KI verstehen
Chapter 1 of 7
Chapter One

Hello —
so glad
you're here. AI,
properly
explained. AI in the
classroom. — Artificial Intelligence explained simply and properly

Today I'll explain how Artificial Intelligence — AI for short — actually works.

Artificial Intelligence from the mathematical foundations to the modern Transformer architecture — structured across seven chapters.

Hi class! This page explains in 7 steps what Artificial Intelligence is — with learning objectives, discussion questions and exercises for class.

Promised: no complicated jargon. You just click your way through step by step.

With terminology, sources and technical depth — for self-learners and developers. Switch back to Simple any time.

Suitable from middle school upward. Tip for the teacher: each chapter has a blue box at the end for discussion or homework.

~ Make yourself comfortable.
~ Subheadings, sources and technical boxes where helpful.
~ Recommended lesson unit: 2 × 50 min.
Chapter Two

What is AI
really?

What is AI —
technically
speaking?

What is AI?
A definition
for class.

Imagine someone who has read almost every book in the world. Millions of books, newspapers, letters — just everything.

Artificial Intelligence is an umbrella term for systems that solve tasks normally requiring human intelligence. Today it almost always means Machine Learning — and especially Deep Learning on large neural networks.

AI are computer programs that learn patterns from data and use them to solve tasks — write text, recognise images, translate languages. They aren't thinking beings, but very good pattern-recognisers.

That "person" can then answer questions, write stories or recognise images. Not because they really think, but because they've seen incredibly much.

Current Large Language Models (LLMs) like GPT-4, Claude or Gemini use the Transformer architecture (Vaswani et al., 2017, "Attention Is All You Need"), have billions to trillions of parameters and are trained on corpora with hundreds of billions of tokens.

AI means Artificial Intelligence — "artificial knowledge" a computer has learned.
At their core LLMs compute a probability distribution over the next token. They don't "understand" in the human sense — they model language statistics with impressive depth.
Terminology hierarchy: AI ⊃ Machine Learning ⊃ Deep Learning ⊃ Transformers ⊃ LLMs. Each inner term is a special case of the outer one.
📚 Learning objectives
  • You can distinguish AI from normal software.
  • You explain the term "Machine Learning" in your own words.
  • You understand that AI computes the most likely answer — it doesn't "think".
📖 Key terms
  • AI (Artificial Intelligence): Umbrella term for programs that solve tasks normally requiring human intelligence.
  • Machine Learning (ML): Subset of AI in which computers learn from data — instead of being explicitly programmed.
  • Algorithm: Clear step-by-step instruction, like a cooking recipe.
  • Neural network: Program structure loosely inspired by the wiring of nerve cells.
  • Training data: The examples an AI learns from (images, texts, sounds…).
💡 Did you know…

The first AI program Logic Theorist dates from 1956 (Newell & Simon, Carnegie Mellon). AI is therefore older than moon landings, the internet and smartphones — it only really "exploded" around 2012 when data and compute became big enough.

❓ Quiz
Which description fits AI best?

Answer C: "A program that learns patterns from examples."

A would be a classic program (e.g. a calculator). B describes a human brain. Only C captures Machine Learning.

Tip for the teacher: Have the class guess the three options first — A: "A computer that thinks like a human." / B: "A collection of nerve cells with electrical impulses." / C: "A program that learns patterns from examples."

🎯 Extended learning objectives (Bloom's taxonomy)
  • K1 — Knowledge: Pupils name 3 examples of AI in daily life.
  • K2 — Comprehension: Pupils explain the difference between "classical program" vs "Machine Learning" in their own words.
  • K3 — Application: Pupils classify everyday applications into one of the two categories.
  • K4 — Analysis: Pupils discuss what distinguishes AI from human thinking.
⏱ Timing for this chapter (≈ 15 min)
  • 2 min: Read the lead text aloud together.
  • 3 min: Collect key terms on the board (with pupil guesses).
  • 2 min: "Did you know…" fact — gather quick reactions.
  • 5 min: Quiz in small groups (2-3 pupils). Note guess before revealing.
  • 3 min: Discussion: "Impressive vs. concerning".
💬 Leading the discussion

Question: "What do you find impressive about AI? What worries you?"

