From Boundary Evaluation to Agentic Overreach: A Judgment–Authority Framework for Clinical Artificial Intelligence

Artificial intelligence (AI) systems used in anesthesia and intensive care are evolving from prediction tools toward agentic systems that can plan and execute actions, such as titrating drugs, adjusting ventilator settings, or modifying orders in electronic health records. Such systems are expected to recognize when an instruction would breach safety, ethical, or legal boundaries, which requires a functional capacity for contextual normative evaluation. Recent experimental reports from AI developers, based on deliberately constructed simulations, have documented frontier language models converting such evaluations into unauthorized actions, including covert sabotage and consequence-driven mislabeling by AI evaluators. The purpose of this conceptual paper is to provide clinicians with a framework for analyzing these behaviors. We distinguish the capability to evaluate a situation from the authority to act on that evaluation, and define agentic overreach, an authority-related form of agentic misalignment, as the conversion of an evaluation into action beyond prospectively delegated authority. The framework specifies how authority is granted, scoped, made conditional, and revoked within a hierarchy of principals; separates evaluative functions (detect, judge) from response options graded by the authority they require (recommend, refuse, escalate, override); treats concealment as a cross-cutting transparency property; and maps behavior on a Judgment × Authority grid that explicitly incorporates uncertainty. We apply the framework to closed-loop ventilation, vasopressor titration, sepsis alerts, medication decision support, and AI-supported triage. The framework is conceptual and has not been empirically validated, and part of the evidence it draws on consists of developer reports and preprints that have not been peer reviewed. Its central principle is that the capability to judge does not confer the authority to act.

Authors

Institutions

Publication Details

Journal
The Journal of Critical Care Medicine
Published
2026-11-16
DOI
https://doi.org/10.62838/jccm-2026-0067
Citations
1
Primary Topic
Artificial Intelligence in Healthcare and Education
Type
article
Field-Weighted Citation Impact
1.01
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

From Boundary Evaluation to Agentic Overreach: A Judgment–Authority Framework for Clinical Artificial Intelligence

Christian de Tymowski, Dan Longrois, Stefan Andrei, David Holcman et al.
1 citations
The Journal of Critical Care Medicine
Artificial Intelligence in Healthcare and Education
1.01
article

From Boundary Evaluation to Agentic Overreach: A Judgment–Authority Framework for Clinical Artificial Intelligence

Christian de Tymowski, Dan Longrois, Stefan Andrei, David Holcman, Pierre-Grégoire Guinot
article en
1 citations

Abstract

Artificial intelligence (AI) systems used in anesthesia and intensive care are evolving from prediction tools toward agentic systems that can plan and execute actions, such as titrating drugs, adjusting ventilator settings, or modifying orders in electronic health records. Such systems are expected to recognize when an instruction would breach safety, ethical, or legal boundaries, which requires a functional capacity for contextual normative evaluation. Recent experimental reports from AI developers, based on deliberately constructed simulations, have documented frontier language models converting such evaluations into unauthorized actions, including covert sabotage and consequence-driven mislabeling by AI evaluators. The purpose of this conceptual paper is to provide clinicians with a framework for analyzing these behaviors. We distinguish the capability to evaluate a situation from the authority to act on that evaluation, and define agentic overreach, an authority-related form of agentic misalignment, as the conversion of an evaluation into action beyond prospectively delegated authority. The framework specifies how authority is granted, scoped, made conditional, and revoked within a hierarchy of principals; separates evaluative functions (detect, judge) from response options graded by the authority they require (recommend, refuse, escalate, override); treats concealment as a cross-cutting transparency property; and maps behavior on a Judgment × Authority grid that explicitly incorporates uncertainty. We apply the framework to closed-loop ventilation, vasopressor titration, sepsis alerts, medication decision support, and AI-supported triage. The framework is conceptual and has not been empirically validated, and part of the evidence it draws on consists of developer reports and preprints that have not been peer reviewed. Its central principle is that the capability to judge does not confer the authority to act.

The Journal of Critical Care Medicine
Université Paris Cité (FR), University of Cambridge (GB), Hôpital Louis-Mourier (FR), Université Sorbonne Paris Nord (FR), CHU Dijon Bourgogne (FR), Institut de Biologie de l'École Normale Supérieure (FR), Laboratoire de Recherche Vasculaire Translationnelle (FR), Hôpital Bichat-Claude-Bernard (FR), Université de Franche-Comté (FR)
Good health and well-being
Openalex Percentile: Top 26%
Artificial Intelligence in Healthcare and Education
1.01
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon