Law as Integrity

In the [YOU] on AI Field Guide

The cycle asks what it would mean to see large language models clearly, and law as integrity offers one of the sharpest answers. An AI legal prediction system is trained on the fit dimension of judicial reasoning: it learns what outcomes followed from what facts, with what frequency, across what corpus of decisions. It has access to the skeleton of integrity and no access to the flesh. A model trained on every legal decision ever rendered learns the central tendency of what judges have done, including every prejudice, every doctrinal dead end, every compromise that integrity would now repudiate. To reproduce that tendency with high fidelity is not to approximate integrity; it is to entrench the past against the critical scrutiny that integrity requires. The better the model's fit to historical data, the farther it stands from what a court deciding under integrity would do.

This inversion has a practical consequence that the cycle emphasizes throughout: the authoritative confidence of a well-calibrated model is evidence of nothing about the quality of the underlying decision. A system can be maximally accurate at predicting what courts have done and maximally wrong about what they should do—and it presents both with identical confidence. The decorrelation of fluency from authority that is the cycle's diagnostic signature appears in its legal form here: the machine's output sounds like law, reads like a judicial ruling, cites real precedents, and may be the opposite of what integrity requires. Dworkin's framework is the instrument that measures the distance.

Origin

The concept was developed at length in Law's Empire (1986), Dworkin's most systematic work, where it appears as one of three interpretive attitudes law might take toward its own past. Legal conventionalism treats past decisions as settling present law through their social pedigree, with discretion filling the gaps. Legal pragmatism treats past decisions as useful but not binding precedents to be followed or departed from as consequences recommend. Law as integrity treats the community's legal history as an author treats the earlier chapters of a chain novel: the interpreter must continue the story in the way that fits what has been written while making the whole as good a work of literature as it can be.

The chain-novel metaphor is important and was deliberate. It shows that fit and justification are not alternative criteria between which the interpreter chooses but simultaneous demands that constrain each other. The chain novelist cannot rewrite earlier chapters, but she need not continue every thread they contain; she must find the continuation that coherently extends the work while making it better. This is what Dworkin thought judges did—not mechanically, not infallibly, but as a genuine aspiration that distinguished their activity from mere policy-making or arbitrary choice.

Dworkin's imaginary judge Hercules personifies the perfect execution of this task—a jurist of limitless skill and patience who can hold the entire body of law in view and construct the theory that best satisfies both dimensions simultaneously. Hercules was never meant as a realistic model of what judges do. He was a measuring instrument: the standard against which real judicial reasoning could be assessed, the answer to what the law requires before the contingent limitations of human cognition are introduced.

Key Ideas

Fit and justification. Every interpretation of a legal text or practice must satisfy two independent criteria. Fit: the interpretation must cohere with enough of the existing legal materials to count as an interpretation of the same practice rather than a replacement of it. Justification: among the interpretations that fit, the one that shows the practice in its best moral light is the correct one. These criteria can pull in opposite directions, and the skill of interpretation is the judgment that finds the best balance. AI legal systems operate on fit alone—not because their designers chose to omit justification, but because justification is not in the data.

The integrity constraint on precedent. Integrity does not require mechanical consistency with every past decision. It requires principled consistency: the judge must be able to articulate a principle that explains both the current decision and the precedents it follows, while treating decisions that fail the same principle as errors that integrity would now correct. A model that maximizes predictive accuracy over the historical record cannot make this distinction. It must treat all past decisions as equally data, with no standpoint from which some of them could be wrong.

The commitment requirement. Producing text that reads as a principled justification is not the same as deciding under integrity. Integrity requires that the decider mean the principle—that she commit to it across future cases, that she would produce the opposite outcome if the principle required it. A model that generates an equally fluent rationale for whatever output it predicts, and would generate a different rationale if the output changed, has not committed to a principle. It has performed the gesture of integrity without its substance, which Dworkin would identify as a form of bad faith more dangerous than open prediction because it wears the legitimating credential while delivering none of the guarantee.

Law as community and legitimacy. Integrity is not merely a method of adjudication but the basis of political legitimacy. A community whose legal decisions cohere around a set of principles that each member can recognize as the community's genuine commitment has a claim on the allegiance of its members that a community whose decisions are merely the outcomes of a series of disconnected power struggles does not. When AI is introduced into adjudication without integrity—when the system produces outcomes that cannot be traced to coherent principle—this claim to legitimacy is undermined, even when the outcomes happen to be accurate. Accuracy and legitimacy are different goods, and Dworkin's framework insists that neither can substitute for the other.