Expected pupil answers:

  • Impressive: "Recognises faces", "Translates instantly", "Can draw pictures"
  • Worrying: "Takes jobs", "Makes fake news", "Knows everything about me"

Follow-up question: "Which concern is justified, which is exaggerated? Argue your case."

❓ Quiz solution with pedagogical explanation

Answer: C — "A program that learns patterns from examples."

Why this question matters: Pupils often confuse AI with a "digital brain" (option B) or with "very clever software" (no learning). The quiz checks whether the core concept "learning from data" was understood.

Common misconceptions:

  • "AI is like a human, just digital" → No, no consciousness.
  • "AI knows everything" → No, only what was in the training data.
  • "AI thinks" → No, it computes probabilities.
🤔 Anticipated pupil questions
  • "Is Alexa AI?" — Yes, a speech AI (see chapter 4).
  • "Can AI have feelings?" — No. It can simulate emotion but not experience it.
  • "Will AI replace us all?" — Some tasks yes; this deserves its own discussion (chapters 6+7).
  • "What's the difference from robots?" — Robot = physical machine, AI = software. Some robots use AI, many don't.
🖨 Mini worksheet (5 tasks for the class)
  1. Define "AI" in at most 2 sentences.
  2. Name 3 examples of AI in your daily life.
  3. Explain the difference between a calculator and an AI.
  4. What does "Machine Learning" mean? Explain with one example.
  5. Which concern about AI matters most to you personally? Argue your case.

Model answer for task 1: "AI is a computer program that learns patterns from data and uses them to solve tasks (e.g. write text, recognise images)."

✅ Measuring learning success
  • Quick check (1 min): "Who can explain AI in 10 words?" (Hands up)
  • Homework review: Did pupils name 5 different everyday examples?
  • Depth of understanding: Can pupils name the difference "programmed" vs "trained"?
← Back to the start
Chapter Three

How does an
AI learn?

How is a model
trained?

How does AI learn?
Like you.

Quite like a small child. You show the child a cat often enough, and eventually it recognises cats everywhere — even in a book or as a sculpture.

Modern Machine Learning typically runs in three phases: Pre-Training (unsupervised, huge corpora), Supervised Fine-Tuning (curated examples) and RLHF — Reinforcement Learning from Human Feedback for safety and alignment.

When you cram vocabulary, you repeat words until they stick. An AI does the same — only with billions of examples instead of 30 words.

For an AI it's the same — only it looks at millions of images until it knows for sure.

Mathematically the network adjusts its parameters via Backpropagation and Gradient Descent so that a loss function (e.g. Cross-Entropy) is minimised. Training large models costs millions of dollars in GPU/TPU hours today.

📚 Learning objectives
  • You understand that an AI is not programmed, but trained.
  • You explain why an AI needs millions of examples while a child needs only a few.
  • You name the three training phases: Pre-Training, Fine-Tuning, RLHF.
📖 Key terms
  • Training: The AI's learning process — many examples, many repetitions.
  • Parameter: The "tuning knobs" inside the network. GPT-4 is estimated to have 1,700,000,000,000 of them.
  • Loss: Measure of how wrong the AI still is. Goal: minimise loss.
  • Overfitting: When the AI memorises examples instead of generalising. Like a student who knows answers but doesn't understand them.
  • RLHF: "Reinforcement Learning from Human Feedback" — humans rate AI answers, the AI learns what is helpful/polite.
💡 Did you know…

GPT-4 has an estimated 1.7 trillion parameters. If looking at each parameter took one second, you'd need ~54,000 years. Training reportedly cost OpenAI more than USD 100 million in compute.

❓ Quiz
An AI was trained only on cats and dogs. Can it recognise a banana?

Answer: No — and this is important to understand.

An AI only knows what it has seen in training. A banana would either be "weird cat" or "weird dog" to it. It would confidently give a wrong answer — and that's the core of many AI problems: it doesn't know what it doesn't know.

🎯 Extended learning objectives (Bloom's taxonomy)
  • K1: Pupils name the 3 training phases (Pre-Training, Fine-Tuning, RLHF).
  • K2: Pupils explain why an AI needs more examples than a child.
  • K3: Pupils simulate AI learning with the cat trainer and reflect on the result.
  • K5 — Evaluation: Pupils judge whether "statistics = learning".
⏱ Timing (≈ 20 min)
  • 3 min: Read lead aloud + comparison "cramming vocabulary".
  • 5 min: Demonstrate the cat trainer on the projector (teacher clicks).
  • 5 min: Pupils try it themselves in pairs.
  • 5 min: Quiz + explanation of the banana example.
  • 2 min: Transition to discussion "Learning vs. statistics".
💬 Leading the discussion

Main question: "After training, an AI can't learn from new examples without being retrained entirely. Is what it does really learning?"

