AI and privacy
What the rules say about patient data in an AI chatbot, what supervisory authorities and medical organisations advise, and what to arrange per use.
For AI in healthcare, the answer for a free or personal account is no. There is no data processing agreement and professional secrecy is not safeguarded. With a business subscription, it is possible only if the healthcare provider has assessed and documented it beforehand, with an agreement, a DPIA and arrangements about what may and may not be entered.
Client file summaries, consultation reports, plain-language letters: this already happens in practices and institutions. This article concerns what is permitted with patient data. The page about AI chatbots and the GDPR covers the general question of what happens when you upload a client file. Here we explore healthcare.
ChatGPT, Copilot and Claude each have consumer and business versions, with different terms per subscription. Find the comparison in which subscription for business use. Those terms change regularly. Read them yourself before deciding and record the date.
Article 9 GDPR lists health data among the special categories of personal data. Paragraph 1 prohibits processing. Paragraph 2 gives the exceptions. Read about other categories in the article on special category data.
For healthcare professionals, paragraph 2(h) is the main provision: processing necessary for, among other things, medical diagnosis and providing healthcare. Paragraph 3 adds a condition. The data must be processed by or under the responsibility of a professional subject to professional secrecy, or by another person under a duty of confidentiality.
The Netherlands elaborates this in Article 30 of the Dutch GDPR Implementation Act. Paragraph 3 lifts the prohibition for healthcare professionals, institutions and facilities insofar as processing is necessary for proper treatment or care of the data subject, or managing the institution or professional practice. Paragraph 4 adds that only persons subject to confidentiality may then process the data.
The exception covers what is necessary for treatment or practice management. It does not cover everything convenient. Anyone putting a complete patient file into a chatbot to draft one letter must explain why the entire file was needed. Usually, it was not.
On its page about using and sharing health data, the Autoriteit Persoonsgegevens (AP), the Dutch data protection authority, writes that a healthcare provider processing health data on a large scale carries out a DPIA and appoints a data protection officer. According to the AP, even a small healthcare practice usually requires a record of processing activities. See what an entry looks like in the article on the record of processing activities.
Professional secrecy applies alongside the GDPR. For medical treatment, it appears in Article 7:457 of the Dutch Civil Code, part of the rules known as the WGBO.
Paragraph 1 requires the healthcare professional to ensure that no information about the patient, or access to or copies of patient file data, is given to others without the patient's consent. Paragraph 2 says who is not considered “others”: those directly involved in performing the treatment agreement and the professional's substitute, insofar as disclosure is necessary for their duties.
The legislation does not mention software suppliers. Healthcare providers have long worked with patient record system suppliers that may see data as processors. Whether an AI chatbot provider similarly fits within professional secrecy depends on the contract, technology and purpose. This is a legal assessment you make for each application with your data protection officer or a lawyer. This article offers no judgement on it.
One thing is clear. A personal account without an agreement, where the provider decides what to do with input, differs from a processor acting only on instructions. Read how other professions with confidentiality duties handle this in the article on professional secrecy and AI.
Article 7:454 of the same code requires the healthcare professional to establish a patient file and include data on the patient's health and procedures performed insofar as necessary for proper care. If AI creates a report, it remains your file and your report.
On 6 August 2024, the AP published a warning about AI chatbots and data breaches. It had received several data breach notifications because employees shared personal data about patients or clients with an AI chatbot. The AP names ChatGPT and Copilot as examples of digital assistants used in workplaces.
In one reported breach, a general practice employee entered patients' medical data into an AI chatbot against the arrangements. The AP writes that medical data is highly sensitive, receives extra statutory protection for good reason, and sharing it casually with a technology company is a major violation of the people concerned's privacy. The statement does not name the chatbot or number of patients.
The AP draws a distinction. If an employee uses a chatbot independently against the arrangements and enters personal data, there is a data breach. If use forms part of organisational policy, it is not a data breach, but according to the AP is often not legally permitted. Organisations must prevent both situations.
The article on a data breach through an AI chatbot explains what to do afterwards.
On 10 February 2025, the IGJ urged healthcare providers to take care when introducing generative AI. The inspectorate lists automated documentation through speech recognition, patient file summaries and patient communication chatbots as initial applications. According to the inspectorate, the same frameworks apply as for any digital care, with additional specific risks: generative AI can produce incorrect or incomplete information, potentially causing a medication error or incorrect patient assessment. It says healthcare providers using generative AI are legally responsible for careful implementation and use.
The publication contains questions for three stages: before purchase, during implementation and after deployment. The inspectorate also notes that an AI application supporting diagnosis or treatment may be a medical device, subject to separate European rules.
The KNMG medical federation has an AI resource, updated on 9 April 2026, with two practical dilemmas. On AI documenting patient conversations, the KNMG writes, among other things:
Asked whether a doctor may use ChatGPT for administrative tasks, the KNMG answers that this first depends on the healthcare institution's rules, great caution is required and the user remains responsible for use and its consequences.
