Can you put patient data into an online AI tool?
Not into a consumer AI tool used with a personal account. Health data is protected both by the GDPR and by medical confidentiality, and the practice may only entrust it to a provider with which it has signed a contract that protects it. You can, however, use AI for anything that contains no patient data, and use a properly governed professional tool for the rest. A one-page written rule and fifteen minutes explaining it to the team are enough to get started.
The scenario in brief
Example, made up for illustration. In a fictional GP practice, the receptionist receives a letter from a specialist: the patient's name, date of birth, diagnosis and treatment. To produce the three-line summary the doctor wants, she copies the whole letter into a consumer AI assistant, opened with her free personal account. The summary is good, and done in ten seconds.
Nothing seems to have leaked. Yet a patient's health data has just left the practice for a service with which the practice has signed no contract, which may keep the conversation, and whose account nobody at the practice controls. That is exactly what the GDPR and medical confidentiality ask you to avoid, without forcing you to give up on AI.
What the law actually protects
The GDPR, the European regulation on personal data, places health data among the ‘special categories’: processing it is prohibited in principle, with exceptions. For care, the data must be processed by a professional bound by secrecy, or under their responsibility. As the Data Protection Authority (DPA) points out, health data enjoys stricter protection.
Medical confidentiality comes on top of the GDPR. The Code of Medical Ethics (Article 25, in French) states that it covers all information the doctor learns in the course of their profession, and that the doctor ensures that their staff respect professional secrecy. The Patients' Rights Act (page in French) also protects the patient's privacy, and the Criminal Code punishes breaches of professional secrecy (Article 352 of the new Criminal Code, in force since 1 September 2026, which takes over the former Article 458).
| Rule | What it says | What it means for AI |
|---|---|---|
| GDPR, health data | Processing prohibited in principle, apart from strictly defined exceptions | The data cannot go off to just any service |
| GDPR, processor | A written contract is mandatory with any provider that processes data on your behalf | No contract with the tool = no patient data in it |
| Medical confidentiality | Covers everything the healthcare professional learns, and binds their staff | Reception is bound in the same way as the doctor |
| Patients' rights | The patient is entitled to privacy and to see their file | What is done with their file must be explainable to them |
Why a consumer AI tool is a problem
A consumer AI assistant is an online service open to everyone: you type a question and it answers. Used with a personal account, it creates four problems for a practice.
- No contract with the practice. As soon as a provider processes data for you, a data processing agreement is always mandatory, the DPA reminds us (brochure in French).
- Settings that keep your texts. The DPA advises checking the settings for conversation history and AI training, which are often switched on by default. Training here means your texts may be used to improve the model.
- An account outside your control. An employee's personal account is not managed by the practice: if they leave, the history leaves with them. If their password is stolen, so are the conversations.
- Answers that are sometimes wrong. An AI can make up a figure or a drug name with total confidence: everything it produces must be checked.
The test the DPA suggests before typing anything: ‘If I were at a dinner with unfamiliar people, would I be comfortable sharing this information about myself or someone else?’
Removing the name is not enough
Deleting the patient's name before pasting the text is a good habit, but it is almost never enough: a person remains recognisable from a combination of details, especially in a neighbourhood or a small town.
- Date of birth, address, national register number or file number.
- Dates of consultation or hospital stay.
- A rare disease, an unusual job, a particular family situation.
- The name of the GP or hospital, which places the patient.
- Images: an X-ray, a photo of a wound or a scanned document often show the name or date in the image itself.
The DPA recommends hiding anything that identifies a person: names, addresses, age. But if the patient can still be recognised by cross-checking the text with other information, it is still protected health data. If in doubt, don't paste it.