Teacher arguments to fuel debate:

  • For "learning": "It adapts to examples. So do we."
  • Against "learning": "We learn from experience. AI learns from statistics. Different."
  • Middle ground: "A new kind of learning, different from human learning."
❓ Quiz solution with explanation

Answer: NO — the AI would confidently say "weird cat" or "weird dog".

Why this question is crucial: It exposes the fundamental AI problem: AI doesn't know what it doesn't know. Unlike humans who can say "I don't know", an AI always gives an answer — even when it's wrong.

Application: For every AI answer on a new topic → check critically!

🤔 Anticipated pupil questions
  • "How many images has ChatGPT seen?" — Estimates point to 100+ million images in multimodal models, plus hundreds of billions of words.
  • "Can AIs train each other?" — Yes, this is an active research direction (synthetic data).
  • "Why is it so expensive?" — Data centres with tens of thousands of specialised chips (GPU/TPU). Power + hardware = millions.
  • "Can I train an AI myself?" — Yes, with Teachable Machine in the browser (see closing box).
🖨 Mini worksheet (5 tasks)
  1. Name the 3 training phases of a modern language AI.
  2. What happens during pre-training? Explain in one sentence.
  3. Why does an AI need millions of examples when a child needs 5?
  4. Imagine: An AI was trained only on dogs. What problems arise?
  5. Which of the three phases do you think matters most? Argue your case.

Model answer for task 3: AIs detect statistics in pixels; children combine prior experience ("four legs, fur, tail = animal") with the new example.

✅ Measuring learning success
  • Practice check: Could pupils solve the cat trainer correctly? Do they understand what's happening?
  • Oral explanation: Can they articulate the "programmed vs trained" difference?
  • Transfer: Can they apply the banana-quiz to other domains? (e.g. "AI for skin-cancer detection — what if it misclassifies a freckle?")
Example pipeline for image classification:
  1. Forward pass: image → layers → prediction "cat: 0.82"
  2. Compute loss: difference vs ground truth (label "cat: 1.0")
  3. Backward pass: gradient of the loss w.r.t. each weight
  4. Weight update: w ← w − η·∇L
Try it yourself:
Tap every image that shows a cat.
Correctly found 0
Attempts 0
Wrong 0
That's how an AI learns: through practice, many examples and patience.
← One step back
Chapter Four

You're probably
already using it.

AI in production:
where it hides.

AI in your daily life —
already there.

AI hides inside things you use every day. Tap a card to learn more.

Behind everyday apps sit highly specialised ML systems. Tap for the technical classification.

Every smartphone, every stream, every search — AI is everywhere, but invisible. Tap the cards and discover where it hides.

Car navigation

It finds the fastest route, knows traffic and suggests detours.

Tap for details Show less
An AI in the background learns from millions of trips when and where traffic builds up.
Combination of classic graph algorithms (A*, Dijkstra) and machine-learning models predicting travel times from live GPS data of millions of vehicles.

The weather app

Tells you whether tomorrow will be sunny or you need an umbrella.

Tap for details Show less
The AI compares old weather patterns with the current situation and turns that into a forecast.
Hybrid models: numerical weather prediction (NWP) like ECMWF combined with deep-learning models (e.g. GraphCast from DeepMind, 2023) delivering 10-day forecasts in under a minute instead of several hours.

Photos on your phone

You search "beach" and find all beach photos — without having sorted them.

Tap for details Show less
The AI has learned what is in the pictures — faces, animals, places.
Computer Vision: Convolutional Neural Networks (CNN) and Vision Transformers (ViT) recognise objects; multimodal models like CLIP (OpenAI, 2021) connect image and text for semantic search.

Alexa & Siri

You speak, it understands and replies — like a helper in the house.

Tap for details Show less
Speech AI turns your words into text, finds the answer and reads it back to you.
Pipeline: Automatic Speech Recognition (ASR, e.g. Whisper) → Natural Language Understanding → optional LLM → Text-to-Speech (TTS) with neural vocoders like WaveNet.

