Machine learning is the heart of contemporary AI. The basic idea is surprisingly simple: instead of explicitly programming a computer to perform a task, you provide it with enormous amounts of data and let it find patterns on its own. A machine learning model is essentially a very complex mathematical function that, given an input, produces the most "probable" output according to patterns in the training data.
Let's take a concrete example: a system that must distinguish photos of cats from photos of dogs. We don't write rules like "if it has pointed ears it's a cat". We provide the model with millions of labelled photos ("cat", "dog") and the model, through an iterative process called backpropagation, adjusts its internal parameters until it minimises errors. In the end, the model has "learned" to distinguish cats and dogs — not because it understands what they are, but because it has internalised statistical patterns in pixel configurations.
This process works incredibly well for many tasks: image recognition, translation, speech transcription. But it hides a fundamental problem: data is never neutral. Every dataset is the product of human choices — who decides what to collect, how to label it, what to exclude. If a face dataset predominantly contains white faces, the model will be poor at recognising black faces. Not out of malice, but out of mathematics: the model optimises its performance on the data it has seen most.
Researcher Timnit Gebru, before her controversial dismissal from Google, documented how the most widely used machine learning datasets are deeply imbalanced: they over-represent the Global North, the English language, Western cultures, men and white people. When we train models on this data, we are not creating "objective" tools — we are encoding existing inequalities into systems that appear neutral.
Then there is the invisible cost of training. Training a single large language model can consume as much energy as five cars over their entire lifetime, according to a study by the University of Massachusetts Amherst. The water needed to cool data centres, the rare earths for chips, the labour of thousands of data labellers paid a few cents per task — all this is hidden behind a chatbot's elegant interface.
The principle "garbage in, garbage out" is perhaps the most important law of machine learning. But we should update it: "bias in, bias out, amplified". Biased data does not just produce biased results: it produces biased results with an aura of scientific objectivity that makes them even more dangerous.