AI-assisted course authoring: faster, but checked

Much of course creation is routine work: reviewing material, drafting structure, wording texts, writing questions, creating graphics, translating, narrating. AI can speed this up considerably because language models read, organise and transform text and other services generate images, graphics and voices.

The most important design principle for AI in course authoring is: AI suggests, people decide. A generated text, graphic or question should first appear next to the course, with the option to compare several versions and to accept or reject them. Nothing should get into a published course unasked.

This principle protects against errors and keeps responsibility where it belongs. It also makes it traceable which parts of a course have been checked by people. A history that records every step and lets earlier states be restored is a sensible complement.

Language models write fluently and convincingly even when they are wrong. In course authoring this shows in several ways: invented figures or sources, outdated legal positions, generalisations that do not hold for your industry, and additions that were not in the source material. AI can also shift emphasis, for example giving a minor topic a lot of space.

Remedies are close grounding in the imported material, marking of additions, fact-checking of verifiable statements against sources and a final review by a subject expert.

Another, quieter error is uniformity. AI text tends towards a general, smooth tone and examples that fit everywhere and nowhere in particular. A course gains when subject experts contribute their own examples from the business, typical mistakes from practice and the language of the target audience. That is exactly what AI cannot know, and exactly what makes a course credible.

Behind AI-assisted authoring are language models from different providers. They differ in quality, speed, cost and where processing takes place. For some organisations it is essential that material is processed only within the EU; others want the most capable model. Good platforms leave this decision to the customer.

Also ask whether material you enter is used to train the models and what exactly is sent to them. For authoring, the material itself is enough; learners’ personal data has no place there.

Courses created with AI assistance must also respect copyright. Anyone uploading third-party material needs the rights to it, and AI-generated images should be checked for unintended similarity to protected works. The organisation that publishes a course remains responsible for its accuracy.

Where AI-generated voices or avatars are used, labelling is appropriate; Article 50 of the EU AI Act sets transparency obligations for this. And anyone using AI in course authoring should have sufficient AI literacy themselves; Article 4 requires providers and deployers of AI systems to ensure it.

In the Universe workshop, everything an AI generates first lands as a suggestion next to the slide; several versions can be browsed, accepted or rejected, and every acceptance can be undone. Verifiable statements are checked against sources when a slide is generated, before graphics and narration are produced. Each tenant decides which text models the workshop may use; one of them runs in the EU (Paris). Usage is charged to a workshop credit balance with a price table per operation.