On its page about AI in general practice, the Dutch College of General Practitioners writes that there is still limited experience of AI in general practice and limited scientific evidence of effectiveness. The NHG sees opportunities in triage, consultation preparation and speech-driven reporting, and is developing practical recommendations.
The table below helps organise applications. It does not replace a DPIA or legal judgement.
| Application | Which data | What you arrange |
|---|---|---|
| Rewriting a general leaflet or information text | No patient data | A permitted tool; medical accuracy checked by a healthcare professional |
| Searching protocols and work instructions | No patient data, provided the question does not describe a patient | Agreement that questions do not identify patients; check source references |
| Drafting a letter or email to a patient | Name, complaint, treatment | Business subscription with data processing agreement; only data the letter needs; checking before sending |
| Summarising a patient file for handover | The complete patient file | DPIA; professional secrecy assessment; preferably within the patient record system or using pseudonymised text; check omissions |
| Recording and transcribing a consultation | Audio, everything the patient says | Consent as the KNMG describes; do not retain recording; DPIA; check report before filing |
| Chatbot answering patient questions | What the patient enters | Clarify it is AI; limits on what it answers; healthcare professional oversight; DPIA |
| Support for triage or diagnosis | Symptoms, measurements, medical history | Check medical device status and AI Act classification; the doctor decides |
This application requires the most. Audio is recorded, a patient speaks who did not request the tool, and the result enters the patient file. Read what recording and transcription involve, including outside healthcare, in the article on AI meeting notes and the GDPR.
The EU AI Act classifies applications by risk. High-risk system requirements have been postponed, some until late 2027. The AI Act explained sets out the dates and what already applies.
PrivaSearch is neither a patient record system nor a medical device. It is a document search layer: it removes personal data from text before an AI assistant searches it. Names, BSNs (Dutch citizen service numbers), addresses and other detected data are replaced with pseudonyms. A colleague then sees the original alongside the pseudonymised text, adds what detection missed and approves it. Only then is a document searchable. Find the steps on the page on how it works.
Two qualifications matter for healthcare. Detection targets data pointing to a person, such as names, addresses, BSNs and dates of birth. Therefore assume medical content, such as diagnosis or treatment, remains readable. And a combination of details may still identify someone, particularly in a small community. That is why human review is mandatory. Pseudonymisation is not anonymisation: for your organisation, the data remains personal data and health data. Read the difference in pseudonymisation or anonymisation.
You can search within PrivaSearch or connect an assistant: Claude, ChatGPT Business or Mistral Vibe (formerly Le Chat). The assistant then sees only released passages, and what PrivaSearch returns reaches that assistant's provider. Anything you type into the assistant yourself goes there directly. Details appear on the page about AI chatbots and the GDPR.
You assess whether this fits your practice's or institution's professional secrecy, with your data protection officer. PrivaSearch supplies starter documents, including a DPIA starter document and register entry, described on the page about the GDPR and the AI Act. These are not completed assessments. To see it with fictitious patient files, book a demo.
For texts without patient data, such as a general leaflet, this is possible if the practice permits it. Patient data is different: a personal account has no data processing agreement and does not safeguard professional secrecy. The KNMG writes that use first depends on the healthcare institution's rules, great caution is required and the user remains responsible.
According to the Autoriteit Persoonsgegevens, yes, when it happens independently and against employer arrangements. The authority received such a notification from a general practice. In many cases, notification to the AP and data subjects is then mandatory. If the organisation itself permits use, it is not a data breach, but according to the AP is often not legally permitted.
That depends on the application. Processing health data necessary for treatment has a statutory exception requiring no consent. For recording a conversation with an AI application, the KNMG says patient consent is required. Professional secrecy also requires consent for disclosure to anyone not directly involved in treatment.
The GDPR, with extra protection for health data, the Dutch GDPR Implementation Act, medical confidentiality under the Dutch Civil Code and the AI Act. If an application is intended for diagnosis or treatment, European medical device rules may also apply. The Health and Youth Care Inspectorate supervises care quality and safety; the Autoriteit Persoonsgegevens supervises personal data handling.
Not automatically. An assistant being part of existing software says nothing about which documents it can read, who sees the answer and what is retained. The Autoriteit Persoonsgegevens names Copilot alongside ChatGPT in its data breach warning. A healthcare provider assesses it beforehand with a DPIA and attention to who can access which patient files.
AI may create a draft; the healthcare professional remains responsible for the patient file. The KNMG recommends checking the automatic report for factual errors, misinterpretations and who said what, and adding non-verbal communication. The inspectorate notes that generative AI can produce incorrect or incomplete information. Rereading is therefore a fixed part of the work.
Where an uploaded file ends up, how long it stays, who can access it and what deletion means. With four approaches for files containing personal data.
How schools use AI, why pupil data receives extra protection, which applications are prohibited or high risk, and what a school needs to arrange.
What ChatGPT stores, who can access it, how to turn off training, and what temporary chats, memory and shared links do. Includes a practical checklist.
We show it with fictitious client files from your sector: upload, detection, review, search and the connection with your AI assistant. Afterwards you'll know exactly what your team can use safely.
Book a demo →