How Edaptic's AI systems work
This page describes every AI system in Edaptic — what it does, how it makes decisions, what it can and can't do, and how humans stay in control. Required under the EU AI Act for high-risk AI systems.
The AI systems, one by one
Edaptic uses several AI-related systems. Five are described first — the AI tutor, the AI grader, the Evidence-Based Mastery tracker, the rule-based risk indicator, and the adaptive recommendation engine. Two further optional systems are described after them: a speech-to-text system that only runs when your course turns on voice input, and an AI governance layer that only runs when your institution turns it on. Each is described with its purpose, decision logic, known limitations, and the human oversight in place.
AI tutor
The AI tutor is the live, conversational AI you interact with while studying. When you open a slide, you can request an "Intuition" or "Technical" explanation — the system first checks whether your instructor has already prepared and approved an explanation for that slide. If so, that instructor-curated content is shown (with a ✅ Human Approved badge). If not, an explanation is generated on the fly and shown with a 🤖 AI Draft badge, indicating it has not yet been instructor-reviewed.
After the initial explanation, any follow-up questions you type in the chat are answered in real time by a large language model. These live responses are AI-generated without per-message pre-approval. The persistent notice at the top of the chat identifies the model being used.
How it decides what to say: A system prompt is constructed dynamically from the current slide, your instructor's configured persona, your audience level, and your domain. This prompt, together with your message, is sent to the configured large language model (such as GPT-4o, Claude, or Gemini). The model generates a response token by token. It does not access the internet, your grades, or other students' data during this process.
- ⚠ Live chat responses can be incorrect or misleading — always verify important claims against course materials or your instructor's approved explanations
- ⚠ Has no memory between sessions — each new session starts fresh
- ⚠ Cannot access the internet, your grades, or other students' data
- ⚠ An explanation shown as "🤖 AI Draft" has not yet been reviewed by your instructor — treat it with appropriate caution
- ⚠ May reflect biases present in its training data
AI grader
When you submit a written answer on an assignment, the AI grader evaluates it against the question and a rubric set by your instructor. It produces a numeric score and written feedback explaining the score.
How it decides your score: Your answer, the question text, the model answer or key concepts, and the rubric are sent to the grading model. The model returns a score (an integer between 0 and the question's maximum points, typically 10) and a justification in a structured format. The score is type-checked and clamped to the valid range before being stored. If the model fails or returns malformed output, the system retries up to two more times before marking the question as "grading failed".
- ⚠ May misinterpret creative, unconventional, or multilingual answers
- ⚠ Cannot evaluate diagrams, handwritten work, or uploaded images
- ⚠ Rubric interpretation depends on the model — different models may produce slightly different scores for the same answer
- ⚠ A written answer counts as correct for the mastery calculation when the AI score is at least 60% of the question's max points — this threshold is fixed and may not suit all question types
Mastery tracker
After every quiz or assignment answer, the mastery tracker updates a score for each concept tested. This is not an AI model — it's a transparent weighted ratio of correct evidence over total evidence per concept. The full specification lives in doc/evidence_based_mastery_model.tex.
How it calculates your mastery: Each answer becomes one piece of evidence linked to the canonical topics it tests. Your score for a concept is the weighted fraction of correct evidence, with recent work counting more than older work.
Recency decay is piecewise-linear and explainable in one sentence: evidence newer than 30 days counts in full, then tapers linearly to half weight over the next 90 days, and stays at half weight after that. Older work still counts — just less.
The score is paired with a band (Needs support 0–39, Developing 40–69, Almost secure 70–84, Secure 85–100) and a confidence level (None / Low / Medium / High) derived from how much evidence has accumulated. A 90% score on three questions is correctly tagged as low-confidence.
- ⚠ Assumes each answer is independent — doesn't account for fatigue, stress, or contextual factors
- ⚠ Treats written and report answers as correct for the mastery calculation if the AI grader scores them ≥ 60%, which folds grader uncertainty into the score
- ⚠ A small number of recent answers can move the score quickly when there is little prior evidence — the confidence level flags this
- ⚠ Source weights and decay are course-wide rather than personalised per student
Risk indicator
The risk indicator is an advisory label shown to your instructor that estimates whether you might benefit from additional support. It is not an AI system — it's a rule-based engine that evaluates a small set of explainable rules against your Evidence-Based Mastery state.
