Before a machine can recognize a face, a road sign or a sign of illness on a scan, it has to learn to see. It does this by studying millions of photographs, each one labeled by a person. The machine learns that this is a dog, that is a shoe, and so on.

This sounds like a purely technical job. But it is also a moral one. What a machine is shown, and who describes it, shapes how the machine understands the world.

In 2021, MIT Technology Review reported on a surprising idea from Japan. Researchers asked whether a machine really needs photographs of our world to begin learning. Their answer was to teach it with fractals instead.

What is a fractal?

A fractal is a pattern that repeats itself at every size. Look closely at a fern, a coastline or a cloud, and you will see the same shapes again and again, smaller and smaller. Nature is full of them.

The researchers used a computer to create an endless supply of fractals. Some looked like leaves, others like snowflakes or snail shells. The computer sorted similar patterns into groups and labeled them automatically. No human had to do it.

They called this collection FractalDB. They used it to give a machine its first lessons in seeing. Then they finished its training with real photographs.

The result was impressive. The machine performed almost as well as machines trained on the biggest photo collections, such as ImageNet, which holds more than 14 million pictures. The team also tried other kinds of abstract images, but fractals worked best. The lead researcher, Hirokatsu Kataoka, believes this is because fractal geometry is part of the hidden knowledge of the natural world.

Why does this matter for ethics?

Today, most machines begin with big collections of photos gathered from the internet. These collections have real problems.

First, the labeling is usually done by workers who are paid very little, and their work is rarely noticed.

Second, the labels can be unfair. Some contain racist or sexist descriptions, and the machine can absorb these without anyone realizing it.

Third, some photos show real people who never agreed to be included.

And because many other systems are built on top of these first lessons, the problems spread. A small hospital or a young company can use a ready-made machine instead of building one from nothing. That is a good thing. But it also means any hidden unfairness is passed along to everyone who uses it.

Fractals avoid much of this. A fractal has no face. It has no private life. Nobody needs to be paid a few cents to label it, and nobody’s picture is taken without permission. It carries no stereotypes. For a machine's first lessons, that is a very clean start.

Reasons to be careful

It would be wrong to think the problem is solved. The original report includes a note of caution.

Anh Nguyen, a researcher at Auburn University who was not part of the study, was not yet convinced that fractals can match the big photo collections. He has studied how strange patterns can confuse machines and make them see things that are not there. He sees a possible link between that problem and this new method, and he wants to look at it more closely.

There is a second reason for caution. Unfairness does not disappear just because the first lessons are clean. After the fractal lessons, the machine is still trained on real photographs, and that is where human bias can return. Also, people still decide which fractals to make and how to judge success. These are human choices, even if they feel technical.

The last warning is about words. Terms like “computer-made” and “automatic” can sound neutral and fair. But a machine is only as fair as the way it is used. A system that decides who gets a job or a loan must answer for its decisions, no matter how it learned to see.

A lesson for all of us

The big idea here is simple: Ethics must start early. When we talk about unfair machines, we usually look at the final result, such as a wrong decision or a misidentified face. But many problems begin much earlier, in choices about what the machine first learns from. Few people ever examine those choices.

This matters in the Arab world too. Many countries in the region are investing heavily in artificial intelligence. Having powerful computers is only part of the answer. Leaders must also ask what information their machines learn from, whose lives it includes, and who did the work of preparing it. Experts have long pointed out that many large photo collections show mostly Western settings. A region building its own technology should think about that.

Fractal training will not answer every question. But it shows that there is more than one way to teach a machine, and that we can choose a better way.

Fractals are a good symbol for this. They show how something complex can grow from small, simple steps repeated carefully. Making machines fair may work the same way. There will be no single grand solution. Instead, we need to take care with each small decision, at every stage, from the first image a machine sees to the last choice it helps make.

  • Rafael Hernandez de Santiago, viscount of Espes, is a Spanish national residing in Saudi Arabia.