Truace
Subscribe

User Beware: Inaccuracy and Inconsistency of Large Language Models in Providing Precision Dosing Recommendations for Patients With Kidney Impairment-A Case Series

Introduction Evidence-based dosing guidance for medications in critically ill patients with acute kidney injury (AKI) and receiving continuous kidney replacement therapy (CKRT) is limited. Freely available large language models (LLMs) can generate confident, human-like outputs. The accuracy and reproducibility of LLMs in providing precision drug dosing recommendations in the setting of AKI and CKRT have not been evaluated. Objectives We sought to characterize literature concordance and internal consistency of LL…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1498· cite & verify
User Beware: Inaccuracy and Inconsistency of Large Language Models in Providing Precision Dosing Recommendations for Patients With Kidney Impairment-A Case Series
"Plugged into dialysis" by newslighter is licensed under CC BY 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/2.0/.

In brief

Six investigators put three standardized clinical scenarios to the free versions of six chatbots between July and September 2025. The scenarios covered an adult with acute kidney injury, a child on continuous kidney replacement therapy receiving cefepime, and a toddler on high-effluent therapy receiving meropenem. Answers were compared with the primary literature.

Agreement with published dosing averaged 63% and ranged widely, and asking the same model the same question repeatedly could change the recommended dose dramatically. Dosing evidence for these critically ill patients is already thin, so confident but unreliable chatbot output is a safety concern. The study is a small case series of two antibiotics and three vignettes, so it cannot show how often such errors would reach patients.

Main points

  1. Models tested: ChatGPT, Claude, Google Gemini, Microsoft Copilot, OpenEvidence, and Perplexity, in their freely available versions.
  2. Mean literature concordance was 63%, with individual results from 28% to 94%.
  3. Repeated prompts by the same user produced daily maintenance dose variance from 0 to 1000%.
  4. OpenEvidence universally cited relevant sources, while other models cited inconsistently or confabulated sources.
  5. All models recommended therapeutic drug monitoring, but only Claude consistently acknowledged its own uncertainty.

The problem

Six freely available LLMs matched literature-based cefepime and meropenem dosing only 63% of the time on average, and repeated prompts from the same user produced widely varying daily doses.

The rundown

The study addressed a gap: accuracy and reproducibility of LLM dosing advice in acute kidney injury and continuous kidney replacement therapy had not been evaluated, even though evidence-based dosing guidance for these patients is limited.

Each of six investigators queried all six models with three vignettes, and each also prompted one model three times per vignette to test consistency between iterations. Responses were assessed for concordance, use of pharmacokinetic equations, citation accuracy versus confabulation, acknowledgment of uncertainty, and drug monitoring recommendations.

The authors conclude that the models may help generate hypotheses and retrieve literature, but their outputs should not be used in isolation for drug dosing in critically ill patients with kidney dysfunction.

Sources

  1. Peer-reviewedPharmacotherapy: The Journal of Human Pharmacology and Drug Therapy2026-10-01

The debate