Insights
AI is Making eLearning Cheap, Not Good
A recent thought-piece in eLearning Industry put it bluntly: AI made your eLearning cheap, not good. Here's how AI can make your courses better too.
A course that once took six weeks to build can now be drafted in an afternoon. The script, the slides, the quiz, the voiceover, even the presenter — generative AI will produce all of it before lunch, and it will look finished.
That last part is the problem.
A recent thought-piece in eLearning Industry put it bluntly: AI made your eLearning cheap, not good. It opens with an instructional designer describing a Six Sigma course she sat through on an AI-powered platform as a "dull, AI-generated slideshow with zero depth and zero value". The author's diagnosis is the right one: "AI didn't invent bad eLearning, but it did make bad eLearning easier to produce."
If you buy, commission or sign off on training, this is now your problem as much as your instructional designer's. The cost of producing a course has collapsed. The cost of a course that doesn't change anything has not.
Speed is real. It just isn't the point.
Let's be fair to the technology. AI genuinely speeds up course production.
The UK's Education Endowment Foundation ran a trial across 68 schools in which science teachers used ChatGPT to prepare lessons. They spent 31% less time on planning — about 56 minutes a week instead of 81 — and an expert panel reviewing the resources blind found no drop in quality.
In workplace learning, a 2025 study in Computers & Education compared AI-generated teaching videos with human-made ones across 447 participants. The AI videos took hours to produce rather than days. On the exam that followed, there was no statistically significant difference in what people learned. (Participants did enjoy the human-made videos more — a detail worth keeping in mind if you want people to finish what they start.)
So AI can produce competent content quickly. That's settled. The question for a buyer is what you do with the time and money you've saved.
There are two answers. You can make more courses. Or you can make the same courses better. Most organisations, without anyone deciding it, are doing the first.
The yardstick problem
The reason is measurement. Ask most L&D teams how a program is going and you'll hear about the things that are easiest to count: courses launched, enrolments, completions, satisfaction scores.
ATD's 2025 research on evaluating learning, a survey of 277 talent development professionals, found that 93% of organisations collect participant satisfaction data. Only 30% say they are good at using learning data to make business decisions. Nearly three in ten don't evaluate their programs at all.
When the scoreboard counts output, AI is a gift. It lets a team post a bigger number. Twenty new modules this quarter. Completion up. Everyone liked the narrator.
None of that tells you whether a single person does their job differently. As the eLearning Industry article puts it: "A completed course is evidence of participation, not proof of proficiency."
There's a useful ladder here. From easiest to measure to most important:
| Level | What it tells you |
|---|---|
| Output | How many courses or modules were produced |
| Participation | Who enrolled and who finished |
| Learning | What knowledge or skill people acquired |
| Performance | Whether they apply it on the job |
| Business outcome | Fewer errors, faster onboarding, better customer experience |
AI has made the bottom rung almost free. The top rungs cost exactly what they always did — because they depend on design, not production.
Informing is not teaching
Here's the distinction that matters most when you're reviewing an AI-assisted course.
AI is very good at informing. It will write a clear explanation, a tidy example and ten plausible quiz questions on almost any topic. But information is the cheapest part of learning.
Your people can already find it in a simple search.
Teaching is different. Courses that change behaviour tend to share four things:
- Relevance. The course is built around a real performance problem, not a topic. "Handle a pricing objection without discounting" rather than "Sales fundamentals".
- Practice. Learners make realistic decisions, under realistic conditions, more than once. Decades of research back this up: Dunlosky and colleagues' landmark review of learning techniques rated practice testing and spaced practice among the most effective strategies available.
- Feedback. Learners see the consequences of their choices and understand why. Hattie and Timperley's widely cited review found feedback is one of the most powerful influences on learning — and that its effect depends heavily on how it's given.
- Application. Something carries the learning back into the job. A meta-analysis by Blume and colleagues across 89 studies found that whether training transfers depends on factors like motivation and, crucially, a supportive work environment — not just the course itself.
A slideshow with a quiz at the end has none of these. That was true before AI. The difference is that it used to take six weeks and a budget line to build a bad course, which gave someone time to ask whether it should exist. Now it takes an afternoon.
What AI should be doing in your courses
None of this is an argument against AI. It's an argument for pointing it at the right part of the work.
The expensive parts of good learning design have always been the parts that make courses effective: building realistic scenarios, writing enough varied practice that people can actually get good at something, and giving each learner feedback specific to the choice they made. Those were the first things to be cut when budgets got tight.
That's exactly where AI can be help course designers be more effective. Used well, it lets a design team build branching scenarios that would have been unaffordable, generate dozens of practice cases instead of three, personalise feedback, and analyse what learners are actually getting wrong. The time saved on production gets reinvested in the parts that change behaviour.
Used badly, it produces more of the slideshow.
Five questions to ask before you sign off on an AI-built course
Whether you're building in-house or briefing a vendor, these questions separate capability from content:
- What must people be able to do differently afterwards? If the answer is a list of topics, the course isn't designed yet.
- Is training the right fix? Some performance problems are process, tooling or management problems. A course won't solve them, however quickly it's built.
- Where do learners practise? Count the moments where someone makes a decision, not where they read or watch.
- What feedback do they get when they're wrong? "Incorrect, try again" is not feedback.
- How will you know it worked? Agree before launch what you'll measure beyond completion — a quality metric, a time-to-competence figure, a behaviour a manager can observe.
If a vendor's pitch leads with how fast they can produce courses, ask them those questions too.
The real opportunity
AI has handed L&D something it has never had: cheap production. The organisations that win won't be the ones that use it to flood their LMS. They'll be the ones that take the savings and spend them on the expensive, unglamorous work of design — practice, feedback, application and honest measurement.
Count capability, not courses. That's the standard worth holding AI to, and it's the standard we build to at Credentialcraft.
If you're planning a course and want it to change what your people do, let's talk about what it needs to achieve.
Frequently asked questions
Is AI-generated eLearning less effective than human-made eLearning? Not inherently. A 2025 study of 447 participants found no significant difference in learning between AI-generated and human-made instructional videos. Effectiveness depends on instructional design — relevance, practice, feedback and support for applying learning on the job — not on who or what produced the content.
How should organisations measure the quality of eLearning? Look beyond completions and satisfaction scores. Measure whether people acquired the skill, whether they apply it at work, and whether it moved a business outcome such as error rates, onboarding time or customer experience. Agree those measures before the course is built.
What is the best use of AI in instructional design? Use AI to make the expensive parts of good design affordable: realistic scenarios, varied practice activities, personalised feedback and analysis of learner data. Avoid using it simply to produce more content faster.
What should I ask a vendor building courses with AI? Ask what learners will be able to do differently, where and how often they practise, what feedback they receive, and how success will be measured beyond completion rates.
Sources
- AI-Designed eLearning Quality: AI Made Your eLearning Cheap, Not Good — eLearning Industry
- ChatGPT in lesson preparation: Teacher Choices trial — Education Endowment Foundation
- Netland et al. (2025), Comparing human-made and AI-generated teaching videos, Computers & Education
- ATD Research: Organisations struggle with measuring the impact of training
- Dunlosky et al. (2013), Improving students' learning with effective learning techniques, Psychological Science in the Public Interest
- Hattie & Timperley (2007), The power of feedback, Review of Educational Research
- Blume et al. (2010), Transfer of training: A meta-analytic review, Journal of Management