Using AI to Create Online Courses: What Works and What Doesn't

We build an AI-powered course tool, so treat what follows with appropriate scepticism. But the overclaiming in this space does real damage — people try AI course creation, get plausible-looking nonsense, and conclude the whole category is a con. It's worth being precise about what actually works.

Where AI genuinely helps

Structuring content you already have

This is the strongest use by a distance. Given a policy document, a slide deck or a procedure manual, AI is very good at identifying the natural module boundaries, proposing a sequence, and turning prose into headings, lists and knowledge checks.

It works because the substance is already there and correct. The model is reorganising, not inventing.

Drafting knowledge-check questions

Writing good multiple-choice questions is tedious and AI removes most of the tedium. It's also decent at generating plausible distractors, which is the hard part of question writing.

You still need to check them — models will occasionally produce a question with two defensible answers.

First-draft narration and summaries

Turning written content into spoken-word narration, or producing module summaries and takeaways, is well within reach and saves real time.

Alternative phrasings

Rewriting a paragraph for a different reading level, or making dense policy language plain, is genuinely useful and low-risk because you can see immediately whether it's right.

Where AI fails, reliably

Inventing subject-matter content

Ask a model to write a course about your organisation's return-to-work procedure and it will write a return-to-work procedure — confident, plausible, and not yours. Every specific detail will need checking, at which point you've spent longer than writing it.

The rule: AI is safe when the facts come from you and dangerous when they come from the model.

Anything requiring current or jurisdiction-specific accuracy

Legislation, standards, regulatory thresholds, award rates. Models are confidently wrong about these in ways that are hard to spot precisely because the output looks authoritative.

Judging what matters

AI can summarise a document. It cannot reliably tell you which three things in that document would actually change someone's behaviour on Tuesday. That judgement is the core of instructional design and it isn't automated yet.

Genuine assessment design

AI produces questions that test recall. Designing an assessment that tests whether someone can recognise a situation and act correctly — the thing that matters in compliance and safety training — still needs a human who understands the work.

The workflow that actually works

1. Start with real source material — your documents, not a prompt.
2. Use AI to structure and draft, not to originate facts.
3. Review every factual claim against the source.
4. Rewrite the parts that matter most yourself — the scenarios, the assessment, the framing.
5. Keep a human accountable for the output. 'The AI wrote it' is not a defence anyone will accept.

A test worth applying

Before shipping anything AI-drafted, ask: if this were wrong, how would I know? If the answer is 'I'd have to check it against the source anyway', then AI saved you formatting time, not thinking time — which is still worth having, but it's a different claim.