Shadow AI: Confidential information in unapproved services
Unapproved use obscures data flows. Processing an input, retaining it and training on it are separate questions.
Revisione editoriale:
In questa versione del sito il testo dell’articolo è disponibile in inglese.
Contesto e fonti
The German data protection authorities’ May 2024 guidance provides criteria for selecting and using AI applications. The assessment concerns the particular data processing and provider conditions. An input screen or a cloud label does not establish whether prompts are retained or used for additional purposes. [1]
Ambito e limiti
This article does not claim that a named company’s confidential inputs were proven to enter model training. That would require separate evidence.
Spiegato in modo semplice
A hypothetical employee uploads a confidential contract for summarisation without checking the recipient or conditions. The control gap begins with the unknown data flow, before any later disclosure.
Raccomandazioni pratiche
We recommend listing approved tools and permitted data categories in accessible language. Provide a usable route for common tasks and record relevant contractual commitments and settings. A paid plan or European branding is not a substitute for checking. Make it easy for staff to report uncertainty before more documents enter the same workflow.