BIO and NEN 7510: AI and privacy in government and healthcare

A municipal employee wants to tidy up a letter about a benefits claim and pastes the paragraph into a chatbot, national ID number included. In government and healthcare, what is at stake is not a customer name but a national ID number, a benefits file or a patient record. This is data of the highest sensitivity, and the rules are strict to match. The BIO for government and NEN 7510 for healthcare set those rules. Neither mentions AI, but both apply the moment a staff member puts such data into an AI tool.

#The BIO and NEN 7510 in brief

The BIO (Baseline Informatiebeveiliging Overheid) is the shared information security baseline for Dutch central government, municipalities, provinces and water authorities. NEN 7510 is the Dutch standard for information security in healthcare, supplemented by NEN 7512 and NEN 7513, which cover trust in electronic communication and logging of access to patient records. Both are based on ISO 27001 and adapt it to their sector.

The common thread: sensitive data must be classified and protected with care, and processed only by people authorised to do so, in an environment suited to it.

#Why AI adds risk here

Citizen and patient data is often special-category personal data, such as health data, with extra protection under the GDPR. The AI risk is familiar: a staff member wants to help quickly and pastes too much context into a chatbot. The Dutch Data Protection Authority (Autoriteit Persoonsgegevens, AP) described a case at a municipality where a one-month sample showed that CVs, care files and internal reports had been put into a chatbot. That is exactly the scenario the BIO and NEN 7510 aim to prevent.

The problem is rarely bad intent. It is a useful question with too much identifiable data, in a tool that was not approved for that purpose.

#The measures that count here

• Classification. Citizen and patient data belongs in the highest sensitivity class. That settles it: such data may not enter a public or personal AI tool.
• Acceptable use. The AI policy must make explicit what is not allowed: no patient, case or citizen data in unapproved tools. See enforcing an AI policy.
• Awareness. Staff must know which data never belongs in a prompt. See what you can and cannot share with AI.
• Data minimisation. The task often works just as well without the identifiable data: improving a letter does not need a real national ID number or a real diagnosis.
• DPIA. For special-category data and high risk, a DPIA is often required.

#A ban alone does not work

The reflex in sensitive sectors is to ban AI. But a ban moves the use to personal phones and accounts, where there is no visibility at all. That creates shadow AI, precisely in the environment where the data is most sensitive. The more durable route is to allow safe use and make the risk visible at the moment it arises.

That visibility is exactly what a ban gives up. Instead of guessing which chatbots and AI note-takers have quietly become part of daily work on a ward or in a case team, an administrator can see every tool that has been opened, each one scored for risk, with nothing pre-approved or pre-blocked by default.