Data · AI & Jobs
Which jobs will AI touch, and where?
One map, three economies. Each rectangle is an occupation, sized by how many people do it and coloured by how exposed the work is to generative AI. Switch between Denmark, the European Union, and the United States.
Tap any occupation for detail. Sorted from most to least exposed.
Sources: Eurostat LFS employment by occupation, 2025; ILO Working Paper 140, 2025; karpathy/jobs (US BLS + AI-scored). Occupation groups are ISCO 2-digit for Denmark and the EU; the US view uses detailed US occupations. Armed forces are excluded.
These hands-on AI demos work best on a bigger screen. Open this page on a laptop or desktop to try them.
AI literacy · How it works
So what is AI actually doing?
You just saw where AI is landing on work. Underneath, it is doing two simple things over and over: it reads your text in chunks, then it guesses what comes next. You can try both below.
It reads in chunks, not words
Type anything below. This shows how a model like ChatGPT breaks your text into pieces called tokens. It does not see letters or whole words, it sees these chunks, and it works through them one at a time.
What this means. This is exactly what a model sees: tokens, not letters. Three things follow. It is billed per token, so more text costs more. It can only hold so many tokens at once, a limit called the context window. And because it reads chunks and not letters, it genuinely cannot count the letters in a word like strawberry, which it sees as just three pieces: st, raw, berry. Once it has your text as tokens, the only question left is what token comes next.
Uses o200k, the tokenizer behind GPT-4o and today’s OpenAI models (GPT-3.5 and the original GPT-4 used an older one). Other models split a little differently; for an exact count you can check any text in OpenAI’s own tokenizer. The idea is explained from scratch in Andrej Karpathy’s open minbpe; the tokenizer here is the gpt-tokenizer port of OpenAI’s tiktoken. A dot marks a space.
Then it guesses the next word
Once it has read your words, this is all a model does. It does not look anything up. It looks at the words so far, estimates how likely each possible next word is, picks one, adds it, and does it again. That loop is the whole trick, from this tiny model to ChatGPT. Start a sentence and watch the bars move.
What this means. A model predicts, it does not remember, which is why it can sound completely fluent and still be wrong. Always pick the most likely word and it reads flat and repetitive; let it take chances and it turns creative but wanders. Big models expose that choice as a setting called temperature: low keeps it safe, high lets it get bolder. A hallucination is just prediction going wrong: an unlikely guess that looked likely enough, delivered with the same confidence as a true one. That is why a model always sounds sure, and why it is worth checking what it tells you.
This is a real model, only a small one: an n-gram model built from public-domain books (Alice in Wonderland, Sherlock Holmes, Pride and Prejudice and others). It remembers just the last word or two, so it wanders; models like GPT use the same next-word idea with far more context and a neural network. The method is taught from scratch in Andrej Karpathy’s open makemore. It runs in your browser and works in lowercase.