It is one of the most common questions doctors and practice owners ask today: can I use ChatGPT, and is it even legal with patient data. The topic matters, because AI genuinely saves time, yet health data is protected more strictly than ordinary personal data. The good news is that you can use AI sensibly and in line with GDPR, as long as you know a few rules.
This article explains what you can and can't do, how free ChatGPT differs from a solution deployed in the practice, and how to use AI safely. It complements the broader piece on AI use cases for doctors and is part of AI and systems for healthcare. This is not legal advice, just a practical explanation of the rules.
In the AI projects we run for medical practices, we start from one question before anyone touches a tool: where does the patient data actually sit, and who can see it. In practice that single check tells you more than any feature list, because most compliance problems are not about the model itself but about pasting identifying data into a place that has no data processing agreement and no control over where the data goes. The rest of this guide follows the same order we use with clients: understand the data, then choose the tool.
The short answer
The tool itself is neither compliant nor non-compliant with GDPR. What decides is how you use it and in which variant. For tasks without patient data, ChatGPT is helpful and safe. For work on data that identifies a patient, the free version is not suitable, but a solution deployed in the practice, with a data processing agreement and control over the data, is.
The question is not "is ChatGPT GDPR-compliant" but "for what and in which version am I using it". The same tool can be fine in one scenario and break the rules in another.
Why medical data is a special risk
Health data is special-category data under Article 9 GDPR, meaning it falls under a stricter regime than ordinary personal data. By entering it into a public tool you lose control over where it is processed, whether it is used to train models and who has access to it. Without a data processing agreement you fail a basic requirement, and responsibility for a breach rests with the practice, not the tool's provider.
Anonymization vs pseudonymization
This distinction is key, because it decides whether the data still falls under GDPR.
- Anonymization is the irreversible removal of identifying data, so information can no longer be linked to a specific person. Fully anonymized data is not personal data and is outside GDPR.
- Pseudonymization replaces identifying data with a code or pseudonym while keeping the ability to re-identify via a separate key. Pseudonymized data is still personal data and stays under GDPR, but the risk is lower.
In practice, fully anonymizing medical data is hard, because even a set of seemingly harmless details can lead to identification. That is why with AI we usually talk about pseudonymization plus safeguards, not the assumption that the data is no longer protected.
Who is responsible for the data
Under GDPR the practice is usually the controller of patient data, and the AI provider that processes it on the practice's behalf is the processor. For such processing to be lawful you need a data processing agreement (DPA) defining scope, purpose and safeguards. The key consequence is that the controller, the practice, is accountable to the patient and the regulator even if the breach happened on the provider's side. Free ChatGPT used privately gives no such agreement, so the whole burden stays with the practice.
Risk assessment (DPIA) and the AI Act
Processing special-category data at scale often requires a Data Protection Impact Assessment (DPIA), a documented risk analysis before deployment. On top of that comes the AI Act, the EU's artificial-intelligence regulation, which phases in and places extra obligations on AI used in sensitive areas, including healthcare. You do not need to be a lawyer to draw the practical conclusion: document an AI deployment up front and base it on a solution that supports compliance, rather than improvising on a public chat.
Can you use free ChatGPT with patients
You should not enter data that can identify a patient into the free, public ChatGPT: name, national ID, exact dates and details that together point to a specific person. In that version you have no data processing agreement, no guarantee of processing location and no certainty that data will not be used for training. That does not make free ChatGPT useless in healthcare; it just means it is not the place for patient data.
What you can and can't do
The simplest way is two lists.
You can (without identifying data):
- questions about general medical knowledge and explanations of concepts,
- translating texts and simplifying language,
- drafting documents, procedures and educational materials,
- organizing your own notes stripped of patient data.
You should not (in a public tool):
- paste data that identifies a patient,
- enter fragments of medical records with data,
- upload test results or images with personal data,
- base a clinical decision solely on the model's answer.
How to use AI safely
If you want AI to work on patient data too, you need a solution designed compliance-first, not a free chat. The rules are:
- Pseudonymization where possible, so AI works on data without direct identification.
- A data processing agreement with the provider and hosting in the EU.
- No training on your data, confirmed contractually.
- Role-based access control and an audit trail, that is who accessed what and when.
- A human approves every output before use; AI does not make the decision.
More on organizing data in the practice in the piece on AI in medical documentation, and on choosing the right system in the guide on how to choose practice management software.
How we check whether an AI tool is GDPR compliant
When a practice asks us whether a given tool is "safe", we do not answer from the brand name. We run it through the same short assessment every time, because compliance lives in the configuration and the contract, not the logo. These are the seven points we look at, in roughly the order that matters:
- Hosting. Where the data is processed and stored, and whether it stays in the EU.
- Data processing agreement. Whether the provider offers a DPA that names scope, purpose and safeguards.
- Training on your data. Whether your inputs can be used to train the provider's models, and whether that is contractually excluded.
- Minimisation and pseudonymisation. Whether the workflow lets you send the least data needed, ideally without direct identifiers.
- Access control. Whether there is role-based access and an audit trail of who saw what and when.
- Human oversight. Whether a person reviews and approves every output before it is used.
