Rights as Trumps

In the [YOU] on AI Field Guide

The cycle's central observation about intelligence as an amplifier—indifferent to the signal it is fed—finds its sharpest legal form in the trump. When an optimizer is pointed at a population and told to maximize an aggregate, it amplifies the aggregate logic with perfect fidelity. The individual who holds a right against this logic is not asking the optimizer to be less accurate; she is asserting a claim the optimizer's architecture cannot process. Rights as trumps is the concept that translates the cycle's warning into the specific vocabulary of law and justice: the amplifier does not care what signal you feed it, and some signals—the dignity of the person, her standing as an equal—are not the kind of thing that can be fed into an amplifier at all, but only protected from one.

The concept also illuminates the specific inadequacy of “fairness constraints” in machine learning as currently practiced. A fairness constraint that reduces demographic disparity to an acceptable level is a penalty term with a high price, not a trump. The optimizer will pay the price whenever the gain elsewhere exceeds it. Dworkin's analysis shows that this is not a flaw in the implementation of fairness constraints but a structural feature of the approach: any constraint that can be outweighed is not a right in his sense. A right requires a categorical prohibition, not a sliding scale, and the design of systems that deploy consequential automated decisions must distinguish between these two kinds of protection if the protection is to be real.

Ronald Dworkin

Origin

The trump formulation first appeared in Dworkin's essay “Taking Rights Seriously,” published in 1970 and incorporated into his first book of the same name in 1977. It was a direct response to utilitarian political theory, which treated rights as specially weighted preferences to be balanced against other preferences in the social welfare calculus. Dworkin's objection was not that this balance was performed incorrectly, but that the balancing framework itself mischaracterized the nature of rights. A right that survives only when the balance happens to favor it does no work, because the favorable balance would have produced the same outcome without it. The concept did serious work as a philosophical matter because it explained what made rights rights—their insulation from the aggregate calculus—rather than merely naming them as important considerations.

The trump metaphor drew on H.L.A. Hart's earlier discussion of rights as “protected choices,” but sharpened Hart's account by making the anti-utilitarian structure explicit. Dworkin extended the concept through Law's Empire (1986) and Sovereign Virtue (2000), where he grounded rights in the foundational principle of equal concern and respect. The most fundamental right, in his final account, is not any particular entitlement but the background right to be treated as an equal—the right from which all specific rights derive their force.

Key Ideas

The structure of the trump. A trump is not a preference so strong it overrides all others. It is a claim of a categorically different kind that removes certain considerations from the calculus entirely. When Dworkin says rights trump aggregate welfare, he means that the welfare calculation is the correct method for most political decisions and that rights mark the domain where that method may not operate. The boundary of the domain is the boundary of rights, and rights mark it from the outside, not as one very large term within the calculation.

Rights and the optimizer's architecture. The alignment problem, stated in Dworkinian terms, is that every architecture for building objective functions is intrinsically utilitarian: it weighs, sums, and maximizes. A trump cannot be represented in this architecture without being converted into what it is not—a very expensive penalty term. This architectural mismatch means that the protection of rights in automated systems cannot be achieved by better objective design alone. It requires categorical prohibitions implemented outside the optimization loop: hard rules about what the system is forbidden to decide, regardless of the gain that decision would produce.

Accuracy as the enemy of standing. Dworkin's analysis of what equal concern requires—that each person must be addressed in her own right, not as a token of a type—generates a counter-intuitive implication for AI: a more accurate model that more finely classifies people by group membership may be a greater violation of their rights than a less accurate one, because accuracy at the level of the group closes the door on the individual's claim that “I am different.” The right is not to be classified correctly. The right is to be addressed as an individual.