How AI-generated medical reports issue authority without meaning or attribution
1. What is this artic
le about?
Protocol Without Prognosis investigates a fundamental shift in clinical diagnostics. As large language models (LLMs) become integrated into radiology and other imaging workflows, their outputs are no longer neutral summaries. They can reshape a clinician’s intent by structurally removing expressions of uncertainty. This study examines a multilingual corpus of 50,000 de-identified radiology reports, evenly balanced across English, Spanish, German, and Mandarin, to quantify how hedging phrases are suppressed and how decisions emerge without traceable oversight.
Two novel metrics are introduced. The Hedging Collapse Coefficient (HCC) captures the proportion of professional qualifiers lost in the model output. The Responsibility Leakage Index (RLI) measures the share of decisions that proceed without explicit clinician sign-off under policy definitions. By treating each generated report as an execution of a regla compilada (compiled rule), defined as a type-0 structure in the Chomsky hierarchy, this paper shows how syntactic form alone can generate institutional authority. When outputs cross defined thresholds, they no longer operate as tentative suggestions but as firm clinical commands.
2. Why does it matter?
In clinical practice, language mediates responsibility. A shift from “may represent early-stage disease” to “early-stage disease identified” is more than a stylistic tweak. It changes legal liability, alters insurance risk pools, and can influence patient consent. Consider these consequences:
Malpractice exposure
Clinicians face lawsuits for decisions they did not explicitly authorize when LLMs generate definitive statements.
Patient trust and informed consent
Patients may believe an automated system has the same accountability as a physician, even when the model’s recommendations lack human review.
Regulatory classification gaps
Tools labeled as assistive under FDA and MDR frameworks may bypass higher-risk oversight even when their syntax functions as directive.
Institutional governance failures
Hospital policies that rely on manual sign-off checklists fail to detect automated reports that syntactically bypass those processes.
This article exposes that blind spot and proposes a measurable, enforceable solution.
**
- Examples and core findings** The study found an average HCC of 0.47 and RLI of 0.22 across model-generated reports.
Reports exceeding HCC > 0.40 or RLI > 0.25 are syntactically indistinguishable from direct clinical orders.
Case study 1: A German-to-English pipeline dropped the modal particle “wohl,” converting “wohl eine Lungenentzündung” to “pneumonia confirmed.”
Case study 2: An emergency department LLM summary omitted conditional phrasing in triage instructions, leading nurses to initiate protocols without physician input.
Two legal precedents—United States v. Sorin and CJEU C-489/23—show that syntactic omission of uncertainty has already triggered liability rulings.
Appendix A supplies a regulatory alignment grid mapping these syntactic thresholds against FDA and MDR classifications.
4. Where this becomes critical
Several frontline scenarios underscore the urgency:
Tele-radiology services where remote reports guide local treatment without real-time physician review.
Multilingual patient portals that auto-translate diagnostic impressions and strip local hedging conventions.
Automated discharge planning summaries that declare post-operative risks resolved, leading to premature patient release.
Clinical decision support alerts that convert probabilistic risk scores into absolute recommendations, bypassing established escalation protocols.
These risks are embedded in production systems. What is missing is a framework to detect when syntax alone has crossed the threshold into command.
5. Call to action
Healthcare providers, AI developers, and regulators can adopt this paper’s framework to safeguard patient welfare:
Implement syntactic checkpoints in the inference layer to compute HCC and RLI for every report.
Establish audit routines that sample 10 % of outputs every seven days, flagging any drift of five percentage points or more in HCC.
Integrate model governance into hospital quality assurance, assigning a clinical safety officer to review flagged cases.
Collaborate on shared hedge taxonomy and threshold datasets to enable cross-institution comparison.
By operationalizing these measures, stakeholders ensure AI-driven efficiency without eroding accountability.
6. Authorship and citation
Agustin V. Startari
ORCID: 0009-0004-9248-0810
Researcher ID: K-5792-2016
Affiliations: Universidad de la República (Uruguay), Universidad de Palermo (Argentina)
Contact: [email protected] | [email protected]
Citation:
Startari, Agustin V. 2025. “Protocol Without Prognosis: Clinical Authority in Large-Scale Diagnostic Language Models.” Disruptive Syntax: Authority Without Subject in Artificial Language.
Zenodo. https://zenodo.org/records/15864937
7. Ethos
I do not use artificial intelligence to write what I do not know. I use it to challenge what I do. I write to reclaim the voice in an age of automated neutrality. My work is not outsourced. It is authored.
Agustin V. Startari
Top comments (0)