Decide when and how to use AI responsibly
Does AI add value to this task without replacing the learning I need to do?
Start with the learning goal and the applicable rule, keep the important thinking visible, and finish by proving the understanding without the tool.
In brief
- For assessed work, explicit task and examiner instructions come before general advice.
- A better immediate artefact is not automatically evidence that you learned more.
- Use bounded assistance, preserve checkpoints, and test a comparable task independently.
Use this decision path before opening a tool
- Name the outcome. What must you know, practise, explain, judge, or create by the end? “Finish the task” is not yet a learning outcome.
- Find the controlling rule. For assessed work, check the assignment, module information, applicable FPSO/APSO, Moodle announcements, and examiner or supervisor instructions. If the use is not clearly permitted, ask first.
- Protect the target competence. Identify the reasoning, formulation, coding, analysis, or judgement you need to perform yourself. Do not delegate the evidence the task is meant to collect.
- Classify the input. Remove personal, sensitive, confidential, examination, participant, patient, unpublished research, and unauthorized teaching material before any prompt.
- Divide the work. Give the system a bounded role, such as questions, critique, alternatives or a small draft. Keep decisions and verification with you.
- Set checkpoints. Decide in advance what you will verify, how you will record material assistance, and what comparable task you will complete without AI.
Separate assisted performance from learning
Several recent studies found that AI support could improve the immediate product or practice result without a corresponding gain in independent knowledge or transfer. Their settings were narrow and short-term, so they do not prove that AI always harms learning.
The durable lesson is the distinction itself: ask what improved. Was it the artefact, the speed, an immediate test, an unaided transfer task, delayed retention, or only the learner’s impression?
Offloading is not automatically bad. A calculator can free attention for higher-level reasoning. The problem is offloading the very operation you are supposed to learn, especially before you can recognize a plausible error.
Choose a workflow that keeps you active
Scroll horizontally to view all columns.
| Your task | A bounded AI role | Your independent checkpoint |
|---|---|---|
| Understand a concept | Ask for questions, hints, competing explanations, or a misconception check after your own attempt. | Explain it from memory and solve a new example without the tool. |
| Develop an argument | Generate counterarguments or ask for assumptions you may have missed. | Choose the position, verify evidence, and write the reasoning yourself. |
| Revise writing | Request feedback against named criteria rather than a replacement text. | Accept or reject each change and ensure the final wording still expresses your judgement. |
| Find literature | Use AI to expand search terms or discover candidate sources. | Search suitable databases, open every original publication, and build the evidence trail yourself. |
| Work with code | Ask for a bounded diff, test cases, explanation, or debugging hypothesis. | Inspect every change, run checks, test edge cases, and explain the final behavior. |
Select a route, not a brand
A product name says little about whether a particular use is appropriate. Evaluate the exact account, model, configuration, connectors, and task.
Fit to the task
Does the route explain, retrieve, calculate, execute code, or work over a selected corpus? What evidence can you inspect?
Access and continuity
Will every learner have equitable access for the full activity, without being pushed into an individual paid account?
Data conditions
What is retained, used for training, reviewed for abuse, visible to administrators, or sent to connected services? These are different questions.
Verification burden
What could fail quietly, and do you have the subject knowledge, original evidence, tests, or human support needed to detect it?
Non-AI alternative
Could office hours, peer explanation, official documentation, a database, debugger, reference manager, or manual workflow serve the goal better?
Try a complete learning task
For a concept you are revising, explain it from memory first. Ask the system for a question that could reveal a mistaken assumption, answer it yourself, and check its feedback against your lecture material. Finish with a new example without AI. A fluent explanation you can recognize is easier than an explanation you can produce.
The worked learning workflows include literature notes, code review, a false proof and data analysis. Use the two-version delegation exercise to compare a broad request with a precise specification on the same synthetic dataset. Compare errors and the checking effort, not just how polished the answer looks.
Before submitting any of these outputs, return to verification and declaration. For a course-specific tutor, check Iris and OneTutor access. For input you cannot share with a cloud service, follow the local-model and data guide.
Pause before sharing the input
TUM’s public guidance is conservative even for institutionally offered tools: do not enter personal, sensitive, or confidential information. Institutional access is not permission for every data class.
- Remove names, student numbers, email addresses, grades, credentials, and health or disability information.
- Do not submit participant, patient, examination, disciplinary, confidential, or internal TUM information.
- Do not upload unpublished research, third-party confidential material, or protected code without an authorized route.
- Ask before uploading lecture notes, slides, recordings, previous examinations, or another person’s work.
- Use the minimum context needed and disable unnecessary connectors, browsing, or file access.
Not using AI can be the better decision
- The rules are unclear and the responsible examiner has not answered.
- The task exists to practise the exact thinking the system would perform.
- You cannot verify the result or recognize a plausible failure.
- The necessary input is sensitive, confidential, protected, or not yours to share.
- A conventional source or tool is more direct, reproducible, or trustworthy.
- You need evidence of what you can do independently.
Sources and scope
These sources support the guidance on this page. Product and institutional details can change; use the linked official information for current conditions.
- General Academic and Examination Regulations — readable German consolidation (opens in new tab)
Technical University of Munich · Through the 12th amendment, effective 1 January 2026
Readable consolidation; the officially promulgated German text controls and the applicable FPSO supplements it
Official TUM statutes and regulations database (opens in new tab)
- Using AI Wisely: Guidance for Students and Faculty (opens in new tab)
TUM ProLehre | Media and Didactics · Living guidance; checked 30 July 2026
TUM teaching and learning guidance; not legal advice
- Wie entscheide ich, wofür ich den Einsatz von KI-Tools in meiner Lehre erlaube? (opens in new tab)
TUM ProLehre | Media and Didactics · Version 2, 9 March 2026
German teaching decision aid; not an examination statute
- TUM Citation Guide (opens in new tab)
TUM University Library · June 2026
TUM-wide advisory citation guidance; not examination law
- Empowering Learners for the Age of AI (opens in new tab)
OECD / European Union · 2026
Non-binding AI-literacy framework; adapted for higher education
Licence: Creative Commons Attribution 4.0 International (opens in new tab)
This is an adaptation of an original work by the OECD and the European Union. The opinions expressed and arguments employed in this adaptation should not be reported as representing the official views of the OECD, its Member countries or the European Union.
- OECD Digital Education Outlook 2026 (opens in new tab)
OECD · 2026
Evidence synthesis; findings remain task- and study-dependent
- Generative AI without guardrails can harm learning (opens in new tab)
Proceedings of the National Academy of Sciences · 25 June 2025
One short-term high-school mathematics field experiment
- Beware of metacognitive laziness (opens in new tab)
British Journal of Educational Technology · 10 December 2024
One randomized university writing experiment with 117 students
- Copyright guidance (opens in new tab)
TUM University Library · Living guidance; checked 30 July 2026
German advisory guidance; not a licence to upload teaching material