What you can do instead
AI is genuinely useful in a practice, and most uses need no patient data at all. For the rest, you need a properly governed professional tool: chosen by the practice, covered by a contract and set up to protect the data.
| Need | Without patient data | With patient data |
|---|---|---|
| Draft a template letter or text message | Yes, in a consumer AI tool, then complete it by hand | No, except in a properly governed professional tool |
| Summarise a guideline or a scientific article | Yes, the text is public | Not applicable |
| Prepare a poster, an FAQ, the practice's home page | Yes | Not applicable |
| Summarise a specialist's letter or a patient file | No | Only in a properly governed professional tool |
| Dictate a consultation report | No | Specialist tool with a contract, and the patient informed |
Questions to ask before choosing a professional tool
- Is there a data processing agreement that complies with the GDPR, signed in the practice's name?
- Is the data used to train the tool? The answer must be no, in writing.
- Where is it stored, and for how long? Preferably in the European Union, with a short retention period and the option to delete.
- Who manages the accounts? The practice, with one account per person and multi-factor authentication (a code on your phone as well as the password).
- Can you see who did what? An access log lets you check if in doubt.
- How is the patient informed? The GDPR requires you to tell people clearly what you do with their data.
A one-page rule and fifteen minutes of explanation
Banning AI outright doesn't work: people use it anyway, on the quiet. A clear rule and a solution work better. Since 2 February 2025, the European AI Act also requires staff to have sufficient command of these tools; the European Commission specifies that an internal record of the training is enough, with no certificate needed.
Here is what you can do first thing on Monday:
- Ask the team who uses what, without blame: the point is to know, not to punish.
- Write the rule on one page: what is allowed (texts without patient data), what is forbidden (patient names, files, results or images in a tool that hasn't been approved), and who to contact if in doubt.
- Provide a solution: a properly governed professional tool if the need is there, otherwise the list of permitted uses.
- Switch off history and training in the accounts used for the practice, where the tool allows it.
- Explain the rule in fifteen minutes at a team meeting, and note the date: that is your training record.
- Always check what the AI produces before sending it or filing it in a patient record.
What if it has already happened?
Patient data has been pasted into a tool that hasn't been approved? Don't panic, but don't stay silent either: under the GDPR, this may be a data breach.
- Write down the facts: what, when, by whom, in which tool, how many patients.
- Delete the conversation in the tool and check the history settings.
- Assess the risk with your data protection officer: could the patient be identified from the text?
- Notify the DPA if necessary: where there is a risk to the people concerned, notify without undue delay and, where possible, within 72 hours, through the DPA portal (page in French); notifications sent by email are not processed.
- Fix the cause: the written rule and the right tool stop it happening again.
AI can save a practice valuable time. The one thing to remember: your patients' data stays in tools the practice has chosen, covered by a contract. To review how your team uses AI, choose a professional tool or write your one-page rule, DeepIT supports healthcare practices in Brussels and across Belgium.
Set AI ground rules for your practice→
Frequently asked questions
Can I use an online AI tool if I remove the patient's name?
Only if nothing else can identify them: no date of birth, address, file number, rare detail or name in an image. For anything about a specific patient, use a professional tool covered by a contract.
Does the paid version of an AI assistant solve the problem?
Not automatically. What matters is the contract signed in the practice's name, the written commitment that the data is not used to train the tool, where it is stored, and the practice managing the accounts. An individual subscription paid for by an employee does not provide those guarantees.
Are receptionists and assistants bound by confidentiality?
Yes. The Code of Medical Ethics requires the doctor to ensure that their staff respect professional secrecy. So the rule on using AI must apply to the whole team, reception included.
Do I need to tell the patient if I use an AI tool to dictate my reports?
Yes, the GDPR requires you to inform people clearly about how their data is used: a mention in your patient information and a sentence during the consultation are often enough. The tool must also be covered by a data processing agreement.
What does a practice risk if it has put patient data into a consumer tool?
Depending on the case: a complaint from a patient, DPA proceedings that may lead to a fine, disciplinary proceedings before the Order of Physicians and, if the disclosure was deliberate, criminal prosecution. Acting fast and documenting the facts limits the damage.