Can Doctors Use ChatGPT? How to Use AI With Patient Data Safely
Doctors, therapists and clinics are using AI to write notes and letters. What medical confidentiality and the GDPR mean for AI tools, and how to use AI with patient data without exposing it.
8 min readClinicians spend a large share of their day on paperwork: referral letters, discharge summaries, reports for insurers, notes after each consultation. Many have started using AI chatbots to get through it faster.
Should they, and how can they do it without putting patient confidentiality at risk?
Why health data is different
Under the GDPR, health data is a special category of personal data (Article 9). Processing it is prohibited by default, except under specific conditions such as the provision of healthcare by professionals bound by secrecy.
Healthcare professionals are also bound by medical confidentiality, which in most European countries is a legal and ethical obligation, and in some a criminal one. A doctor who discloses patient information to a third party without a legal basis can face disciplinary action and, in some jurisdictions, criminal liability.
Pasting a patient's history into a consumer AI chatbot sends that information to a third-party company, on servers you don't control, under terms you didn't negotiate.
The problems with a standard chatbot
Consumer chatbots may use conversations to improve their models unless the user opts out. Prompts are processed in plaintext on the provider's servers. There's no DPA between your practice and the AI company, the data may be processed outside the EU, and you don't decide how long it is kept.
Removing the name doesn't fix this. A combination of age, rare condition, dates and town can be enough to identify a patient, and under the GDPR, pseudonymised data is still personal data.
Practical rules for using AI in a medical practice
1. Separate the tasks by sensitivity
| Task | Contains patient data? | What to use |
|---|---|---|
| Explaining a condition in plain language for a leaflet | No | Any reputable tool |
| Summarising published guidelines or studies | No | Any reputable tool |
| Drafting a referral letter from your notes | Yes | Only a tool with end-to-end encryption and no provider access |
| Structuring consultation notes | Yes | Only a tool with end-to-end encryption and no provider access |
| Preparing a report with diagnosis and history | Yes | Only a tool with end-to-end encryption and no provider access |
2. Minimise what goes in
Even with a secure tool, include only what the task needs. Use initials or a code instead of names, and leave out identifiers that don't change the output.
3. Keep the clinician in charge
AI can produce plausible but wrong medical content, so the clinician has to check every output before it goes into a record or to a patient. AI tools that support diagnosis or treatment decisions may also fall under medical device rules and the high-risk category of the EU AI Act. Use a general-purpose assistant for documentation and communication, and keep clinical decisions with the clinician.
4. Document your choice
Your Data Protection Impact Assessment (DPIA) and records of processing should cover the AI tool: what data goes in, where it is processed, how long it is kept and what safeguards apply.
5. Check national rules
Health data rules vary across Europe. Read your national data protection authority's guidance and your professional body's code of conduct before adopting a tool.
What a safe setup looks like
For patient data, nobody other than you should be able to read the content, including the AI provider. That requires end-to-end encryption, so the prompt is encrypted on your device and only decrypted where the model runs. The model should run in a hardware-secured enclave (a Trusted Execution Environment) that neither the provider nor the cloud operator can inspect. Nothing should be kept once the response is delivered, cryptographic attestation should prove the enclave is genuine, and you need a DPA and processing within the EU.
ChatLock works this way. Your chat history stays encrypted on your own device, and the AI models run inside secure enclaves where ChatLock can't see what you wrote. You save the time without handing patient data to a third party. Read more in our article on confidential computing.
FAQ
Is it legal for doctors to use ChatGPT?
Using AI is not illegal in itself. The problem is entering identifiable patient data into a tool that lacks the safeguards required by medical confidentiality and the GDPR. Consumer ChatGPT accounts generally don't meet that bar.
Is anonymising patient data before using AI enough?
True anonymisation is hard to achieve with clinical narratives. Removing the name is usually only pseudonymisation, and pseudonymised health data is still personal data under the GDPR.
Can therapists and psychologists use AI for session notes?
The same rules apply, and mental health data is among the most sensitive there is. Only use tools where the provider cannot access the content, and keep notes minimal.
Can AI help with diagnosis?
General-purpose chatbots are not medical devices and should not be used to make diagnostic or treatment decisions. They can help with documentation, communication and research, with a clinician reviewing the output.