Rules: A Short History of What We Live By

In the AI Story

The book's range is extraordinary. It begins with ancient Greek recipes — technai that bundled procedures with explicit guidance on exceptions, materials, and circumstances. It traces the evolution through Roman legal codes, medieval monastic rules (Benedict's rules for monastic life explicitly include commentary on when and how to deviate from them), Enlightenment attempts to codify legal judgment into algorithmic decision procedures, nineteenth-century bureaucratic rules, and twentieth-century attempts to replace human judgment with algorithmic systems.

Across this range, Daston identifies a recurrent pattern. Thin rules are repeatedly proposed as solutions to problems of scale, consistency, and accountability. They work — partially, and within specified conditions. When conditions vary, they fail. When they fail, human judgment is reintroduced to handle the cases the thin rules could not anticipate. The thick judgment does not replace the thin rule; it supplements it, cleaning up after its characteristic failures. The cycle then repeats, with new proposals for thinner rules that will finally eliminate the need for judgment — and the same outcome.

The book is particularly sharp on the historical relationship between rules and exceptions. Thin rules imagine a world in which exceptions are aberrations — deviations from a norm that could, in principle, be specified more precisely. Thick rules imagine a world in which exceptions are the norm — cases that the rule must handle through the exercise of judgment the rule itself acknowledges is necessary. The history of rules, Daston argues, repeatedly confirms the thick-rule view: the exceptional cases are not marginal phenomena but the substantial bulk of real situations.

The relevance to AI is explicit in the book's concluding chapters. Machine-learning systems implement thin rules at an unprecedented scale — statistical inferences that operate without contextual judgment, without discretion, without awareness of the cases they cannot accommodate. The book's historical argument is that such systems cannot operate reliably without the thick judgment of humans who handle the cases the thin rules fail to predict. The institutional question is whether we are constructing the infrastructure for that thick judgment — and the book's pessimistic answer, developed more fully in the AI volume, is that we are not.

Origin

The book emerged from Daston's decades of research on the history of scientific method, algorithmic thinking, and the relationship between rule-following and judgment. It synthesizes arguments made in earlier works including Classical Probability in the Enlightenment and Objectivity, extending them into a comprehensive historical analysis of rules as a category.

The book was widely reviewed and became an unexpected crossover success, read not only by historians of science but by legal scholars, philosophers, technologists, and policy analysts concerned with the operation of rules in algorithmic systems. It has become a key reference text in discussions of AI governance, machine-learning ethics, and the future of rule-based systems.

Key Ideas

Three millennia of rule-making. The book traces rules from ancient Greek recipes through contemporary algorithms, identifying consistent patterns across radically different domains.

Thick and thin as analytical poles. Individual rules vary in their thickness; the distinction is not binary but identifies two poles of a spectrum that shapes how rules operate.

The thinning tendency is historical. Pressures toward efficiency, scale, and accountability repeatedly drive rules from thick toward thin — but the thinning is never complete.

Exceptions are the norm. Real cases reliably exceed the specifications of thin rules; the exceptional cases are substantial, not marginal.

AI as the latest thinning project. Machine learning implements extremely thin rules at unprecedented scale, requiring thick human judgment that the institutional environment has not adequately developed.

Debates & Critiques

A debate concerns whether the thick/thin framework adequately captures the specific features of algorithmic systems or whether those systems constitute a genuinely new category that resists the old distinction. Defenders argue that the framework remains useful with adjustments; critics argue that statistical inference from large corpora differs qualitatively from rule-following in the traditional sense. A related debate concerns whether Daston's pessimism about the institutional infrastructure for AI-era thick judgment is warranted or whether emerging developments in AI governance, professional standards, and educational curricula will prove adequate faster than the historical pattern suggests.