AI literacy under Article 4

What AI is and is not

18 min

Time: 15 minutes

Learning objectives

1. Explain the difference between analytical and generative AI in your own words
2. Describe why generative AI in particular took off at work
3. Distinguish what AI is good at from what it cannot do

Key concepts

AI

A collective name for computer programs that do things which used to require human thinking. AI is not one technology but a family of techniques.

Analytical AI

Programs that look for patterns in data in order to sort, predict or warn. The spam filter, the bank's credit assessment and the streaming service's recommendations are analytical AI. It answers yes or no, or with a probability.

Generative AI

Programs that create new content, text, image or sound, based on patterns in the material they were trained on. The chat assistants belong here. They write, summarise, translate and suggest.

Language model

The engine behind the chat assistants. A large model trained on vast amounts of text to predict which word is likely to come next.

Prompt

What you write to the assistant: the question, the instruction or the text you want help with. How you write determines what you get.

Content

Two kinds of AI that do different things

Analytical AI has been part of daily life for a long time without being called AI. It filters out spam, suggests the next song and flags unusual card payments. It is trained on structured history and works best on bounded questions with a clear answer.

Generative AI is what has happened in the last few years. It creates something new: a draft of the email, a summary of the report, a proposed agenda. It is trained on unimaginable amounts of text and answers in fluent, human-like language. That is why it feels so different, and also why it sometimes sounds confident when it is wrong.

Why generative AI took off at work

Three things came together. It requires no prior knowledge: ordinary language is enough, you do not have to learn a program. The same tool works for text, analysis, ideas and structure, so the threshold to try is low. And the answer arrives in seconds rather than hours.

Behind it lie cheaper computing power, more text to train on and a technical breakthrough in how the models are built in 2017. When the first simple chat interfaces arrived at the end of 2022, the technology became available to everyone at once.

What AI is and is not

The most important thing to understand is that the AI does not understand the text the way a person does. It calculates statistically likely continuations. That explains both the strengths, fluency and breadth, and the weaknesses, invented facts and built-in skews. You own the result. However much the AI has helped, you are responsible for what you send being correct and reasonable.

Where you already meet AI

Think for a moment. The spam filter in your email, the suggestions in the search box, the spell checker guessing the next word, the map app recalculating the route. All of it is AI of the analytical kind. You may have met the generative kind in a chat assistant, in the tool that summarises the meeting, or in a colleague's unusually well-written email.

Exercises

Exercise 1: Analytical or generative? (5 min)

Decide whether the following is analytical (A) or generative (G) AI.

1. The spam filter in your email
2. A chat assistant's answer to a question
3. The playlist the streaming service suggests
4. AI writing a draft email
5. Automatic credit assessment
6. An image created from a text description

Exercise 2: Your own AI inventory (5 min)

Write down three to five places in your daily life where you already meet AI, directly or indirectly. Tools you use, services that suggest things, systems that make decisions about you.

Reflection questions

1. How has your own picture of AI changed in the last two years?
2. Which part of your work do you think AI could support first?
3. Where do you see risks with AI in your particular work?

Summary

• Analytical AI sorts and predicts, generative AI creates new content.
• Generative AI took off because it needs no prior knowledge, works broadly and answers immediately.
• AI calculates likely continuations, it does not understand. Hence both fluency and errors.
• AI is a support for your thinking. You own the result.

Mark module as complete

How a language model works, without the technology

15 min

Time: 12 minutes

Learning objectives

1. Explain to a colleague what a language model does when it answers
2. Anticipate why it is sometimes wrong with full confidence
3. Draw three practical conclusions for how you use it

Content

A very good guessing game

Imagine reading a sentence aloud and stopping halfway: "The meeting has been moved to next …" Most people guess "week". You did not look anything up, you have simply read enough sentences in your life to know what usually comes next.

A language model does the same thing, but with the whole internet as reading habit. It was trained by guessing the next word in billions of sentences and being corrected every time it guessed wrong. After enough corrections it guesses so well that the answers sound as if someone had thought. But they are still guesses about which word is likely to come next, one word at a time.

Why it sounds so confident

The model has no concept of true and false. It has a concept of likely. When you ask for a source it produces something that looks like a source, because that is what sources tend to look like in the text it learned from. Sometimes the source exists, sometimes not, and the model itself notices no difference.

That is also why it rarely says "I don't know". A confident answer is statistically more common than a hesitant one, so that is what it gives. The hesitation has to come from you.

