ChatGPT and professional secrecy: what medical practices, law firms, fiduciaries and banks need to know

Why AI chats are a problem for secrecy

Whatever you paste into an AI chat goes to the provider and is processed on its servers, often abroad. For anyone bound by secrecy, that is a disclosure to a third party. The Swiss data protection authority (FDPIC) states that the Data Protection Act applies directly to AI-supported processing.

The key rules in Switzerland:

• Professional secrecy (Art. 321 Swiss Criminal Code): covers, among others, doctors, psychologists, nurses, lawyers, notaries and auditors bound to confidentiality under the Code of Obligations, as well as their auxiliary staff. No offence if the person concerned consents or the supervisory authority gives written permission.
• Banking secrecy (Art. 47 Banking Act): covers officers, employees and agents of a bank. Negligence is also punishable (fine of up to CHF 250,000).
• Professional confidentiality (Art. 62 FADP): fine of up to CHF 250,000 for intentionally disclosing secret personal data learned at work, including by staff and trainees.
• Sensitive personal data (Art. 5 let. c FADP): for example health data or data on criminal proceedings.
• Service providers (Art. 9 FADP): only allowed if no statutory or contractual duty of confidentiality prohibits the transfer. Transfers abroad must also meet Art. 16 FADP.

Sector by sector: typical cases

Medical practices and therapists

Risky: pasting a medical report with name, date of birth and diagnosis into ChatGPT for a summary, or having therapy notes rewritten for an insurer.

The Swiss Medical Association (FMH) has published eight recommendations on large language models. The first: prefer locally hosted or controlled models, and enter patient data only where technical and legal safeguards ensure exclusively local or controlled processing. Information that may not be published, the FMH says, does not belong in a language model.

Less risky: general medical questions, templates without patient details, properly anonymised case descriptions.

Law firms and notaries

Risky: uploading a contract with the parties' names, pasting a client email, or describing a matter in detail.

The Swiss Bar Association (SAV) guidance on AI says firms must first establish what happens to input data, who can access it and where it is stored. It names three routes: software on the firm's own network with no data leaving it, compliance with the outsourcing rules, or a consent and waiver from the informed client. Otherwise, confidential information and personal data stay out of AI systems. Firms should adopt an internal AI policy.

Fiduciaries and auditors

Risky: uploading salary statements, tax returns or ledger extracts with names, social security (AHV) numbers and IBANs to an AI chat.

Art. 321 SCC covers auditors bound to confidentiality under the Code of Obligations, and the statutory auditor must keep its findings secret (Art. 730b para. 2 CO). Fiduciary and tax work without an audit mandate is not listed separately. Art. 62 FADP, contractual confidentiality and professional rules still apply: under the EXPERTsuisse rules of conduct, members are bound by professional secrecy for everything they learn in their work.

Banks

Risky: asking an AI to draft a client letter with name, account number and asset details, or to analyse a client list. Banking secrecy also covers negligence, so check with compliance which AI services are approved and which data must never go in.

What may be possible: the conditions

1. Anonymise: if no individual can be identified, the data is no longer personal data (Art. 5 let. a FADP). Removing names is often not enough if the context reveals the case. See our guide on anonymising data for AI.
2. Consent: Art. 321 para. 2 SCC allows disclosure with the consent of the person concerned. It should be informed and documented.
3. A provider with a contract and confidentiality duties: a business plan rather than a personal account, with a reviewed contract, confidentiality, known server locations and clear rules on training. More in customer data in ChatGPT.
4. A local solution: both the FMH and the SAV mention models run within your own infrastructure.

Whether this is enough depends on your profession, contracts and supervisory authority.

A look at Germany and the EU

In Germany, Section 203 of the Criminal Code protects secrets entrusted to doctors, lawyers, notaries, auditors and tax advisers, among others. It deals with service providers explicitly: secrets may be disclosed to other persons who contribute to the professional's work, as far as this is necessary (para. 3). But anyone who fails to ensure that such persons are bound to confidentiality commits an offence (para. 4).

In its AI guidance, the German Federal Bar (BRAK) advises "abstract" prompts that reveal nothing about a specific matter and fully anonymised documents. It considers sending client secrets to language models such as ChatGPT unnecessary given the state of the art at the time.

In the EU, Art. 9 GDPR generally prohibits processing special categories such as health data, with narrow exceptions. For purposes such as medical diagnosis or treatment, Art. 9(3) requires processing by professionals subject to professional secrecy or by another person under a duty of confidentiality.

Catching slips with technology

Many slips happen when pasting in a hurry. AImori checks input locally in the browser before it reaches ChatGPT, Claude, Gemini, Copilot or Perplexity. It detects more than 240 data types such as AHV numbers, insurance numbers and IBANs, and with a local AI model also names, addresses and health information. Findings can be redacted or replaced with placeholders like [IBAN_1] (Alt+Shift+L); files such as PDF or DOCX can be uploaded as a cleaned version.

With "confidential sources", AImori stops text copied from, say, practice software or online banking. Companies can use Intune or Group Policy to block such findings instead of warning. A tool does not replace a legal assessment.

Frequently asked questions