AI in TVET and Employment Promotion
Applying AI sensibly – for people, data and learning
Otto von Guericke University Magdeburg · Engineering pedagogy and technical education
Navigate with ← → or the buttons below · N speaker notes · F full screen · How this keynote was built with AI → · Overview & LiaScript versions
AI is already here. The question is how we use it.
AI arrives quietly
- Spell-check and translation on your phone
- Voice messages turned into text
- Chatbots answering customer questions
- Route planning for delivery drivers
- CVs sorted by recruitment platforms
AI, in plain words: software that has learned patterns from many examples – and uses them to suggest, predict, sort or write.
It does not understand like a person. It can be confidently wrong.
Where have you met AI this week?
Apply, don't build
Building big AI models costs huge amounts of money, data and energy – and only a handful of companies do it.
We do not need to build the engine – we need to train good drivers and build good roads.
A new way to talk to computers
From clicking, to coding, to simply saying what you need.
Click → Code → Prompt
Same task: “Which apprentices missed more than 3 days this month?”
Why this is empowering
Plain language becomes the interface. A trainer, a technician or a labour-market analyst can now describe what they need – without being a programmer.
But: you still have to know your trade to judge whether the answer is right.
Offline AI or online AI – where does your data go?
Measured on this very keynote: offline draft 46 min · cloud draft 4 min · offline draft + cloud polish ≈ 64 % fewer cloud tokens. See the numbers →
The AI never touches the computer
A language model only writes text – sometimes that text is code or a command. A program around it, with permissions you set, decides what is actually done.
Where does our labour market data come from?
From job ad to decision
job ads, surveys, registers
remove duplicates, map to skills
trends, skill gaps
guidance, curricula, policy
Whose data is it?
Job-portal data often sits with private platforms abroad.
people agree – and data is used only for what they agreed to
a system that learned from past hiring can repeat past discrimination
AI for hiring, access to training and assessing learners is high-risk
One option: run open AI models on local servers – the data never leaves the country. (Remember the offline/online switch.)
The informal sector – the blind spot
Imagine a plumber in Nairobi: skilled, busy, trusted by the neighbours. No CV online. No LinkedIn profile.
Share of workers in informal employment. Source: ILO, Women and Men in the Informal Economy (2018) – check for the latest figures.
If it is not in the data, it is not in the decision.
How should TVET prepare technicians for machines that contain AI?
The AI suggests. Grace decides.
Grace is a water technician. The leak-detection system sends her a work order. You are Grace – what do you do?
Same pattern elsewhere: a solar inverter “underperforming”, a camera that flags “faulty parts” on a production line. Illustrative cases.
What changes for skilled workers
Skilled workers now work with the system:
- Read its suggestions: does this make sense?
- Notice when it is wrong
- Feed it good data, look after the sensors
- Report problems to the right person
| Level | What it means |
|---|---|
| 1 Use | operate AI-supported tools safely |
| 2 Understand & check | know what it can and cannot do, spot errors, protect data |
| 3 Maintain & adapt | keep sensors accurate, update data, find and fix faults |
Built into each trade – water technician, electrician, mechatronics technician. Not a separate IT course.
Which tasks change? Sort Grace's week.
Click each task to cycle: AI can do it AI helps people do it
The Job-Futuromat (Institute for Employment Research, IAB) shows which tasks of an occupation could already be automated today – based on German occupations, in German.
How can AI make learning personal, self-paced and open?
Learning that fits the learner
Scripted demo – no AI is called.
AI is a co-pilot – not an autopilot
“Ohm's law says voltage = resistance ÷ current (U = R / I). So a 46 Ω heater on 230 V draws 5 A.” ✅ I'm confident this is correct.
| Teachers keep | Learners need |
|---|---|
| setting goals | critical thinking: “Is this correct? How do I check?” |
| judging quality | problem solving |
| relationship, motivation | communication and teamwork |
| fair assessment | responsibility |
Risks: confident wrong answers · copying instead of learning · unequal access · privacy of young learners
If teachers are not confident with AI, learners will not be either.
Where should GIZ invest?
simple phones, messaging, SMS, voice, local languages
teachers, trainers, analysts, employment services
partners own their data; open source, open formats, local servers where it matters
in data and in training
pilots, then scale what works
permissions, checking, human approval – the AI only writes, people decide
Three take-aways
AI is a tool for real problems – like getting Amina into the right job.
Technicians like Grace and teachers like Sara who can work with AI – and question it.
Local ownership, privacy, and the informal sector in the picture.
Where will you start?
AI will not replace good TVET – but TVET that uses AI well will shape who benefits from it.
Thank you!
Hannes Tegelbeckers · Otto von Guericke University Magdeburg
Play with it – after the talk
IAB: which tasks of your occupation can be automated? (German) Will Robots Take My Job? ↗
Automation risk estimates for US occupations Teachable Machine ↗
Train your own small AI with your webcam – “learning from examples” in 2 minutes Quick, Draw! ↗
Doodle and watch an AI guess – and fail ESCO ↗
The EU's shared list of skills and occupations Cedefop Skills-OVATE ↗
Skills demand from millions of online job ads LM Studio ↗
Run an open AI model offline on your own laptop Ollama ↗
Open models on your own server – the “offline AI” of this talk Behind the scenes →
This keynote built three ways: offline, cloud, combined – with tokens and cost
Also as LiaScript course (with narration and in-browser polls): open in LiaScript ↗ · Source: GitHub ↗