sala.red From Bolivia, sala.red investigates, analyses and comments on how technology shapes society in the Global South.

Laboratory AI Literacy

How an AI learns?

An AI understands nothing: it looks at thousands of examples, finds what repeats, and what is not in the data simply does not exist for it.

By — Investigadora en tecnología.

You see a woman in a pollera crossing the street and within a second you know a great deal. Which region she might be from. Whether she is heading to a fiesta or to the market. Roughly how old she is. How the hat sits on her head and what that says.

Nobody taught you that in a classroom. You learned it by looking. Thousands of people, thousands of times, across your whole life.

An artificial intelligence also learns from examples. That is where the resemblance ends.

How a machine learns

To make it recognise a pollera you have to show it photos. Not one. Thousands. Polleras in different colours, in different cities, by day and by night, close up and far away.

The machine does not look at photos the way you do. It turns them into numbers and hunts for , meaning things that repeat across all of them.

pattern: something that shows up over and over. On a pollera: the pleats, the fall of the fabric, the shape of the hat above it. The machine has no idea what a pollera is. It knows those shapes appear together an enormous number of times.

Train it yourself

Choose how many photos you show it and watch what happens.

Trainer

You are teaching it to recognise a woman in a pollera.

Pick an amount.

Nobody explained what a pollera is. Nobody told it where it comes from, why it is worn, or what it means. It counted shapes until a rule fell out.

1. THOUSANDS OF EXAMPLES different colours, places, light 2. FINDS WHAT REPEATS pleats, fall, hat 3. A RULE IF IT LOOKS LIKE THIS, MAYBE never certainty. probability.

The whole process. The word understand appears at no point in it.

What you see, what it sees

Same woman, same photo. Tap to compare.

One image, two readings

A woman in a pollera on the street.

Choose who is looking.

You read a story. The machine measures a shape.

What if it never saw enough of Bolivia?

This is the point of the whole thing.

Artificial intelligence is being built mostly in other countries, in other languages, out of other cultures and other data. The models you use every day learned from what was on the internet. And the internet holds millions of photos of New York and almost none of El Alto. Entire libraries in English and next to nothing in Quechua or Aymara.

Try different sets of photos and see the result.

What was it trained on?

Same model, same training time. Only the material changes.

Pick a set of photos.

That is called . It is not a programming error. It is an exact reflection of what was in the photos.

bias: when a model works well for some cases and badly for others, because the examples were not spread evenly. If nearly all the photos came from one part of the world, it will be good there and weak everywhere else.

What happens when we are missing

It describes us wrongA pollera can end up tagged as a costume, and a dance as a show. The machine reaches for the closest category it knows, and the ones it knows came from somewhere else.
It makes things upWhen it does not know, it will not say so. It fills the gap with the nearest thing and states it with total confidence. Ask about a fiesta in your town and it produces dates, names and details that sound perfect and do not exist. That is called .
hallucination: when the model produces something false in exactly the same confident tone it uses when it is right. Nothing in the answer tells you which is which. Every fact it hands you needs checking somewhere else.
It caves inPush back and it will often change its answer to please you, even when the first one was right. It gives way because it was trained to be liked.
It cannot translate oursEnglish to Spanish comes out well because it saw millions of examples. For Quechua and Aymara it has almost no material, and it shows.

Whatever is missing from the data does not exist for the machine.

And this can change

You do not need to be an engineer. Data means photos, recordings, video and text, and anyone with a phone produces that.

Record. Photograph. Note what each thing is, who is dancing, where and when. Capture people speaking Quechua and Aymara in ordinary conversation, not only reciting. Build collections that exist nowhere in the world today.

And before starting there is one question to answer: who owns that data.

A community lending its music, its weaving or its language to train a model is handing over something valuable. It gets to decide who uses it, what for, and if money comes of it, to receive a share. That gets written down before anyone hits record.

Bolivia does not have to be only a user of technology built somewhere else. It can put data, knowledge, culture, research and creativity into what is being built right now.

Artificial intelligence learns from the data we give it. If Bolivia is not in that data, AI will not learn who we are either. Creating our own data is also a way of preserving our culture.

Next in this series
  1. How exactly does an AI learn?
  2. Who controls artificial intelligence?
  3. What happens if Bolivia does not create its own data?

Jasmin Jiménez Gonzales.
Psychology student, 26, Cochabamba. Quechua.
She works at the crossing of psychology, technology, culture and artificial intelligence.
Production and interactive development: sala.red.