[YOU] on AI insists that the amplifier does not choose what it amplifies—the human brings the signal. O'Neil's framework shows what the signal has historically contained: the assumptions and prejudices of the builders, dressed in the costume of neutrality. Every choice in the construction of a model—what data to train on, what variable to optimize, what group to treat as the default—is a choice that reflects someone's interests. When those choices are encoded in mathematics and deployed at scale, they do not become less partial. They become less visible. And invisibility, in a system that wields real power over people's lives, is a form of political insulation.
The concept operates as the foundation for O'Neil's entire diagnostic program. Once you see that a model is an opinion, you can ask the right questions: Whose opinion? In whose interest? What is being measured, and what is being ignored? What does the model define as success, and for whom? These are not technical questions. They are political ones, and O'Neil's insistence on their political character is what distinguishes her from reformers who believe that better data or more diverse teams will solve the problem without changing the structure of accountability.
O'Neil illustrates the concept with an example drawn from ordinary life: the informal model she uses to feed her family. The data is what is in the kitchen. The constraints are time and budget. But the definition of success—what counts as a good meal—is hers, shaped by her values. Her young son might build a very different model, one that optimizes for the quantity of candy consumed. Same data, same household, radically different models—because the definition of success is not given by the data but imposed by the modeler. Every model embeds a definition of success, and that definition is a choice that reflects someone's interests.
The insight became politically sharp in O'Neil's years inside finance, where she watched mathematical models that were formally elegant and factually accurate be used to obscure rather than reveal, to lend a veneer of scientific rigor to practices that were neither rigorous nor honest. A model does not lie. But the choice of what to model, what to optimize, and what to count as a good outcome can be deeply dishonest, and the model's formal honesty provides cover for that dishonesty. This is the mechanism O'Neil spent a career exposing.
The choices that cannot be made by data alone. Someone must decide what the model is for, what it should predict, and what counts as a successful prediction. Someone must decide which variables to include and which to leave out. Someone must decide what data to train on, knowing that available data reflects the world as it has been, with all its inequities intact. Each decision is a judgment, and each could have been made differently. The model that results is not a neutral mirror of reality but a particular construction, shaped at every step by the priorities of its makers. The mathematics is real. The choices behind it are human, and therefore contestable.
Absorbed bias. Models learn from the past, and the past is full of human decisions shaped by prejudice and inequality. A hiring model trained on a company's historical decisions will learn to replicate those decisions, including whatever discrimination they contained. It does not know it is discriminating. It is finding patterns in the data it was given, and the patterns encode the injustice of the world that produced them. In this way models do not transcend human prejudice—they absorb it, formalize it, and project it forward, while wearing the mask of mathematical objectivity.
The false authority of the formula. When a person makes a biased decision, the bias can be named, challenged, and held to account. When an algorithm makes the same decision, the bias is hidden behind mathematics. The person affected is left to absorb the verdict without recourse, because there is no one to argue with and no reasoning to contest. This absence of recourse is itself a form of harm—a denial of the basic dignity of being heard—and it falls, like other harms, on those with the least power to demand it. O'Neil's phrase is an attempt to restore the recourse: to insist that behind every model there is a choice, and behind the choice there is a person with interests, and that person can be held responsible.