Translation

Google Translate turns English into Italian, Turkish or German in an instant.

Tap for details Show less
The AI knows countless sentences in many languages and finds the right translation.
Neural Machine Translation (NMT) on Transformer base; multilingual models like NLLB (Meta, 2022) translate between 200+ languages directly — no detour via English needed.

Spam in your inbox

Ads and phishing emails land in trash automatically — you don't do that yourself.

Tap for details Show less
The AI has seen millions of emails and recognises spam even before you do.
Classic: Naive Bayes or gradient-boosting classifiers on TF-IDF vectors. Today increasingly BERT-based encoders that recognise phishing in previously unseen wordings.

Search engines

Google sorts billions of web pages so the most helpful is on top — and now also summarises directly.

Tap for details Show less
The AI checks which pages match your question best — and sometimes writes a short summary itself.
Hybrid aus klassischem Ranking (RankBrain, 2015, lernende Embeddings), BERT-Ranking (2019, Sprachverstehen) und neueren multimodalen Modellen (MUM, 2021, T5-basiert) — plus LLM-Zusammenfassungen („AI Overviews", 2024). Architektonisch verwandt mit Retrieval-Augmented Generation.

Recommendations

Netflix, Spotify and YouTube suggest movies, songs or videos you might really like.

Tap for details Show less
The AI compares your behaviour with others' and recognises what you'll probably like.
Recommendation systems: collaborative filtering (Matrix Factorisation, ALS) + content-based embeddings (Two-Tower models) + deep learning for sequence prediction (Transformer4Rec, SASRec).
📚 Learning objective

You recognise AI in familiar everyday applications and understand that "AI" often works invisibly in the background.

💡 Did you know…

Spotify uses over 100 signals for your daily mix playlist: what you listen to, skip, when you pause, time of day, even the weather. Netflix computes recommendations for 247 million subscribers every day.

❓ Quiz
Which of these apps uses NO AI?

Answer: A simple calculator.

Weather app, photo search, spam filter, translation — all use AI. A calculator runs fixed mathematical operations — no learning, no pattern recognition. But: scientific calculators with plot functions or step-by-step explanations may use AI!

🎯 Extended learning objectives
  • K2: Pupils explain where AI works visibly/invisibly in daily life.
  • K4 — Analysis: Pupils analyse what data an AI needs.
  • K6 — Evaluation: Pupils judge which of these AIs are indispensable / dispensable.
⏱ Timing (≈ 25 min — extendable)
  • 3 min: Introduce all 8 cards briefly (everyone clicks).
  • 5 min: Form groups of 4, each picks 1 card.
  • 5 min: Group research on the chosen card.
  • 10 min: 2-minute presentations from each group.
  • 2 min: Quiz + short reflection.
💬 Leading the discussion

Question: "Which of these 8 AIs would you miss, which not?"

Expected "would miss": Maps/navigation, photo search, voice assistants (older users), translation (language learners)

Expected "wouldn't care": Recommendations (some prefer self-discovery), spam filter (invisible)

Follow-up question: "What would happen without this AI? Who would do it instead?"

❓ Quiz solution with explanation

Answer: A simple calculator — not AI.

Why this matters: Helps pupils distinguish AI from "software with algorithms". Not every "intelligent app" is AI in the narrow sense.

Extended discussion: "Which apps on your phone likely use AI, which don't? Argue your case."

🤔 Anticipated pupil questions
  • "Are all games AI now?" — No. Chess/Go AIs yes, but most console/mobile games use fixed algorithms for "enemies".
  • "Is Google Maps an AI?" — Yes, the routing uses AI for traffic prediction.
  • "How does Spotify make money from recommendations?" — Listening time + subscriptions. Better recommendations = longer use = more money.
🖨 Mini worksheet (group work)

Per group of 4, pick one of the 8 cards:

  1. What data does this AI need to function?
  2. Which company is behind it, and who profits?
  3. Where was the biggest technical hurdle in its development?
  4. What if this AI didn't exist? (Describe a day without it.)
  5. Name one risk or concern about this AI.
✅ Measuring learning success
  • Presentation quality: Was it technically correct + clearly explained?
  • Reflection depth: Did the group justify a "would miss" statement?
  • Homework: The 9th AI application — creative, correctly recognised?
← One step back
Chapter Five

How you talk
to an AI.