Weak-skill threshold: a topic is "below threshold" when its mastery score is under 70%.
Medium-risk rule: fires when three or more required topics are below the threshold.
High-risk rule: fires when at least three required topics are below the threshold and at least two of those skills are needed for the next assignment.
When a rule fires, the engine records the rule that triggered it (main reasons) and a list of explainable next steps (recommended actions) such as "review Linear Regression" or "complete the next practice set". Both are surfaced to your instructor; nothing is hidden behind a black-box score. These labels are advisory — they do not trigger any automated academic consequence.
- ⚠ Does not account for extenuating circumstances (illness, part-time study, late enrollment)
- ⚠ A student with very little evidence may be flagged simply because there's no signal yet — the confidence level on the underlying mastery is the better indicator early in a course
- ⚠ Default thresholds (70% mastery; three weak skills; two required for the next assignment) may not suit every course structure equally
Adaptive recommendation engine
The adaptive engine suggests personalised settings for your learning experience: difficulty level, teaching style, and learning persona. These are suggestions — you can override all of them on your Preferences page.
How it decides: Three signals feed the recommendation, ordered by influence: your assignment mastery scores (primary), your session behaviour such as message length and question frequency (secondary), and your feedback ratings on AI responses (a small confirmation signal). The output is a recommended audience level and teaching persona, not a binding decision.
- ⚠ Early in a course, with limited data, recommendations may be generic or inaccurate
- ⚠ Infers learning style from behavioural signals which may not reflect your actual preference
Speech-to-text (optional)
If your instructor turns on voice input for your course, you can speak instead of type. Your short recording is sent to OpenAI's Whisper service, which converts the speech to text. Only the text is kept — it becomes your answer, exactly as if you had typed it; the audio recording is not stored. This feature is controlled at two levels: your instructor must turn it on for your course, and the platform operator must allow it for the whole deployment. An operator that keeps all data inside the EU/EEA can switch it off platform-wide, in which case it is unavailable regardless of any course setting.
whisper-1), servers in the United States- ⚠ Audio cannot be cleaned of personal identifiers the way typed text is — so this feature is off unless your course enables it and the operator allows it platform-wide
- ⚠ Transcription can mishear accents, background noise, or technical terms
- ⚠ Edaptic does not store the recording; it is transcribed and discarded
AI governance layer (optional)
If your institution switches AI governance on, a copy of the tutor prompt and the tutor's response — with the same personal-data placeholders already applied as for any other AI request — together with a pseudonymous account reference and your course identifier is sent to TrustLayer. TrustLayer runs policy checks and keeps an audit record. It does not generate any content you see and it does not change what the tutor says to you. This is off by default.
- ⚠ Off by default; only active if your institution turns AI governance on
- ⚠ When the operator sets it to "enforce" mode, all tutor traffic — including EU-hosted providers — is routed through TrustLayer, so its hosting region then determines where processing happens
How AI content is marked
You will always be able to distinguish AI-generated content from human-authored content. Here's the marking system:
| Where | Marking | What it means |
|---|---|---|
| Chat interface | 🤖 AI notice | A permanent strip at the top of every chat showing the AI model name and linking to this page. Cannot be dismissed. |
| Slide explanations | 🤖 AI Draft or ✅ Human Approved | "🤖 AI Draft" means the explanation was generated by AI and has not yet been reviewed by your instructor. "✅ Human Approved" means your instructor confirmed accuracy via a mandatory checkbox in the approval modal. |
| Assignment feedback | 🤖 Graded by AI | Badge on each AI-graded answer. Plus a pre-submission disclosure banner (EU AI Act Art. 22): "Written answers were graded by an automated AI system." This appears before you submit any written assignment. |
| Quiz questions | AI-generated | Questions were generated by AI. Your instructor must confirm review in the "Confirm & Publish" modal — ticking "I have reviewed all questions in this assignment and confirm they are accurate and appropriate for students" — before students can see them. |
| Mastery scores | Calculated | Not AI — a weighted ratio of your correct evidence over total evidence per topic (Evidence-Based Mastery) |
| Risk indicator | Calculated | Not AI — produced by a rule-based engine over your mastery state. Visible to your instructor only. |
Human oversight at every decision point
The EU AI Act requires human oversight for high-risk AI systems. Here's where humans are in the loop in Edaptic:
Quiz and assignment questions
Instructors must confirm review in the "Confirm & Publish" modal — ticking "I have reviewed all questions in this assignment and confirm they are accurate and appropriate for students" — before students can access any assignment. Assignments can also be saved as drafts for staged review before publishing.
