From using AI to thinking with it. Progressive course based on the Anthropic framework: fundamentals, the 4Ds method and ethical and safe use.
Objective: move from seeing AI as an "intelligent mind" to understanding it as an expert system in probabilities. Inductive methodology: Experience → Concept → Term.
🧠 Opening activity — "The Autocomplete-Brain": The teacher says an incomplete sentence and students finish it: "Tell me who you walk with and I'll tell you who…" Objective: demonstrate that humans also predict by probability. Just like AI.
Doesn't search for answers. Calculates which word has the highest probability of coming next based on millions of texts.
AI breaks text into minimal units (tokens) and chooses the piece that best fits each gap.
It has no consciousness, feelings or common sense. It doesn't seek the truth — it seeks coherence.
When AI invents false data with total confidence because it "sounds coherent". Prioritises grammatical coherence over truthfulness.
Large Language Model. Mathematical model trained on massive text. Not a mind, it's a probability formula.
Minimal unit of text that AI processes. Can be a syllable, a short word or a punctuation mark.
AI's "workbench". Everything it can remember from the current conversation. Has a maximum size.
Low = predictable and literal responses. High = more creative and risky responses.
The instruction you write. The clearer and more specific, the better the response.
False data presented with confidence. AI prioritises coherence over truthfulness.
AI fluency is not knowing how to use tools. It's the ability to work with AI systems in an effective, efficient, ethical and safe way. The goal is to shift from "using a tool" to "thinking with AI".
The "how". Using tools and prompting techniques to get real results.
The "what and why". Understanding how systems work to anticipate their failures.
Critical judgment to discern the quality and accuracy of what AI produces.
The ethical compass for responsible use. You are the moral author of everything AI produces under your direction.
Before delegating, it's vital to know what an LLM can and cannot do. Delegating without knowing the limits is the main source of errors.
Language mastery: Summaries, translation, explaining complex concepts in plain language.
Conversational flexibility: Adapts tone, format and level of detail to context.
Tool integration: Can connect to APIs, databases and external services via MCP.
Hallucinations: The model can invent facts that sound plausible. Never assume a technical fact is correct without verifying.
Knowledge cutoff: They don't know events that occurred after their training date (e.g. November 2024). For current news, use web search.
Context Window: They have a limit of information they can process before starting to "forget" earlier content.
Complex reasoning: They can fail at multi-step logical or mathematical tasks. Ask it to reason step by step.
The 4Ds are the core of the course. A framework for making intelligent decisions at each step of your work with AI.
🗺️ Definition: Strategically deciding which tasks the human does and which the AI does. It's not about giving everything to the machine, but knowing who does what better.
✍️ Definition: The ability to communicate effectively with AI. AI doesn't read minds — it needs a clear map.
What do you want to receive? Define format, style and target audience before starting.
How should AI work? Give it steps or rules to follow. Ask it to "think before responding".
How should it behave? Define tone, attitude and expected level of detail.
Clear context: Explain who you are, what the result is for and who the audience is.
Few-shot (examples): Provide 1–3 examples of the expected result to calibrate the format.
Chain of thought: Ask it to reason step by step: "Before answering, think out loud."
Output constraints: Define limits: maximum length, format (table, list, paragraph), language.
🔎 Definition: The critical filter. The ability to review what AI delivers and decide if it's useful, real and appropriate. Protects your work from reasoning errors, hallucinations and inappropriate behaviours.
| Type | Key question | What to watch for |
|---|---|---|
| Of the Product | Is the result useful and truthful? | Accuracy, coherent structure, audience appropriateness, objective compliance. |
| Of the Process | Is the logic solid or was it an accidental hit? | Logical errors, unnecessary steps (noise), circular reasoning (sycophancy loop). |
| Of the Performance | Was the communication fluid and efficient? | Overly long responses, apology loop, too many unnecessary questions. |
Circular Reasoning: AI repeats the same idea without progressing or agrees with arguments you had already discarded. Solution: restart the chat with a new instruction.
Noise: AI complicates the simple by adding unnecessary phases or analysis. Example: asking for an email and receiving 3 paragraphs on the history of telecommunications.
Apology loop: Apologises repeatedly but commits the same error again. Signal that the description needs to be reformulated.
🛡️ Definition: Ethical and safe use. Ensuring that the human factor always prevails over automation. "Diligence creates a containment wall between the power of technology and the risks of the real world."
Accountability: AI has no moral agency. If you validate a result, you assume its successes and errors as your own.
Security and Privacy: Sensitive information (company secrets, customer data) should never be uploaded to public models.
Active verification: Distrust technical data. Validate facts and citations in external sources. AI can be very convincing when it's wrong.
Ethics and biases: Don't delegate the final decision. Apply moral filter: does it respect values? does it avoid biases? what impact does it have on customers and society?
Converting instructions into reusable tools (Gems in Google Gemini) and executing a real project in Description–Discernment loops.
A customised chatbot with specific instructions and defined personality. Reusable for repetitive tasks.
1. Title + instructions (detailed prompt). 2. Select model. 3. Test and iterate until precision is refined.
Choose a real challenge → Delegation plan → Work loop → Final responsibility for the result.
Using AI is not bad. What's bad is letting it think for you. Rules of the game for using it with integrity.
✍️ Principle: Using AI to understand concepts, structure ideas or correct grammar is honest. Copying and pasting generated text saying you wrote it yourself is not. If it didn't come from your head, it doesn't carry your name.
🟢 Green — Free: Use AI as a "study companion". Ask it for examples, summaries of texts you've already read or ideas to get started.
🟡 Yellow — With care: Ask it to draft parts of an assignment. If you do, review every word, change it to your style and verify the data is real.
🔴 Red — Forbidden: Ask it to solve an exam, do a complete task from start to finish or invent bibliographic citations.
When you use AI on an important assignment, be brave and say so. Add a note at the end: "For this work I had the support of [AI Name] to organise ideas and review spelling." That shows you're an intelligent and honest user.