How an AI Output Acquires Institutional Weight

In 2016, the Wisconsin Supreme Court permitted only a bounded use of the proprietary COMPAS risk score in State v. Loomis: it could inform sentencing, but could not determine imprisonment or sentence severity, or be decisive in assessing whether someone could be supervised safely in the community.¹

The score became consequential because a presentence report carried it into court, the prosecution used it as an argument, and the judge referred to it alongside other reasons. Its precise weighting remained protected as a trade secret. The case therefore does not show that “the algorithm decided.” It shows how an organization turns a technical output into a reason for a decision.

For organizational AI, that translation matters more than the technical category of the system. Whether the output is a score, text, or recommendation, report formats, responsibilities, and rules of use determine its status, who can contest it, and who remains answerable for its effects.

The narrow legal boundary

COMPAS was not a generative AI system. It combined information from a criminal file and an interview to assign risk groups. The court saw Loomis classified as high risk on three scales but also relied on offense severity, criminal history, and prior supervision when ruling out probation. Loomis could contest inputs and results, not the proprietary weighting; he also objected to the use of gender. The Supreme Court permitted this specific use because independent reasons supported the decision.

The required cautions narrowed the score’s claim further. It described groups rather than an individual probability of reoffending, had not then been cross-validated for Wisconsin, and was designed for treatment, supervision, and parole rather than sentencing. Reports also had to surface possible disproportionate classifications of minority defendants and the need for continuing monitoring and renorming.¹ The judgment did not resolve the hidden weighting; it permitted only a use whose effect was constrained by procedural rules.

Why a compact output travels so easily

Niklas Luhmann uses uncertainty absorption for the way later communication accepts the result of an earlier selection without reconstructing all the uncertainty behind it.² Organizations need this compression. Otherwise every decision would have to repeat all the data, alternatives, and reasoning that preceded it.

COMPAS compressed file and interview data, comparison groups, model assumptions, and undisclosed weights into three risk bars. “High recidivism risk” traveled easily into a report and from there into an argument. Its group-level basis, validation limits, and calculation traveled less easily.

For system design, this identifies the sensitive handoff: when is a compressed result considered reliable enough for the next action? Its purpose, uncertainty, and excluded uses must remain visible at that point. A challenge capable of changing the subsequent process belongs there too.

The decisions made before the individual case

Organizations rely on decision premises: programs, responsibilities, communication routes, and personnel arrangements shape later decisions without fully determining each one.² In Loomis, the report format determined which categories reached the court, while corrections practice shaped the expected use. The Supreme Court changed these premises for later cases through prohibitions, warnings, and a requirement for independent reasons.

Comparable choices in contemporary AI systems are often made long before an individual output appears. Someone defines categories and thresholds, determines follow-on actions, and gives some roles but not others authority to override a result. Those choices can govern many cases even when their authors are absent from the later decision.

Luhmann’s decision communication draws another boundary. An output is not yet an attributable organizational selection. In Loomis, the classification acquired institutional weight through the report, advocacy, and judicial reasoning; the sentence remained the court’s decision.² Saying that “the system decided” hides the roles that authorized, classified, and used the output.

What follows for AI processes

In Artificial Communication, Elena Esposito shifts attention away from whether learning algorithms think like people and toward how people communicate with their outputs and carry them into further social communication.³ That perspective explains why an output’s presentation and circulation belong to its effects. It does not supply a legal rule for Loomis, just as Luhmann’s theory of organization is not already an AI governance standard.

For Intelligence Architecture, three design tasks follow:

1. Declare the status: An output may inform, trigger additional review, operate as a rebuttable presumption, or activate a binding rule. Permitted and excluded uses belong in procedure and where the output is displayed.
2. Make challenge effective: Affected people and reviewers can correct inputs, dispute interpretation, and obtain a reasoned reconsideration. The record shows which output was used, by whom, in what role, and how an objection was handled.
3. Assign responsibility: A named role owns the validation basis, permitted consequences, and authority to change or suspend the rules. The individual case remains the responsibility of the role that adopts the output as a reason.

Applicable law determines appeal rights; technical and organizational architecture must make those rights usable. An AI output does not acquire institutional force from the model alone. That force comes from decisions that display it, give it a status, and translate it into action.

Footnotes

1. Wisconsin Supreme Court, State v. Loomis, 2016 WI 68, official opinion filed 13 July 2016. The account of COMPAS, the procedural history, and the required limitations and cautions follows the majority opinion. ↩ ↩²

2. Niklas Luhmann, Organization and Decision, Cambridge University Press, 2018; German original Organisation und Entscheidung (2000), especially the chapters on uncertainty absorption and decision premises. ↩ ↩² ↩³

3. Elena Esposito, Artificial Communication: How Algorithms Produce Social Intelligence, MIT Press, 2022. The normative requirements stated here are this article’s synthesis with the Loomis case, not claims quoted from Esposito. ↩