Four things that shape the answer

What you write. The model continues your text. Give context, role and desired format and you get a continuation that fits. Give a single line and you get an average continuation.

What it was trained on. Mostly English, mostly from the internet, up to a certain date. Swedish conditions, recent events and things rarely written down are covered less well. Ask about something recent and it may answer with last year's situation.

Chance. The same question gives different answers at different times, because the model chooses among several likely words. That is a feature, not a bug: it provides variation. But it means an answer is not an answer key.

Memory within the conversation. The model sees what is in the ongoing conversation, up to a limit. Long documents can fall out. The next conversation usually starts from zero unless the tool explicitly saves something.

What the tools add

The chat assistant you use is more than the model. The tool can search the web, read your documents, retrieve from an archive or follow rules your organisation has set. That makes the answers better and fresher, but the foundation is the same: a model guessing likely continuations from what it has been given. Give it the wrong material and it guesses fluently from the wrong material.

Three conclusions for daily work

1. Give context. Say who you are, what the text is for and what the answer should look like. It is the cheapest way to get better answers.
2. Check what has to be right. Names, figures, dates, legal references and sources. Ask for sources by all means, but look them up.
3. Treat the answer as a draft. A good first proposal to work on, not an answer key to send.

Reflection questions

1. When have you noticed an AI answer that sounded confident but was wrong? What gave it away?
2. What context would you need to give to get a useful answer in one of your regular tasks?

Summary

• A language model guesses the next word, one at a time, trained on enormous amounts of text.
• It has a concept of likely, not of true. That is why it sounds confident even when wrong.
• The answer is shaped by what you write, what it was trained on, chance and memory within the conversation.
• Give context, check what has to be right, treat the answer as a draft.

Mark module as complete

The safety rules: three rules that are enough

15 min

Time: 12 minutes

Learning objectives

1. State the three rules from memory and apply them to a task of your own
2. Know where to turn when unsure, before and not after
3. Explain why the authorities have started saying yes, using Kalmar as the example

Content

Three rules are enough

All the law you need in daily work fits on one card.

1. No names, personal identity numbers or sensitive information in tools that the organisation has not approved. Approved tools here: the AI tools your organisation has approved.
2. A person always reviews, before anything goes out or becomes a decision.
3. Unsure? Ask before, not after. Ask the person responsible for AI questions in your organisation, or your line manager.

Why the rules look like this

Rule 1 protects the people we are here for. A tool without an agreement with the organisation is like leaving a file on the bus, even when it feels private. Information about the people we help is always sensitive.

Rule 2 exists because AI can sound confident and still be wrong. Your knowledge of the work cannot be replaced. The decision is always a person's, and you are responsible for what you send and do, not the computer.

Rule 3 means a doubt stays a question instead of becoming an incident. Asking is never wrong. No one is upset about a question.

The authorities have started saying yes

In April 2026 the Swedish Authority for Privacy Protection approved AI support for case notes in social services in Kalmar, after a trial in the authority's sandbox. The conditions were exactly the three rules: no sensitive data in the wrong tool, a person reviews every draft, and clear responsibility. The law is no longer a reason to wait. It says how, not whether.

The hard part is the moment

The rules are simple on purpose. The hard part is not understanding them, it is remembering them in the moment when the AI has already answered and the answer looks good. That is why the card exists. Keep it by the keyboard.

Reflection questions

1. Which of the three rules is easiest to forget in your daily work, and why?
2. Take a task you did yesterday. Which of the rules would you have needed to think about?
3. Do you know today which tools are approved in your organisation? If not, who do you ask?

Summary

• No names, personal identity numbers or sensitive information in tools that are not approved.
• A person always reviews before anything goes out or becomes a decision.
• If unsure: ask before, not after.
• The authorities have started saying yes. The law says how, not whether.

Mark module as complete

Hallucinations and source criticism

18 min

Time: 15 minutes

Learning objectives

1. Recognise the most common warning signs of invented information
2. Have a routine for checking what has to be right
3. Rewrite a text when a claim cannot be verified

Key concepts

Hallucination

When the AI produces something that sounds credible but is invented: a figure, a report, a person, a quote. It is not a bug that can be switched off. It is the same mechanism that makes it write fluently.

Verification

Checking a claim against a source that exists independently of the AI. Not asking the AI whether it is sure.

Content

Why AI makes things up

A language model guesses likely continuations. When the text calls for a source, a year or a percentage, it produces something that looks like a source, a year or a percentage. Often it is correct. Sometimes not, and the difference does not show on the surface. This applies to the best tools too, and it applies especially to anything specific: exact figures, names of people and reports, sections of law, quotes.