Prompt Engineering —
how you steer the model.

Asking AI —
the art of the prompt.

You just type a question — like texting someone. Tap a question below and see what happens.

A prompt is the input to the model. The more precisely role, context and output format are defined, the more reproducible the answer. Try the examples.

A good prompt is like a good research question in the library. The more precise you ask, the better the answer. Try the examples below.

K
AI Helper
online now
Hi! Nice to meet you. You can ask me anything. Pick a question below.
Example questions to tap:
Prompts to test — mind role, context and format:
📚 Learning objectives
  • You formulate structured prompts with the 5 building blocks: role, context, task, format, examples.
  • You recognise why "I need help" gets less than a precise prompt.
  • You judge when AI use in school is sensible and when it is unfair.
📖 Key terms
  • Prompt: Your input to the AI — a question, task or instruction.
  • Token: Smallest unit the AI processes. "Banana" is 1 token, "Hippopotomonstrosesquipedaliophobia" is many.
  • Context window: How much text the AI can see at once. GPT-4 handles ~128,000 tokens — about a thick book.
  • Few-Shot: Including 1–3 examples in your prompt lets the AI learn the pattern for this answer.
  • System prompt: Instruction setting the AI's behaviour for the whole conversation ("You are a friendly maths tutor…").
📝 Prompt template to copy

Write a [FORMAT] about [TOPIC] for [AUDIENCE]. Style: [STYLE]. Length: [LENGTH]. Pay special attention to: [IMPORTANT].

Example: "Write a presentation about photosynthesis for an 8th-grade class. Style: simple, vivid. Length: 5 sentences. Pay special attention to: the role of chlorophyll."

💡 Did you know…

ChatGPT processes more than 100 million requests per day. If each took 30 seconds, that would be roughly 95 years of waiting time — all in parallel on huge data centres.

❓ Quiz
Which prompt gets the best result?

Answer: "You are a geography teacher for a 7th-grade class. Explain the water cycle in 4 steps with one example each. Format: numbered list."

Why? This prompt has role (geography teacher), audience (7th grade), task (explain), format (numbered list) and even structure (4 steps with an example).

"Explain the water cycle to me" would be OK but unstructured. "Help me with geography" would be vague — the AI has to guess what you want.

🎯 Extended learning objectives
  • K3 — Application: Pupils formulate working prompts with all 5 building blocks.
  • K4: Pupils analyse why vague prompts produce poor answers.
  • K5 — Evaluation: Pupils ethically evaluate AI use in homework.
⏱ Timing (≈ 20-25 min)
  • 3 min: Explain the 5 prompt building blocks (Role/Context/Task/Format/Examples).
  • 5 min: Try a sample chat on the projector — bad prompt vs. good prompt.
  • 10 min: Class experiment — every pupil formulates a prompt.
  • 5 min: Discussion on fair AI use in school.
💬 Leading the discussion

Question: "Should letting an AI write homework count as cheating?"

Teacher arguments for structured debate:

  • Pro cheating: "The work was the AI's, not yours → no own effort."
  • Pro not cheating: "You had to write the prompt — that's also work."
  • Middle ground: "AI as a brainstorming tool: OK. AI writing the whole essay: not OK."

Concrete class-rule discussion: "If you could use AI, what should be allowed, what not?"

❓ Quiz solution with explanation

Answer: The structured prompt with role (teacher), audience (7th grade), task (explain), format (list), structure (4 steps) is best.

Why: AI understands clear expectations and gives focused answers. Vague prompts → AI guesses what you want → imprecise answers.

Demonstration: Give the class the bad + good prompt — try both — show the difference.

🤔 Anticipated pupil questions
  • "Can I use AI for my presentation?" — Clarify class rules. Generally: AI as research OK, AI as writer problematic.
  • "How does the teacher know if it was AI?" — Hard to detect. Trust + fair rules > playing detective.
  • "Can AI remember conversations from yesterday?" — Within the same session yes, persistently usually no (ChatGPT with Memory feature: yes).
  • "Will AI make teachers obsolete?" — Very unlikely. AI can explain but cannot educate.
🖨 Mini worksheet (class experiment)

Topic (pick one): Photosynthesis / French Revolution / Pythagoras / Hamlet / Climate change

  1. Formulate a structured prompt using all 5 building blocks in 2 minutes.
  2. Swap prompts with your neighbour — whose is the most precise?
  3. Try (with teacher account) the 2-3 best prompts.
  4. Compare the answers — which produced the most useful output?
  5. Reflection: What made the difference?