Slide explanations
AI-generated explanations show a "🤖 AI Draft" badge until the instructor explicitly approves them. The "✅ Human Approved" badge only appears after the instructor checks "I have reviewed this AI-generated content and confirm it is accurate and appropriate for students" in the Confirm & Approve modal.
Concept taxonomy
The AI proposes a concept structure. The instructor must review, edit, and explicitly apply it — including a safety confirmation for replacements.
AI grading
Every AI-graded answer shows a "🤖 Graded by AI" badge. Students can click "Request Human Review" on any answer. Instructors have a dedicated review queue with per-question score and feedback override. Instructor-overridden grades always replace AI scores.
Risk indicators
Risk labels are advisory only. No automated academic actions are taken. The instructor interprets and acts on the signal using their own judgment.
Adaptive settings
All adaptive recommendations can be overridden by the student in their Preferences page. The system respects manual overrides.
Which AI provider sees your data?
Edaptic doesn't host its own large language models. When you chat with the AI tutor, when an assignment is graded, or when an instructor generates content, the request is sent to one of the platform's configured providers. For these chat, grading, and authoring features, the set of providers your deployment can reach is controlled by the platform admin via the API-key configuration; an API key that isn't configured cannot be reached at all.
One further OpenAI service sits slightly outside that key-controlled set today: the optional voice transcription described above (whisper-1). It is listed in the table below for completeness.
| Provider | Where inference runs | EU/EEA? |
|---|---|---|
| OpenAI (GPT) | United States | No (EU-US Data Privacy Framework) |
| Anthropic (Claude) | United States | No (EU-US Data Privacy Framework) |
| Google (Gemini) | United States | No (EU-US Data Privacy Framework) |
| Mistral | France | Yes |
| NTNU HPC (Mistral / Qwen on-prem) | Norway | Yes (EEA) |
OpenAI Whisper (whisper-1) — voice transcription, only when your course enables it | United States | No (EU-US Data Privacy Framework) |
Before any of these providers receives user-typed text, the platform runs a PII scrubber that replaces categorical personal data (email addresses, phone numbers, Nordic national identifiers, credit-card-shaped numbers) with typed placeholders. So even when a request is sent across the Atlantic, the recipient does not see those specific identifiers.
You can see which provider made any individual decision affecting you by asking your instructor for the decision trace for that response — the audit log records the provider name on every AI grade.
Your right to contest AI decisions
Under both the EU AI Act and GDPR, you have the right to understand and challenge decisions that affect you. Here's how:
You can always request human review
No AI-generated score, label, or recommendation in Edaptic is final. You have specific mechanisms to contest each one:
- ✓ AI-graded assignment: Click "Request instructor review" on any graded answer. Your instructor will re-evaluate it personally.
- ✓ Mastery score: Each score is auditable — your instructor can show you the evidence events that fed into it. If you believe a piece of evidence is wrong (for example, a misgraded written answer), request human review on that answer; the override deletes the affected evidence and replays the corrected verdict through the mastery pipeline.
- ✓ Adaptive recommendations: Override any recommendation in your Preferences page. Your manual settings always take precedence.
- ✓ Risk indicator: If you're aware your instructor has flagged you as at-risk and you believe the classification is wrong, contact your instructor directly. The label is advisory and carries no automated consequences.
- ✓ AI-generated content: If you find an error in an AI-generated explanation or quiz question, use the feedback button to report it. Your instructor will be notified.
Questions about AI in Edaptic?
Contact the development team
If you have questions about how any AI system works, want to understand a decision that affected you, or want to report an error, contact us at adil.rasheed@ntnu.no.
For GDPR-specific data rights (access, deletion, portability), see the Your data in Edaptic page.