Course & learning material generation

Preparing courses is slow: outlines, slides, handouts and exercises must be written, levelled, checked against the curriculum and kept up to date. Generative AI drafts these materials from learning objectives and trusted source texts in minutes. The educator stays the author: they steer, verify and approve. The main risks are factual errors, copyright and missing curriculum alignment.

Problem & value

• Today: teachers, trainers and lecturers build materials by hand, copy from old decks or buy publisher content. Differentiated versions (levels, languages, accessible formats) are rarely produced because of lack of time.
• AI value: first drafts in minutes, consistent structure (objectives, activities, check questions), quick adaptation to level, language or length, and refreshing outdated material.
• Evidence: Oak National Academy (England) reports from qualitative research with about 60 teachers that its AI lesson assistant Aila saved roughly 3.5 hours of planning time per week; this is a vendor-reported, self-selected figure and not an independent trial [1][2].
• Pre-condition: value depends on grounding in vetted material. Free-form generation without sources is the main cause of errors.

Variants

• School teacher: lesson plan + slides + worksheet aligned to a Lehrplan/curriculum; see EDU-009 for differentiation and teacher assistant functions.
• Higher education: lecture slides, reading lists, case studies and exercise sheets from the lecturer's own scripts and papers; often via LMS plug-ins (Moodle, ILIAS).
• Corporate training: onboarding and compliance modules from SOPs and policies; see EDU-006.
• Publishers & OER: large-scale drafting and localisation of open educational resources.
• Private learners: self-made study guides from own sources; see EDU-016.

How it works

1. curriculum / objectives + target group
2. source ingestion (PDF/DOCX/slides
3. text; OCR where needed; chunk + embed)
4. retrieval of relevant passages from vetted sources (RAG)
5. outline generation (modules, objectives, timing)
6. section drafting with citations to sources (text, examples, activities)
7. format rendering (PPTX, DOCX, H5P, SCORM/xAPI, LMS import)
8. automatic checks (reading level, curriculum coverage, source support)
9. educator review & edit
10. publish to LMS

• Grounding: retrieval over the educator's own or openly licensed material, with every claim linked to a passage, reduces hallucination (see KNOW-002).
• Structured output: the LLM returns JSON (objectives, slides, items) that templates turn into files.
• Human-in-the-loop: the educator approves every learning object before students see it. Published AI-generated text should be marked as such where legally required (Art. 50 AI Act, see below).

Data preparation

The main inputs are curricula, learning objectives, own scripts, slide decks, textbooks and OER. In an SMB or school they sit on file shares, in the LMS, in teachers' personal folders, in e-mail attachments and partly on paper. Typical gaps: outdated duplicates of the same deck, scanned handouts without text layer, missing learning objectives per document, unclear copyright status of third-party material and knowledge that exists only in one trainer's head.

Preparation steps: collect an authoritative set of current source material in one folder or knowledge base (see PREP-013); OCR scanned handouts (see PREP-006); add simple metadata such as topic, level and licence (see PREP-010); keep a few gold-standard examples of good materials as style templates (see PREP-016); remove personal data such as learner names from reused materials. Convert slides and PDFs to clean text with a document parser (see DOC-014).

Effort: low. For pure drafting from prompts almost nothing is needed; for grounded generation expect roughly 10 to 20 percent of the total effort, mostly for collecting and tagging sources. Check the quick readiness steps in GUIDE-003.

Implementation paths

Which path for whom

• SMB: Path A, because mature EU-hostable authoring tools exist at SMB prices and drafting materials is a standard process.
• Enterprise: Path B, because the drafts should be grounded in the own LMS, curricula and content library via API.
• Developer: Path C, because an open-weight stack with RAG over own sources gives full control and runs on HW2.
• Public: Path B on a sovereign EU or German cloud, because procurement rules and school data rules favour hosting in the EU.
• Private: Path A, because ready-made apps cover lesson planning for individual teachers and learners without setup.

Path A — Buy (recommended for most)

• Teacher/LMS tools: Microsoft Copilot in PowerPoint/Word/Teams for Education, Google Gemini for Education (Classroom, Slides, NotebookLM), Canva Magic Studio for education, MagicSchool, Diffit (levelled texts), Curipod, Twee (language teaching), Oak's Aila (UK curriculum, free).
• EU-based or EU-hosted: fobizz (DE, school-oriented AI tools incl. text and image generation), Mebis/telli-style state-run school offers where available, Gamma is US-based - check hosting.
• Authoring suites: Articulate Rise/Storyline AI Assistant, Genially, H5P with AI plug-ins, Moodle AI subsystem (provider-configurable, can point to a self-hosted model) [3].
• Choose A when: a few teachers or trainers, low volume, no special integration; verify a data processing agreement, EU data residency and that inputs are not used for training.

Path B — Build on cloud APIs

• Models: Mistral Large/Medium (EU provider), Claude, Gemini, GPT via EU regions (Azure OpenAI EU, Vertex AI EU, AWS Bedrock Frankfurt/Paris); structured outputs for JSON schemas.
• Stack: embeddings + vector store (pgvector, Qdrant), document parsing (Docling, Mistral OCR), PPTX/DOCX generation (python-pptx, docx libraries), H5P/xAPI export, LMS integration via LTI 1.3.
• Image/audio: image generation for illustrations (check licensing and Art. 50 marking), TTS for audio versions of texts.
• Choose B when: you want curriculum-specific prompts, your own content library or LMS integration.