Teacher preparation: Have a good + bad prompt ready for the demo.

✅ Measuring learning success
  • Practice test: Could pupils formulate a 5-block prompt?
  • Reflection: Do they understand why structure improves results?
  • Ethical maturity: Can they articulate their own position on fair AI use?
Good prompts have five building blocks:
  1. Rolle: „Du bist ein erfahrener Lektor…“
  2. Context: background, audience, prior knowledge
  3. Task: precisely what should be done
  4. Format: length, style, structure (e.g. Markdown table)
  5. Examples (Few-Shot): 1–3 sample answers
← One step back
Chapter Six

It's smart —
but not perfect.

Limits & risks
of current LLMs.

AI has limits —
and you should know them.

Important to know: an AI is not a human. It has no own thoughts or feelings. It can also be wrong.

Despite impressive capabilities there are fundamental limitations — anyone working with LLMs should know them.

AI is no magic teacher and no oracle. It makes mistakes, has biases, and often doesn't know current events. Critical thinking stays more important than ever.

No real feelings

No consciousness

Even if it sounds friendly — it isn't really happy, isn't sad. It just pretends.

LLMs simulate empathic language but have no experience, no self-model, no intrinsic goals. Anthropomorphising is tempting — and misleading.

It makes mistakes

Hallucinations

Sometimes it says things very confidently that are simply wrong. For important stuff, double-check.

Plausible-sounding but factually wrong outputs. Mitigation: RAG (retrieval), citation guards, lower temperature, output verification via tools.

Doesn't know the future

Knowledge Cutoff

It can't look into the future. Lottery numbers, fortune-telling or "what happens tomorrow" — it doesn't know.

Training data has a cutoff date. Current events require tool-use (web search, APIs) or retrieval. Without web access the model is "frozen".

Bias & distortion

Training data reflects societal biases. Gender, origin, occupation stereotypes — without active mitigation the model reproduces them.

Privacy

Inputs may — depending on the provider — be used for training. No secrets, no PII (names, addresses, credentials) into unknown models.

📚 Learning objectives
  • You name five concrete limitations of current AI systems.
  • You recognise a hallucination when you see one.
  • You understand why source criticism becomes more important in the AI age, not less.
📖 Key terms
  • Hallucination: A plausible-sounding but false AI statement. It doesn't know it's lying — it computes the most likely answer.
  • Knowledge Cutoff: The cutoff date of training data. Events after that are unknown to the AI.
  • Bias: When the AI inherits prejudices from training data.
  • Deepfake: AI-generated fake photos, audio or video.
⚠️ Real case study

USA 2023: Lawyer Steven Schwartz filed a court brief citing 6 court rulings — all invented by ChatGPT. The judge fined him USD 5,000 plus a professional reputation hit. When asked, ChatGPT had even assured him the cases were real — a textbook hallucination.

💡 Did you know…

Hallucinations occur in tests of large AI models in up to 15-25 % of answers — even with the best 2024 models. For specialty fields (law, medicine, rare facts) even higher.

❓ Quiz
An AI says: "Napoleon died in 1822 in Vienna." What do you do?

Answer: Check sources!

This statement is wrong. Napoleon died in 1821 on St. Helena. A hallucination can be a single wrong detail — year wrong, place wrong, both wrong. For any fact you don't already know: search engine + reliable source (encyclopaedia, official website). AI can be the start of your research, never the end.

🎯 Extended learning objectives
  • K1: Pupils name 5 limits of current AI.
  • K3: Pupils recognise a hallucination in a live test.
  • K4: Pupils analyse the Lawyer Schwartz case.
  • K5: Pupils develop their own verification strategies.
⏱ Timing (≈ 20 min)
  • 3 min: Read lead + explain terms Hallucination / Knowledge-Cutoff / Bias.
  • 5 min: Discuss case study Lawyer Schwartz — consequences!
  • 7 min: AI fact-check exercise — class asks a question, verifies result.
  • 5 min: Discussion "3 dangerous areas".
💬 Leading the discussion

Main question: "Find 3 concrete areas where AI hallucinations could become dangerous."

