Patterns from Data Instead of Programmed Rules
Traditional software runs on fixed rules that a person wrote down. For example: "If the invoice amount exceeds 10,000 euros, route the invoice to the department head for approval." The program does exactly that — nothing more, nothing less. Same input, same output, fully traceable at any time.
Machine learning turns this principle on its head: instead of specifying rules, you show the system a large number of examples, and it derives the patterns itself. A spam filter, for instance, does not consist of thousands of hand-written if-then rules. It was trained on tens of thousands of emails that people had previously labeled as "spam" or "not spam" — and learned from them the statistical features that typically characterize spam.
This has an important consequence: with a learned system, nobody can point to a line of code and say exactly why a particular decision came out the way it did. The behavior depends on the training data — if that data is incomplete, outdated, or skewed, so are the results. This property is the root of many opportunities, but also of many risks, which we cover in the lessons that follow.
What a Language Model Does: The Most Likely Continuation
Tools like ChatGPT are built on so-called language models. Their core function can be described in one sentence: the model calculates which word is most likely to follow the text so far — and repeats this word by word until a complete answer has been produced. You know the principle in a simple form from your smartphone's word-suggestion feature; a language model is a dramatically more capable version of it.
What matters most is what a language model is not: it is not a reference work and not a database. It does not retrieve a stored answer; with every request, it generates new text that statistically fits the question. The answers sound so assured because the model has learned from millions of well-written texts what convincing answers sound like — not because it "knows" the answer.
Whether such a system truly "understands" anything is a matter of debate among experts. What counts for your work is the practical part: the model has no intent and no built-in notion of true or false. A helpful mental image is an extremely well-read phrasing machine — not a digital encyclopedia.
Training vs. Use: Two Separate Phases
A language model is created in the training phase: the provider has it process enormous amounts of text, gradually adjusting billions of internal parameters so that its predictions fit better and better. This happens at enormous computational cost in the provider's data centers — long before you open the tool for the first time.
When you then use the model, it is finished and frozen. It does not learn anything in that moment: your conversation does not change the model, and a new session starts with no memory of the previous one — unless the provider has deliberately added extra features such as a session memory.
Two practical points follow from this separation. First: the model's knowledge ends with its training data — it does not know about more recent events unless the tool has additional features such as built-in web search. Second: depending on the provider and the contract terms, your inputs may be stored and potentially used for future training. This is why many companies set rules about which data may be entered into AI tools — more on this in Lesson 4.
Hallucinations: Not a Defect, but How the System Works
A "hallucination" is when a language model fluently and confidently states something false: an invented source, a wrong number, a court ruling that never existed. Anyone experiencing this for the first time takes it for a bug the manufacturer should simply fix.
In fact, it is a direct consequence of how the system works. The model always does the same thing: generate the most plausible continuation. There is no built-in truth check — correct and incorrect answers are produced by exactly the same mechanism. Where the training data is thin — say, for exact legal provisions, specific figures, or niche knowledge — the most plausible answer is not always the right one. Better models and connected tools reduce the problem; based on the current state of the art, it will not disappear entirely.
For your daily work, this means: how convincing a statement sounds says nothing about whether it is correct. Check everything you intend to reuse — figures, names, legal references, quotations — against a reliable source. This risk awareness is exactly what the European Commission names as a core element of AI literacy under Article 4.
What AI Does Well — and What It Doesn't
The way these systems work translates into a clear profile of strengths. Language models shine wherever the task involves language, structure, and patterns, and a human reviews the result afterwards:
By contrast, language models are weak wherever verified precision matters and nobody double-checks the output:
The rule of thumb: AI is an assistant, not an oracle. It delivers drafts and suggestions in seconds — the judgment, the corrections, and the responsibility stay with you. A sales team that uses AI drafts for proposal letters and reviews every draft before sending is using the tool exactly right.
- Strong: first drafts (emails, minutes, announcements), summarizing and structuring long documents, translating and adjusting tone, brainstorming and phrasing variants, spotting patterns in large volumes of text or data
- Weak: exact facts without a cited source (figures, dates, legal provisions, quotations), current events beyond the training data, binding legal, tax, or medical advice, precise calculations and consistency across very long documents
- Never delegable: responsibility — it always remains with you and your company
Your Mental Model for Everyday Work
To sum up: AI is neither magic nor an all-knowing counterpart, but a statistical tool that extends patterns learned from data. It writes impressively, works fast — and offers no guarantee of truth. Once you have internalized this picture, you can use AI productively and safely at the same time: boldly for drafts and ideas, rigorously when checking facts.
The next lessons build on this foundation: in Lesson 3 you will see where this tool creates real value in small and medium-sized businesses, and in Lesson 4 which risks it carries and which rules apply for using it safely.