- Documentation. Whether the deployment is written down up front, often with a DPIA for special-category data.
A tool that clears all seven can usually be configured for compliant use around patient data. One that fails the first two, hosting and a DPA, is rarely worth the rest of the conversation. The same lens also tells you how much risk a given use carries, which is easier to read as a table:
| Use case | Identifying data | Risk level | What we usually recommend |
|---|---|---|---|
| General medical questions, explanations | none | Low | A public tool is fine |
| Translation, drafting, simplifying text | none | Low | A public tool is fine |
| Notes worked on as pseudonymised data | pseudonymised | Medium | Deployed solution with a DPA |
| Records, results or images with patient data | direct identifiers | High | Deployed solution, DPA, EU hosting, no training on your data |
This is a simplification meant to guide a first decision, not a legal classification. The thresholds and the final call for your specific setup belong with your data protection officer or a lawyer; what we bring is the process for getting the technical safeguards in the right place before that conversation.
Free ChatGPT vs a deployed solution
| Criterion | Free ChatGPT | Solution deployed in the practice |
|---|---|---|
| Data processing agreement | none | yes |
| Data hosting | out of your control | EU |
| Training on your data | possible | contractually excluded |
| Access control and audit trail | none | yes |
| Patient data | not acceptable | acceptable with safeguards |
| Use | general knowledge, drafts | documentation, practice processes |
Examples: do and don't
- Do: a doctor asks ChatGPT to simplify a patient leaflet or translate recommendations into another language, with no personal data.
- Do: drafting an internal procedure or a letter to a payer based on general points.
- Don't: pasting a visit note with the patient's name and national ID into the free chat to generate a record.
- Do this instead: the same record created by a solution deployed in the practice, with a data processing agreement and data control.
Compliance checklist
- ☐ I do not enter identifying patient data into public tools.
- ☐ For work on data I use a solution with a data processing agreement and EU hosting.
- ☐ I have confirmation that data is not used to train the provider's models.
- ☐ I apply pseudonymization where possible.
- ☐ I have role-based access control and an audit trail.
- ☐ A human approves every AI output before use.
- ☐ The team knows these rules and what must not be pasted.
If you want to deploy AI in your practice in a GDPR-compliant way, we help choose the solution and set up the process from the security side. See AI and systems for healthcare or book a call.
FAQ
Can a doctor use ChatGPT?
Yes, for tasks without patient data: general knowledge, translation, drafting documents, organizing notes without identifying data. The problem starts when data that can identify a patient is entered into a public tool. Then you need a solution with a data processing agreement.
Are you allowed to enter patient data into ChatGPT?
You should not enter identifying patient data into the free, public ChatGPT. Health data is special-category data under Article 9 GDPR, and the free version has no data processing agreement or control over processing. Pseudonymized data or a deployed solution with safeguards is acceptable.
Is ChatGPT GDPR-compliant?
The tool itself is neither compliant nor non-compliant; it depends on use and variant. The free version is not suitable for patient data. Deployed solutions with a data processing agreement, EU hosting and no training on your data can be configured to be GDPR-compliant.
How do you use AI safely with medical data?
Do not paste identifying data into public tools, use pseudonymization, choose solutions with a data processing agreement and EU hosting with no training on your data, apply access control and an audit trail, and always have a human approve the AI output.
How does free ChatGPT differ from a solution deployed in the practice?
Free ChatGPT is a public service without a data processing agreement and without control over location or use of data. A deployed solution runs with contractual guarantees: EU hosting, no training on your data, access control, an audit trail and integration with your systems.
What are the consequences of a GDPR breach with medical data?
Health data is protected more strictly, and breaches can lead to high financial penalties as well as loss of patient trust. That is why using AI with patient data calls for a deliberate decision and safeguards, not pasting data into the first available chat.
Can I use Gemini, Claude or other models instead of ChatGPT?
The rule is the same regardless of provider. In a public, free version do not enter identifying patient data. To work on data you need a business or deployed variant with a data processing agreement, EU hosting and a guarantee of no training on your data. What matters is the configuration and the contract, not the model's name.
Is patient consent required to use AI?
It depends on the legal basis and purpose. Using AI as a support tool within record-keeping does not always require separate consent, but the patient should be properly informed, and in some cases consent or another Article 9 GDPR ground is necessary. Confirm the scope with your data protection officer.
Is anonymization enough to use AI safely?
If data is fully and irreversibly anonymized, it stops being personal data and GDPR no longer applies. The catch is that fully anonymizing medical data is hard, because a set of details can lead to identification. So in practice you use pseudonymization plus safeguards, rather than assuming the data is already outside GDPR.
How do you check whether an AI tool is GDPR-compliant?
In practice we run a tool through the same short assessment: where data is hosted and whether it stays in the EU, whether there is a data processing agreement, whether the provider trains on your data, whether you can minimise and pseudonymise the input, whether there is role-based access control and an audit trail, whether a human approves every output, and whether the deployment is documented (often with a DPIA). A tool that passes all seven can usually be configured for compliant use, but the final call on your setup belongs with your data protection officer.
Sources
- European Commission: Artificial intelligence in healthcare — EU framework and guidance on the use of artificial intelligence in healthcare.
