ChatGPT for clinical and care reports: Separate confidential inputs from plausible errors
Fluent AI writing is not evidence of a correct report. Approved processing and professional review matter more than a product label.
Révision éditoriale :
Le texte de cet article est disponible en anglais dans cette version du site.
Contexte et sources
AI-assisted writing involves both submitting information and using the output. The German Data Protection Conference recommends defining purposes, data categories, legal grounds and responsibilities in advance. A personal account is not an approved documentation system simply because it writes fluently. [1]
Professional secrets additionally require assessment under section 203 of the German Criminal Code. Necessary service-provider involvement can be lawful under conditions; neither automatic criminality nor automatic permission should be assumed. [2]
OWASP identifies plausible false output as an LLM risk. Reports can be unsafe if a model adds unobserved findings, removes uncertainty or turns an observation into a diagnosis. [3]
Portée et limites
No fixed claim about training or retention across all ChatGPT plans is made. Assess the actual service and settings. Removing names alone does not establish anonymisation.
En termes simples
Fictional example: A model expands a brief note by adding a normal examination that never happened. The problem is an invented fact, not the quality of the writing.
Recommandations pratiques
Our recommendation: Approve specific services and data types after checking contracts, access, retention, transfers and secondary use. A paid plan is not complete evidence of compliance.
Start with fictional cases and test instructions to leave missing information missing. Compare every draft with the original evidence, especially negations, timing, diagnoses and interventions. Assign an accountable reviewer and preserve traceable corrections. Report accidental submissions to unapproved services for incident assessment.