Using AI responsibly¶
online:academy uses AI features — the AI learning assistant (tutor), the AI-assisted grading of free-text answers, and the AI quality check of courses. On this page we explain what the AI can do, where its limits are, and what responsibility stays with you. The page is aimed at lecturers, content managers and company administrators.
In addition, we point out that companies that use AI systems must take measures to ensure, to the best of their ability, that their staff who are involved in the use of AI systems have a sufficient level of AI literacy (Art. 4 EU AI Act).
Why this matters¶
Grading in education counts as a high-risk use under the EU AI Act. That does not mean AI grading is forbidden — but the final decision has to lie with a human, and learners have to know that an AI was involved. That is exactly what the features below are for.
How our AI works¶
- Self-hosted, in Germany. The language model runs on our own server in the Hetzner data center. No answers from your learners go to OpenAI, Google, or other external services.
- No training with your data. The model does not learn from the answers — it only grades them.
- Logged. Every AI call is logged in an audit-proof way (without storing the plain text, only a cryptographic fingerprint).
The three features and their limits¶
1. AI grading of free-text answers¶
The AI proposes a score and feedback. That is a proposal, not a grade.
- ⚠️ The AI can be wrong — with answers that are factually correct but unusually phrased, with irony, with non-native speakers, or with very terse answers.
- ✅ Your job: Review the AI gradings under "AI gradings" in the menu. You can filter by score range (e.g. all borderline cases at 60-70 %) and approve or override each grading (change the score + feedback).
- 🔒 For certificate-relevant courses you can enable "AI gradings require approval by an authorised person" in the course. Then no AI grading counts towards completion before you have approved it.
Rule: Don't wave gradings through unchecked. The most common mistake is to trust the AI blindly — that is exactly what the approval step is meant to prevent.
2. AI learning assistant (tutor)¶
A chat assistant that helps learners with questions of understanding.
- It is labelled as AI; learners see a notice the first time they open it.
- It does not grade and makes no decisions.
- Answers can be inaccurate or made up ("hallucination"). Point out to your learners that they should double-check important statements.
- Conversations are private — you as a trainer do not see them.
3. AI quality check of courses¶
Evaluates only the structure and metadata of your course (title, description, lesson setup) — not the media quality of videos or embedded content. Use it as a source of hints, not as a verdict.
What you owe your learners¶
- Transparency: The labelling happens automatically — don't undermine it (e.g. by passing off AI feedback as your own).
- Right to human review: If learners object to an AI grading, look at it and correct it where necessary. That is their right.
- Data minimisation: Ask learners not to give sensitive personal data in free-text answers.
Dealing with poor AI gradings¶
- Systematically poor/biased AI gradings (e.g. always worse for certain groups) → report them to the generation:l team, with examples and screenshots.
- Inappropriate or wrong tutor answers → report those too; we will check the model.
In a nutshell¶
| The AI proposes | The human decides |
|---|---|
| Score + feedback | You approve or override |
| Answer in the tutor chat | Learners check important statements |
| Quality score | You judge the course |
The AI takes work off your hands — the responsibility stays with you.