Not all data deserves the same protection.
Data protection law grades personal data by risk: the greater the potential harm from misuse, the higher the legal requirements — and the more important masking becomes before data reaches an LLM.
The highest tier: data whose disclosure can stigmatise or discriminate against a person. The Swiss nDSG (Art. 5 lit. c) and the GDPR (Art. 9) name these categories explicitly; processing them usually requires explicit consent, and passing them to third parties — including LLM providers — is particularly delicate. Such content should never reach an LLM unmasked.
- Health data (diagnoses, findings, medication)
- Religious, philosophical and political views
- Genetic and biometric data
- Data on criminal proceedings and sanctions
- Data on social assistance measures
- Ethnic origin, intimate life
Not a “special category” in legal terms, but high-risk in practice: official identifiers such as the AHV number are strictly regulated by law and allow records to be linked unambiguously. Financial data such as IBANs and account activity are a direct target for fraud. When leaked, the damage is immediate — which is why the masking treats them like highly sensitive data.
- AHV number (official personal identifier)
- IBAN and account numbers
- Credit and debit card data
- Salary and tax data
- Insurance and contract numbers
Name, contact details or job title count as “ordinary” personal data — protected, but with lower legal hurdles. Their risk lies in linkability: a name alone says little; combined with an employer, a phone number and the content of a conversation, a profile emerges. That is why the demo masks this data by default, too.
- First and last name
- Email address
- Phone number
- Postal and home address
- Role and employer
What does this mean for masking?
As a rule of thumb: the higher the protection level, the more conservative the masking should be — a lower threshold, and redaction rather than pseudonyms when in doubt. The integrated TekAI package covers all three tiers: specially protected statements (e.g. about health, religion or proceedings) are masked as a whole clause, identifiers such as AHV numbers, IBANs, credit cards and customer numbers as well as names, contact details, addresses and dates of birth. Every finding carries a confidence score, visible in the entity table.