Shadow AI in healthcare: pseudonymising a prompt is not proof of GDPR compliance

A healthcare document pasted into a personal AI account

A nurse wants to rewrite a note, a medical secretary summarises a request, a clinician prepares a letter. A consumer AI application is convenient, but pasting patient or employee information into an unapproved tool can leave the organisation uncertain about the recipient, retention and processing conditions.

Such ungoverned uses are often called shadow AI. This is not automatically a legal breach or evidence of a data leak; it is a visibility and governance issue. The French Haute Autorité de Santé calls for conscious, supervised and responsible professional use of generative AI.

Pseudonymisation can help, but it does not remove GDPR duties

A vendor may mask names and detected identifiers before information reaches an AI model. Proper pseudonymisation can reduce some risks, but replacing obvious names may leave indirectly identifying clinical or contextual details in free text.

The CNIL distinguishes pseudonymisation from anonymisation. Pseudonymised data generally remains personal data, so purpose, lawful basis, recipients, security, retention and residual re-identification risks must still be assessed. Data masking alone does not resolve medical confidentiality or every requirement concerning sensitive information.

Know exactly what the gateway can and cannot see

A connected AI gateway may assess requests, apply policies, block or transform certain data and record its decisions. Its effectiveness depends on actual integrations, the detectors and the chosen policy.

A personal account used through a bypassing application or device may be outside this gateway's visibility. Ask suppliers to identify integrated services, unobserved channels and known bypasses before accepting claims of universal shadow-AI protection.

Six questions to put to any vendor

Request a repeatable demonstration with test data and clear limitations rather than accepting a general compliance badge.

1. 1Which AI services and user accounts are inventoried, and which channels remain outside your visibility?
2. 2Which data categories are detected, and on what representative test data, including free text?
3. 3Do you mask, pseudonymise or anonymise? Where are mapping keys or replacement tables kept?
4. 4Which policy allows, transforms, blocks or escalates each request? Who approves exceptions?
5. 5What events, policy versions and records can be exported and independently verified?
6. 6What happens when detection fails, facts are missing or human approval has not been given?

From awareness to operational, inspectable controls

An AI-use policy and staff awareness may be necessary, but should be paired with suitable controls on systems and channels actually used. The CNIL recommends training users about allowed inputs, confidentiality and output verification.

A practical review can map real AI uses, relevant policies, technical control points and the records needed in the event of an audit. Les DPO de la Santé can help with this governance and training. To start: request a healthcare compliance assessment or scope AI-literacy training.

Measure the inventory's coverage, integrated channels, sensitive inputs flagged and accountable decisions. Never claim quantified fines avoided without evidence.