Principlist AI Ethics

In The You On AI Field Guide

Principlism in bioethics was developed by Beauchamp and Childress in the 1970s as a response to the need for practical guidance in medical decision-making that did not require agreement on deeper metaphysical or theological questions. The four principles — autonomy, beneficence, non-maleficence, justice — were presented as capturing widely shared moral intuitions and providing a framework for navigating particular cases without requiring agreement on their ultimate grounding. The approach has been enormously influential and has been extended to AI ethics in various forms.

MacIntyre's critique applies to AI ethics principlism with particular force. Consider fairness. A utilitarian specifies fairness as the distribution that maximizes aggregate welfare. A Kantian specifies it as treatment that respects the equal rational agency of all affected parties. A virtue ethicist specifies it as distribution according to merit within a practice. A capabilities theorist specifies it as provision of conditions for the exercise of central human capabilities. These are not variations on a single theme; they are incompatible specifications that can yield opposite practical recommendations in the same case. The principle "AI should be fair" is consistent with all of them and guides none of them.

The institutional attractions of principlism are obvious. It allows disparate stakeholders — ethicists, engineers, regulators, executives — to endorse shared principles without committing to the substantive traditions that give them content. It produces documents that look like ethical frameworks without requiring the hard substantive work of ethics. It generates compliance without demanding transformation. And it conceals the power relations that determine which tradition's specification of the principles actually governs: in practice, the specification is determined not by philosophical argument but by whoever controls the implementation.

The MacIntyrean alternative is not to abandon principles but to embed them in explicit substantive traditions. This does not produce consensus — in fact, it produces the acknowledgment of disagreement that principlism conceals — but it does produce determinate guidance. A virtue-ethics-based AI policy specifies fairness in terms of the internal goods of particular practices and the conditions for their preservation. A utilitarian policy specifies it in terms of welfare metrics. The policies differ, but each provides actual guidance for decisions rather than empty principles that can be interpreted to support any decision.

Origin

The principlist approach in bioethics was developed by Tom Beauchamp and James Childress, Principles of Biomedical Ethics (Oxford, 1979). Extension to AI ethics appears in various forms, notably in the Montreal Declaration (2017), the OECD AI Principles (2019), and the EU High-Level Expert Group's Ethics Guidelines for Trustworthy AI (2019).

Key Ideas

Tradition-independent pretense. Principlism presents itself as offering guidance that transcends particular substantive traditions.

Content-dependent on tradition. In reality, principles acquire content only within traditions; the pretense conceals rather than transcends tradition.

Institutionally attractive. Principlism allows diverse stakeholders to endorse shared principles without committing to substantive traditions.

Operationally empty. Principles without traditions cannot guide decisions in contested cases — the guidance comes from whichever tradition implementers happen to endorse.

Virtue ethics alternative. Explicit tradition-embedded ethics produces determinate guidance at the cost of acknowledging disagreement.