The Unseen Labour Behind "The AI Figured It Out"

The phrase "the AI learned to do it" hides a great deal of human effort. Before a model can recognise a defect, moderate a post, judge whether an answer is good or extract a field from a document, people have to show it what right looks like — thousands or millions of times. They label images, mark which of two responses is better, write the reference answers, flag the harmful examples, and correct the model's mistakes so it can be trained not to repeat them. This annotation work is the foundation the whole edifice sits on, and it is performed by people, at scale, largely out of sight of the users who benefit from the result. The reason it matters beyond fairness to the workers is that the quality of the labels sets a hard ceiling on the quality of the model. A system trained on sloppy, inconsistent or biased annotations learns sloppiness, inconsistency and bias with perfect fidelity — it has no way to be better than the examples it was shown. If annotators disagree about what a category means, the model inherits the confusion. If they were rushed, underpaid and given no clear guidance, the errors they made under those conditions become the errors the model makes forever. "Garbage in, garbage out" is not a slogan here; it is the mechanism, and it explains a surprising share of why some deployed systems quietly underperform their impressive demos. This has consequences for anyone building or fine-tuning a model on their own data, which is an increasingly common thing for ordinary businesses to do. The temptation is to treat labelling as a low-value chore to be done as cheaply and quickly as possible, and it is precisely the wrong instinct. The examples a model is trained on are more determinative of its behaviour than almost any other choice, so the guidelines the labellers follow, the checks on their consistency, and the care taken over the ambiguous cases are not overhead — they are the product. A smaller, cleaner, well-judged dataset frequently beats a larger, noisier one. The uncomfortable ethical dimension travels alongside the practical one. The human judgement that makes these systems work is often bought at the lowest price the market allows, sometimes involving reviewing disturbing material to keep it out of the model, and the people doing it are rarely visible in the story of a clever machine. As the field matures, the organisations that get the best results are the ones that stop treating annotation as an invisible commodity and start treating it as skilled work that deserves clear instructions, fair conditions and real quality control — because the shortcut taken at the labelling stage is the flaw that shows up, unfixably, in everything the model does afterwards.