It happens to professionals as well. In 2026 a large audit firm had to withdraw a report on AI that turned out to contain invented sources, produced with AI. The author had not checked.

The warning signs

Learn to recognise them, and you will automatically check in the right places.

• Exact figures without a source. "Studies show that 73 percent of …"
• Named reports, authors and researchers. Always check that they exist and that they said it.
• Quotes. Especially well-phrased quotes that fit a little too well.
• Detailed descriptions of something you did not ask about. The model likes to fill in.
• Authoritative claims without nuance. Reality more often has an "it depends".
• Recent events. The model may answer with an older situation without saying so.

The routine that is enough

You do not need to check everything. You need to check what has to be right.

1. Underline everything in the text that is a checkable claim: figures, names, dates, sources, legal references.
2. Check each one against a source that exists independently of the AI. The authority's website, the report itself, the register.
3. If you cannot find the source: rewrite without the claim. "Many municipalities have started using …" instead of "67 percent of municipalities have …".
4. Never ask the AI whether it is sure. It says yes.

When it matters most

The more serious the consequence of an error, the stricter the check. An internal list of ideas can tolerate an error. A decision document, a letter to an individual, a press release or a text quoting the law cannot. There the rule is: every checkable claim checked, or removed.

Exercises

Exercise 1: Review the paragraph (10 min)

The AI has written the following for a briefing:

"According to the National Audit Office report 'Digital maturity in the municipal sector 2024', 67 percent of Swedish municipalities have introduced AI-based chatbots for citizen services. Researcher Maria Lindgren at the University of Gothenburg argues that 'municipalities that do not digitalise will fall behind within three years'."

1. Which warning signs do you see?
2. What would you need to check before using the paragraph?
3. How do you rewrite it if you cannot verify the claims?

Reflection questions

1. Which text you write at work would have the most serious consequences if a figure were invented?
2. Have you ever passed on an AI claim without checking? What happened?

Summary

• AI makes things up because it guesses likely continuations. It cannot be switched off.
• Warning signs: exact figures, named sources, quotes, details you did not ask for, overconfidence.
• Check what has to be right against an independent source. If you cannot find the source, rewrite without the claim.
• Never ask the AI whether it is sure.

Mark module as complete

The law in daily work: the AI Act, Article 4 and the Swedish guidelines

18 min

Time: 15 minutes

The legal position in this section was checked on 13 September 2026. Laws and guidelines change; the sources are in the module's metadata and are reviewed every quarter.

Learning objectives

1. Describe in broad terms what the AI Act requires of employers who use AI
2. Explain what the AI literacy requirement in Article 4 means for you
3. Know which rules apply when AI meets personal data, confidentiality and employees

Key concepts

The AI Act

The EU's common law on AI, in force since August 2024 and introduced in stages. It sorts AI use by risk: prohibited, high risk, limited risk and minimal risk, with different requirements at each level.

Article 4, AI literacy

The employer's responsibility for ensuring that people who use AI at work know enough, adapted to role and context. Since 27 July 2026 the requirement is that the employer takes measures to promote AI literacy.

The DIGG and IMY guidelines

Sweden's national guidance for generative AI in public administration: 18 guidelines in 7 areas, launched on 21 January 2025.

Content

The risk classes, in brief

The AI Act does not treat all AI alike. What is prohibited is things like social scoring of people and manipulation that exploits vulnerability. What is high risk is AI that seriously affects people's lives: recruitment and assessment of employees, access to welfare and education, credit assessment, law enforcement. There, risk assessment, documentation, human oversight and supervision are required, with full effect from 2 December 2027 for stand-alone systems and 2 August 2028 for AI built into products. Most of what you use in daily work, chat assistants and writing aids, is limited or minimal risk with transparency requirements but not much more.

The point for you: ordinary daily use is permitted and easy. It is when AI is to assess people that the law sharpens its tone, and you do not decide that on your own.

Article 4: the knowledge is the employer's responsibility

Since 2 February 2025, organisations that use AI have been responsible for making sure that those who use it know enough. In the summer of 2026 the EU rewrote the article through the so-called Digital Omnibus, in force from 27 July 2026. The organisation must now take measures to promote AI literacy among its staff, with no fixed level of knowledge per person and no certificate. Since 2 August 2026, shortcomings can lead to sanctions, and supervision is handled by national authorities. In practice the organisation must have thought through who uses AI for what, what risks that involves, and provide measures that fit. The European Commission expects an internal register of what has been done.