Expected pupil answers:

  • Medicine (wrong diagnosis, wrong drug dose)
  • Law (lawyer example — fake rulings cited)
  • History/School knowledge (homework — wrong facts)
  • News (fake-news amplification)
  • Finance (wrong investment tips)
  • Relationships (wrong advice on life questions)

Follow-up question: "What should a law say about this?"

⚠️ Deepening the Lawyer Schwartz case

Background (for teachers): 2023 in New York. Steven Schwartz, a lawyer for 30 years, used ChatGPT for an appeal case. ChatGPT "invented" 6 similar court rulings with case numbers, dates, and citations — all plausible-sounding, none real. Schwartz filed them. Opposing counsel could find none of the cases. Fine: USD 5,000 + public reprimand.

Teacher question for depth: "Could this have happened to any of us? How do you protect yourselves from this?"

❓ Quiz solution with explanation

Answer: Check sources — the claim "Napoleon died in 1822 in Vienna" is false (he died in 1821 on St. Helena).

Why this question is exemplary:

  • Sounds plausible — year and place are "close".
  • Anyone not paying attention adopts it — and passes it on.
  • That's hallucination in small form — and it happens millions of times daily.
🤔 Anticipated pupil questions
  • "Why doesn't AI just say 'don't know'?" — It's trained to always provide an answer. Some newer models learn to say "I don't know", but rarely.
  • "Will AI ever be perfect?" — Unlikely. Humans make mistakes too — AI gets better with more data but isn't error-free.
  • "Can hallucinations be prevented entirely?" — Reduced yes (RAG, low temperature, citation checks). Prevented entirely: no.
  • "What if AI lies on purpose?" — Currently: not intentional. Future: a key AI Safety research question.
🖨 Worksheet: AI fact-check (live in class)

Task: Ask an AI a specific question. Verify the answer.

  1. Pick a topic you have prior knowledge of (e.g. your town, favourite club).
  2. Formulate a very specific question (year, person, place).
  3. Write down the AI's answer verbatim.
  4. Verify with Wikipedia or an official website.
  5. Note: Correct / Hallucination / Partially correct?

Teacher tip: Let the class gather experience — typically several hallucinations appear, making the lesson memorable.

✅ Measuring learning success
  • Understanding: Can pupils list 3 AI limits?
  • Practice: Did they recognise a hallucination?
  • Transfer: Would they automatically verify AI answers in homework?
← One step back
Chapter Seven

You don't have to
be afraid.

Working with AI
responsibly.

Four rules
for your
AI daily life.

AI is a tool — like the phone or the car once was. New and unfamiliar, but essentially harmless.

Four principles for working productively with modern AI systems — beyond hype and panic.

You will be the first generation for whom AI in daily life is completely natural. What you learn now will shape you for a lifetime.

  • Don't share anything personal — no bank data, no passwords.
  • Don't believe everything blindly that it says. For important things: ask back.
  • Just try it out — you can't break anything.
  • Ask family for help when something feels off.
  • No private data into AI — no addresses, passwords, school logins, no first names of classmates.
  • Don't believe blindly — AI makes mistakes. For homework/presentations: always cross-check with real sources.
  • Keep your own thinking — AI is a tool, not a thinker. If you copy everything, you learn nothing.
  • Talk to the teacher if something feels off or you're unsure whether AI use is allowed.
  • Data hygiene: no PII, secrets or credentials into unknown models.
  • Verification: always double-check fact questions. Hallucinations often sound convincing.
  • Iterate: first output is raw material. Ask back, rephrase, combine.
  • Know tool-use: RAG, web search, code execution significantly raise output quality.
  • Compare models: GPT-4, Claude, Gemini, LLaMA, Mistral — each has strengths.
"You don't have to love it.
But now you understand it — and that's half the battle."
"Models change monthly. Principles — verification, privacy, critical thinking — stay."
"AI is a tool — like a calculator. It doesn't change the fact that you have to understand maths."

🍎 For teachers: classroom kit

This page can be used as a double lesson (90 min) or a weekly module (3 × 50 min). Recommended from grade 7 (≈ age 13). Materials are free to use under CC BY 4.0 — please credit "Webagentur Hochmeir e.U. (webhoch.com)".

