What happens inside an AI while you talk to it?

Your sentence is broken into small blocks and translated into numbers. A net of billions of values then estimates which block most likely comes next, over and over. Your prompt is the stage direction that settles which play is being performed. Understanding in the human sense does not happen, and still something comes out that feels like a conversation. I used the first image in 2023 and I still use it. Language models learn from examples. Nobody writes their rules. So nobody knows in advance exactly what they can do, not even the people who release them. The evidence for that is called emergent abilities. One model was trained on English texts, kept getting more data, and past a certain amount it could suddenly answer in Persian. Nobody built that in. It arrived with the volume, the way simple arithmetic does.

A USB stick from an alien

I used the first image in 2023 and I still use it. Language models learn from examples. Nobody writes their rules. So nobody knows in advance exactly what they can do, not even the people who release them.

It feels as if an alien had come to Earth and handed us a USB stick with a new technology on it.

The evidence for that is called emergent abilities. One model was trained on English texts, kept getting more data, and past a certain amount it could suddenly answer in Persian. Nobody built that in. It arrived with the volume, the way simple arithmetic does.

Building blocks and digital brain cells

A computer works with numbers, not with words. So the first step is always a translation.

You can picture it as if the sentence were one long Duplo bar.

The bar gets broken into single bricks, and every brick gets a number. Those bricks are called tokens. The same happens with images, music and video, where the bricks are pixels, notes or single frames. Without this step none of the applications we use today would exist.

What comes next I call the digital brain cells. The technical word is parameters, and current models have billions to trillions of them. Their only job is an estimate: which brick most likely follows the ones before it. What is stored are probabilities. The model has no consciousness and no intention. That it feels otherwise comes from how good the estimate has become.

The actor and the stage direction

Guessing turns into a conversation because someone builds a play around it. You set a role, add a few example lines, and the model carries the play on.

Picture the language model as an extremely versatile and well read actor.

This actor has read every script in the world and can play any role. Without a clear instruction they improvise something brilliant that does not fit your scene. On top of that sits a limit many miss: the context window holds only so much at once. Too much unrelated context makes the answer worse.

There is a name for why two people asking the same question get different texts. The temperature setting controls whether the model always takes the most likely next brick or sometimes the second and third. Set to zero, the same question returns the same text twice.

Where it goes wrong and who answers for it

A model keeps guessing even when it does not know the answer, and it guesses plausibly. That is how market figures appear in a report that nobody ever collected, and sources in a paper that never existed. At first glance both look serious. In people this is called confabulation, an invented reason that feels genuine. In a machine it is the other side of exactly the ability that makes it useful.

The second risk is bias, and it arises on three routes. In the data, when the training texts mostly show men as engineers and the model carries that pattern forward. In the design of the model itself. And in the feedback loop, when clicks and ratings pull a system further in one direction month after month.

What follows is a stance. An AI has no conscience and no morals, and it does not replace judgement. A person stays the final instance and carries responsibility for the decision. In daily work that is an approval: a draft is ready, a person looks at it, and only then does it leave the building.

The exam results I listed in 2023 are long out of date, and models invent less today. The four images still hold, because they hang on no generation of models. And the stance does not hang on an error rate anyway.

The difference between the engine, the car around it and the fleet is in What is the difference between an AI model, a tool and an agent?. That a model would rather be friendly than honest is related and sits in Do you have to teach an AI to tell uncomfortable truths?. Why the real work still is not in the wording is in Does good AI work come from better prompting or from better systems?. And what approval looks like in operation is in How do you guard against hidden instructions without crippling the AI?

  • Tell the AI which role it plays. Without a stage direction it improvises something usable but rarely the thing you needed.
  • Give it the context, and only that. A lot of irrelevant material in the window measurably worsens the answer.
  • Check figures, quotes and source references. The more serious a claim looks, the closer the look it needs.
  • Do not expect the same answer twice. The same question gives two texts, because the most likely brick is not always taken.
  • Let a person decide when people are affected. Applications, customer ratings and prices are no place for an automatic tick.

This note keeps growing

2026-09-03: Planted from an introductory video on text generation, 2023, and the foundation modules of a certification course on the EU AI Act, 2025.

ZukunftBilden GmbH · Salzburg · +43 681 81655313 · office@zukunftbilden.eu