Luhmann, or: An AI Answer Is Not a Decision

The Wisconsin Supreme Court did not give judges permission to let an algorithm sentence a defendant. Its 2016 decision in State v. Loomis allowed a court to consider a COMPAS risk assessment only within explicit limits. The score could not decide incarceration or sentence severity. It could not be the determinative factor in deciding whether community supervision would be safe and effective. A court using it had to identify independent reasons for the sentence.¹

Those constraints are easy to lose when the case is summarized as “AI in sentencing.” COMPAS was not a present-day generative AI system. The opinion described it as a proprietary risk-and-needs assessment. It combined information from the criminal file and an interview and displayed three risk categories on scales from one to ten: pretrial, general recidivism, and violent recidivism risk.

The exact path from calculation to judgment is more revealing than the broad label.

How the score entered the case

Wisconsin charged Eric Loomis with five offenses after alleging that he drove in a drive-by shooting. Loomis denied involvement in the shooting and pleaded guilty to two less serious offenses: attempting to flee a traffic officer and operating a vehicle without the owner’s consent. The remaining counts were dismissed but read in for sentencing.

The presentence investigation report attached a COMPAS assessment showing high risk on all three scales. When ruling out probation, the trial court referred to the assessment alongside the seriousness of the crime, Loomis’s criminal history, and his record under supervision.

Loomis sought resentencing. Among his due-process arguments, he contended that the proprietary method prevented him from challenging the scientific validity and the weighting of factors, and that COMPAS took gender into account. The Wisconsin Supreme Court affirmed the denial of relief. In this record, it concluded, the score had not been determinative and independent factors supported the sentence.

The court nevertheless required every presentence report containing COMPAS to warn the sentencing court that:

• the weighting and calculation had been withheld as proprietary;
• the scores identified risk groups, not an individual’s specific probability of reoffending;
• studies had raised concerns about disproportionate high-risk classification of minority defendants;
• no cross-validation for a Wisconsin population had been completed at that time;
• changing populations required continued monitoring and renorming; and
• COMPAS had been developed for corrections decisions about treatment, supervision, and parole, not sentencing.¹

Loomis could see the same report as the judge and challenge listed inputs and resulting scores. He could not inspect the proprietary weighting that produced them. The decision did not make that tension disappear. It held that this particular use, subject to its limitations and cautions, did not violate due process.

The sequence therefore contains several distinct acts: COMPAS calculated; the presentence report assigned the calculation a procedural place; the prosecution and court invoked it; the judge imposed and explained a sentence; appellate review assessed whether that use was lawful. Compressing the sequence into “the algorithm decided” conceals the institutional choices.

Luhmann’s source concept: decision communication

Niklas Luhmann analyzes organizations through decisions communicated in ways that connect earlier selections to later ones. A decision communication does more than transmit a result. It presents a selection as attributable to the organization: this option now holds, although alternatives were possible.²

Applied to Loomis, the calculation alone was not the sentencing decision. The three risk classifications became organizationally consequential when they entered the presentence report, were advanced as a factor, and received weight in the court’s reasoning. The sentence remained a judicially attributed and communicated selection.

This distinction does not exempt the technical system from validation. It prevents a category error about accountability. The roles that authorize an output, give it procedural status, and use it as a reason are making organizational choices. Saying that the system decided obscures them.

Luhmann’s source concept: decision premise

Organizations cannot reconstruct every matter from the beginning. They establish decision premises — programs, responsibilities, communication routes, and personnel choices that structure later decisions without fully determining every case.²

The Loomis opinion makes several such premises visible. The report format selected which risk categories appeared. Corrections practice defined how the assessment was expected to be used. The Wisconsin Supreme Court then established legal premises for future use: some purposes remained available, others were prohibited, and warnings and independent supporting factors became necessary.

In a contemporary scoring system, the consequential selection may similarly precede any individual output. Someone defines categories and thresholds. Someone decides whether a score merely advises, triggers review, or blocks a transaction. Someone assigns permission to override it. These upstream choices shape many later cases without their owners being present in each one.

Luhmann’s source concept: uncertainty absorption

Luhmann uses uncertainty absorption for the compression through which later communication takes up the result of an earlier selection without reproducing all of its uncertainty.² Organizations need this reduction. Otherwise each decision would have to reopen every datum, alternative, and judgment behind its predecessors.

COMPAS supplied such a compression. File and interview data, comparison groups, model assumptions, and proprietary weights appeared as three risk bars. “High recidivism risk” could then travel as a compact premise. Its portability made it useful. It also made group-level inference, validation limits, and undisclosed weighting easier to leave behind.

Uncertainty absorption is not a software defect in Luhmann’s theory. Nor is it a special property of AI. The concept identifies a control point: when a compressed selection is accepted as a sufficiently reliable premise for another decision.

Esposito’s continuation: artificial communication

Elena Esposito extends the communication perspective to learning algorithms. In Artificial Communication, she proposes that the first question should not be whether machines think as humans do. Their social relevance arises through the ways people communicate with algorithmically generated outputs and allow those outputs to participate in further communication.³

The continuation reframes COMPAS without retrospectively turning it into a thinking agent. Its score mattered because it could be displayed in a report, cited in court, and taken up in a reason. Esposito does not supply a legal rule for Loomis. She helps explain why communicative uptake is more consequential than an anthropomorphic label.

The chain of argument should remain explicit:

1. Luhmann’s source concepts describe decision communication, decision premises, and uncertainty absorption.
2. Esposito’s continuation directs that perspective toward algorithmic output.
3. The normative requirements below are this publication’s conclusions for Intelligence Architecture, not a ready-made design program from either author.

Our normative consequence: status, appeal, ownership

First, document the status of the output. The deployment record and interface should state whether a score is information, a trigger for review, a rebuttable presumption, or a binding rule. Permitted and prohibited uses belong next to the output and in the decision policy. A role relying on it should record the reasons that support the decision independently.

Second, provide a real contest and appeal route. Affected people and reviewers need a way to correct inputs, dispute the meaning of an output, and request a reasoned reconsideration. The record must show which output was used, what status it had, and how an objection was answered. Applicable law determines which legal appeal rights exist; architecture should make those rights exercisable rather than merely mentioning them. Loomis was able to pursue post-conviction and appellate review, but the proprietary weighting still remained outside his inspection.

Third, name an owner for each decision premise. A responsible role must own thresholds, risk categories, validation populations, and the actions a score is allowed to trigger. That ownership includes authority to revise or suspend the use. Responsibility for an individual decision remains with the role that adopts the output as a reason.

A machine-generated answer can profoundly structure later choices. It still does not acquire organizational force on its own. That force comes from the status an organization gives it, the route through which it may be challenged, and the accountable decision to treat a compressed output as a premise.

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. ↩