This training is such a measure. What you are reading now is documented in your employer's register. Literacy in the sense of the law is not technology: it is understanding opportunities and risks well enough to use AI wisely in your role. Exactly what the modules before this one are about.

The Swedish guidelines: seven areas

In January 2025 DIGG and the Swedish Authority for Privacy Protection published 18 guidelines for generative AI in public administration, arranged in seven areas: management and responsibility, data protection, labour law, procurement, information security, copyright and ethics. They are written for everyone in the administration, from employees to management, and are available at digg.se/ai. Three principles recur: a person is in control, openness about when AI has been used, and data protection from the start.

If you work outside the public sector, the guidelines are still the best Swedish summary of what responsible use means in practice.

Personal data, confidentiality and employees

The General Data Protection Regulation applies in full when AI processes personal data. That means information about people only goes into tools the organisation has an agreement with, that decisions based solely on automated processing require human review, and that anyone using AI in decisions must be able to explain the basis for the decision. In spring 2026 the Swedish Authority for Privacy Protection showed in Kalmar that AI support for case notes can be done right: there is a legal basis when a person reviews, the data is protected and the staff are trained. The law says how, not whether.

Public access and secrecy applies in the public sector. Classified information may not leave the authority for a tool without an agreement, and what you get out of AI can become a public document.

The Co-determination Act applies when AI is introduced in a way that affects work. The employer must inform and negotiate with the union before deciding, and AI that measures or assesses employees is also high risk under the AI Act.

What it means for you

Reflection questions

1. Which row in the table is hardest to live up to in your daily work, and why?
2. Is there anything you do with AI today that, after this section, you want to check with someone?

Summary

• The AI Act sorts AI by risk. Daily use is permitted and easy, assessing people is high risk.
• Article 4: the employer must take measures to promote your AI literacy. This training is one such measure.
• The 18 DIGG and IMY guidelines in seven areas are Sweden's guidance, digg.se/ai.
• Personal data only in approved tools, a person reviews and decides, ask before when unsure.

Mark module as complete

Bias and fairness

15 min

Time: 12 minutes

Learning objectives

1. Explain how skew arises in AI and why it is hard to see
2. Recognise the most common types of bias in everyday use
3. Ask four questions that protect against one-sided answers

Key concepts

Bias

Systematic skew in what the AI produces, arising from skew in what it was trained on or in how it was built. Can lead to unfair or misleading results, especially where people are concerned.

Content

How the skew arises

The model learns from text written by people. That text reflects how the world has looked, with prejudices, inequalities and gaps. The model learns the patterns without knowing which ones are unfair, and reproduces them in its answers. The chain is simple: skewed training data gives a model that has learned the skew, which gives skewed answers. The hard part is that the answers sound just as fluent and confident as all the others.

Five types to recognise

The last one is the one you most often meet yourself. Ask "why is proposal A best?" and you get reasons for A. Ask "what objections are there to A?" and you get those too. The model continues in your direction.

When it matters most

Bias is unpleasant in a text summary and dangerous when AI affects decisions about people. That is why the AI Act classifies exactly such use as high risk: recruitment, assessment of employees, access to welfare and education, credit assessment. There, risk assessment, documentation and human review of every outcome are required, and that is not something an individual employee introduces on their own.

An example: AI that helps sort applications can systematically mark down candidates on the basis of name, educational background or other things that correlate with gender or origin. The right use is to anonymise the material, let the AI only extract information and never rank, and let people assess. The wrong use is to let the model choose.

Four questions that protect

Ask them of every answer that concerns people or will form the basis of a decision.

1. Whose perspective is missing in this answer?
2. Does this apply to us, in our operations and our conditions?
3. Did I ask a leading question? Try the opposite question and see what you get.
4. Is there another interpretation the model did not raise?

Reflection questions

1. Have you noticed AI agreeing with you a little too easily? How would you have asked instead?
2. Where in your operations would a skew in an AI answer actually affect someone?

Summary

• Skew comes from what the AI was trained on, and does not show on the surface.
• Confirmation bias is the most common in daily work: the model continues in your direction.
• AI that assesses people is high risk. Anonymise, let the AI extract but never rank, let people assess.
• Four questions: whose perspective is missing, does this apply to us, did I ask a leading question, is there another interpretation.

Mark module as complete

Your result code

Follow-up

Add up