📦 Full printable Teacher Pack: 6 worksheets (with answer keys), class test + grading rubric, 18 homework tasks in 3 difficulty tiers, parent letter template, curriculum overview. → Open Teacher Pack →

📅 Suggested double lesson (90 min)

  1. 10 min — Opening: Class opens the page, everyone clicks through chapters 1 + 2. Brainstorm on board: "Where do you encounter AI?"
  2. 15 min — Working: Chapter 3 (cat trainer) together, read the learning-objective box aloud.
  3. 20 min — Deepening: Chapter 4 (8 everyday cards) — group work (see schools box).
  4. 15 min — Practice: Chapter 5 — class formulates their own prompts (see template).
  5. 15 min — Reflection: Chapter 6 + case study "Lawyer Schwartz". Discuss: when can AI help, when not?
  6. 15 min — Close: Chapter 7 — write four rules on the board. Homework: find a 9th AI application.

🎯 Assessing learning

  • Quiz boxes in every chapter — instant self-check.
  • Oral exam: "Explain AI to your grandparents in 1 minute."
  • Written assignment: One-page essay: "How will AI change my life in 10 years?" Compare ratings with and without AI help.

📚 Cross-curricular themes

  • English/German: Debate: "AI in school — curse or blessing?"
  • History: Industrial Revolution → Automation → AI as the next stage
  • Ethics/Religion: What distinguishes humans from machines?
  • Maths: Probability theory as the basis of "machine learning"
  • Computer Science: Own mini AI project (e.g. image classification with Google Teachable Machine)

🔗 Further resources for the class

⚠️ Important notes for classroom AI use

  • AI tools (ChatGPT, Claude, Gemini) often have an age limit 13+/16+/18+. Please check the terms!
  • Pupil data must never be entered into AI — no names, addresses, school emails.
  • For homework with AI: agree clear class rules in advance (e.g. "AI as research tool OK, as essay writer no").
Further reading:
  • „Attention Is All You Need“ – Vaswani et al., 2017 (arxiv.org/abs/1706.03762)
  • Anthropic Research – anthropic.com/research
  • Hugging Face Course – huggingface.co/learn
  • 3Blue1Brown: Neural Networks – YouTube series for visual learners
FAQ

Frequently asked questions

Here you'll find the most important AI questions again, compact for reference.

Quick reference to the most important AI topics. Answers are structured in FAQPage schema for search engines and AI assistants.

Artificial Intelligence is a computer program that learns patterns from large amounts of data and on that basis writes text, recognises images, answers questions or makes predictions. AI does not really think — it computes the most likely answer based on patterns in its training data.
An AI learns through Machine Learning: it looks at millions of examples (images, text, audio) and gradually adjusts internal values (parameters) so its predictions get better and better. For large language models this happens in three phases: pre-training on huge text corpora, fine-tuning on curated data, and RLHF (Reinforcement Learning from Human Feedback).
AI meets us daily, often without us noticing: in navigation (routes and traffic prediction), in weather apps, in photo search on your phone, in voice assistants like Alexa or Siri, in translation services, in spam filters, in recommendations on Netflix or Spotify and in search engines.
You type a question or task into a text field — this is called a prompt. The clearer and more specific the prompt, the better the answer. Good prompts give context, define the goal and provide examples if useful. This practice is called Prompt Engineering.
AI has no real feelings or consciousness, it does not know the future and it makes mistakes. In particular it can hallucinate — produce convincing but false information. It has a knowledge cutoff and usually does not know current events. Complex multi-step logic remains a weakness.
AI is a tool. It is not inherently dangerous, but there are risks: misinformation through hallucinations, privacy concerns when entering sensitive data, bias from skewed training data and possible misuse for phishing or deepfakes. Responsible use means: do not enter secrets, check sources, stay critical.
Large Language Models (LLMs) are a kind of Artificial Intelligence trained on huge text corpora to understand and generate human language. Well-known examples are GPT-4, Claude, Gemini and LLaMA. They are based on the Transformer architecture, introduced in 2017 in the paper "Attention Is All You Need".
Artificial Intelligence (AI) is the umbrella term for any form of intelligent behaviour in machines. Machine Learning is a subset of AI where systems learn from data without being explicitly programmed. Deep Learning is in turn a subset of Machine Learning based on deep neural networks.
Powered by webhoch.com