Path C — Self-host (open source)

• Models: open-weight LLMs such as Mistral Small/Medium open releases, Qwen3, Llama, Gemma or gpt-oss, served with vLLM or Ollama; multilingual embeddings such as BGE-M3.
• Tools: Open WebUI or LibreChat as front end, Moodle/ILIAS/Open edX as LMS, H5P as content format, Docling for ingestion. Oak's AI lesson assistant is published as open source on GitHub [4].
• Hardware: HW2 serves a university department or school network with a ~30B quantised model; HW3 for larger models and many concurrent users.
• Choose C when: a school authority or university must keep pupil and staff data on premises.

Path D — Custom training

Rarely needed. Fine-tuning pays off only for a publisher or ministry with a large, consistently styled corpus (house style, curriculum taxonomy, rubric formats). Prompting + retrieval covers almost all cases.

State of the art (2026)

• Frontier and strong open models draft coherent lesson sequences, slides and levelled texts; tools chain outline → slides → quiz in one flow, and "notebook"-style products create summaries, audio overviews and study guides from uploaded sources.
• Source-grounded generation with citations is the established good practice; curriculum-aligned retrieval (as in Aila, which mainly draws on Oak's own checked resources) is what makes drafts usable [1].
• Interactive content (H5P, SCORM, simulations) can now be generated as code or JSON.
• Limits: factual errors and invented references, shallow pedagogy (lists of activities without sequencing), cultural and gender bias in examples, uneven quality in smaller languages, wrong difficulty levels, and inconsistent images. Maths and science content needs expert checking.
• Generated material tends toward homogeneity; teachers report the need to edit and adapt before use.

Shared resources & combinations

• Same stack: EDU-002 (quiz items), EDU-009 (lesson planning, differentiation), EDU-006 (corporate microlearning), EDU-016 (study tools from own materials), KNOW-011 (FAQ/knowledge base generation).
• Building blocks: RAG pipeline (KNOW-002), document parsing (DOC-014), translation/localisation (LANG-009), accessible output (DOC-026: PDF/UA, alt text).
• One self-hosted LLM + vector store on HW2/HW3 can serve all of these.

Effort, cost & KPIs

• Path A: days; licence per teacher or institution, often bundled with existing suites.
• Path B: 1-3 months for a pipeline with LMS export and review workflow; token costs are small relative to staff time.
• Path C: 2-4 months incl. operations; one GPU server.
• KPIs: preparation time per lesson/module, share of AI drafts accepted with minor edits, error rate found in review, curriculum coverage, learner outcomes and satisfaction, accessibility conformance.

People & change

Involve management for the AI policy and budget, the department lead or head of training as process owner, and the teachers or trainers who will use the tool from the start, since they remain the authors. IT handles accounts and LMS integration; the DPO is consulted early, mainly if learner data or class lists could end up in prompts (AVV with the vendor, DPIA usually not needed for pure content drafting).

Works council: no in the typical case, because the tool produces materials and does not evaluate staff. It becomes relevant if usage logs per employee are evaluated (§87(1) no. 6 BetrVG, objective suitability suffices), and the council must be informed in good time about planned AI-supported work processes (§90(1) no. 3 BetrVG). In schools and public bodies the staff council (Personalrat) has comparable rights.

AI literacy: the AI Act Art. 4 requires measures to support staff's AI literacy; offer a short workshop on prompting, fact-checking and copyright. Start with a pilot group of two or three motivated teachers, name AI champions and communicate openly that the educator signs off every material. Typical concerns are quality, copyright and fear of deskilling; address them with review rules and an acceptable-use policy. See GUIDE-007, GUIDE-006 and GUIDE-008.

Risks, regulation & compliance

• EU AI Act: content generation itself is generally limited/minimal risk. Art. 50 requires machine-readable marking of synthetic image/audio/video and disclosure of AI-generated text published to inform the public, unless a human has editorially reviewed it and takes responsibility; deepfake-style content must be labelled. Art. 4 AI literacy applies to staff [5]. If generated content feeds assessment, admission or steering of learning (Annex III point 3), that downstream use may be high-risk (see EDU-002, EDU-003, EDU-007). Emotion recognition in schools is prohibited under Art. 5.
• GDPR: avoid uploading pupil data; use anonymised or fictional examples. Student work as source material needs a legal basis; minors need extra care (data protection impact assessment for school-wide deployments, authority approval in German Länder).
• Copyright: uploading publisher textbooks to AI tools may breach licences; check school/university licences and the text-and-data-mining exception (EU DSM Directive Art. 3-4). Output copyright is uncertain; prefer openly licensed (CC BY) sources.
• Quality: hallucinated facts and bias; mitigate with grounding, reviewer checklists and subject expert sign-off.

Getting started

1. Pick one subject or course and list learning objectives and the vetted sources to use.
2. Test two Path A tools on a real lesson; compare time saved and the number of factual corrections.
3. Define a review checklist (accuracy, level, bias, licence, accessibility) and an AI-use note for learners.
4. Check data protection: contract, EU hosting, no pupil data in prompts, works council or school authority approval.
5. If volume grows, add retrieval over your own library and LMS export (H5P/SCORM).

Sources

1. Schools Week - Meet Aila, Oak Academy's new AI assistant - https://schoolsweek.co.uk/meet-aila-oak-academys-new-ai-assistant-2/
2. Oak National Academy - Free, AI-powered lesson assistant - https://www.thenational.academy/ai
3. Moodle Docs - AI subsystem - https://docs.moodle.org/en/AI_subsystem
4. Oak National Academy - oak-ai-lesson-assistant (GitHub) - https://github.com/oaknational
5. Regulation (EU) 2024/1689 (AI Act), Art. 4, 5, 50 and Annex III - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. Oak National Academy - Building AI that's safe for the classroom - https://www.thenational.academy/blog/building-ai-that-s-safe-for-the-classroom-what-we-have-learned-with-aila