The people who build AI, and whom we never see

An AI learns from annotated examples. Someone did that annotation. It is massive human work, and it is the part of the pipeline we talk about least.

What these people actually do

• Annotate: say what is in an image, whether an answer is good, which of two answers is better. It is this last task that makes a model pleasant to use.
• Moderate: decide what an AI must refuse. To teach it to refuse horror, you have to show it horror, and that means people have to look at it, in volume, all day long.

This second job has documented consequences for the mental health of those who do it. Lawsuits are under way on this subject.

Where and under what conditions

This work is largely outsourced to countries where labour is cheaper, Kenya and the Philippines in particular. Some states are starting to impose a minimum wage for these jobs, which is a sign that they exist in large numbers and that their conditions are a problem.

Why this lesson is here

Not to make anyone feel guilty. For a simpler reason: "AI does this all on its own" is false, and that phrase is repeated everywhere. There are people behind it, at every stage, and not knowing that skews your understanding of what this technology really is.

It also changes how you read the news. When a model suddenly becomes more polite, more cautious, better at refusing a dubious request, that is not a scientific breakthrough: it is very often the result of thousands of hours of human judgement.

What you can do about it

Nothing heroic. Knowing it exists, and mentioning it when someone claims in front of you that an AI "learns on its own". That is already putting the thing back in its place.