explanation is simple: teachers quickly identify the tasks where the tool adds value and stop using it elsewhere.

This is exactly the opposite of what a complete-course generator promises. Competence is not about delegating more, but knowing what to delegate and where to retain control. A tool that produces everything in one block prevents that judgment from being exercised, because it leaves no choice of granularity.

Another study listed by Stanford adds a nuance we find more honest than the usual promise. An automated system for grading papers did not reduce teachers' total working time. However, student outcomes improved and teachers discussed more papers with students individually. Time does not disappear; it moves towards what has the most value.

Our approach: AI prepares, you decide

This is the principle on which we built PimenkoAI's Creation mode.

The instructor describes their learning intention and chooses an activity type. PimenkoAI proposes a first version, which the instructor previews before anything else. From there, the exchange continues as a conversation: ask for a more direct tone, add two questions, simplify an instruction or change the difficulty level. Every adjustment is immediately visible. When the result is right, the instructor approves it, and only then is the activity integrated into the course. Nothing is added to Moodle without that confirmation.

This reproduces, point by point, the two conditions identified by CHI researchers: control remains with the user, and the preview is immediate.

Today, 23 formats are available: 7 native Moodle activities and 16 interactive H5P contents. The catalogue may evolve, but the principle remains the same: one activity discussed, previewed and approved at a time, never an entire course produced in one go.

This is not a technical limitation; it is a deliberate pedagogical choice. We do not believe in the prompt that generates a complete course. Not because the technology would be incapable, but because the resulting deliverable contains dozens of pedagogical decisions that no one has made.

See how PimenkoAI's Creation mode works →

References