Privacy, KVKK and Protection of Client Data

Gains:

• Ability to classify client data within the framework of professional confidentiality and KVKK and determine which data can be given to artificial intelligence tools
• Ability to create secure context in cloud-based tools by anonymizing session content and identity information
• Ability to adapt information, explicit consent and data storage/deletion responsibilities to the use of artificial intelligence

There is a saying at the core of the mental health profession: "What you tell me stays there." A client trusts you to tell you the most fragile information of his life (traumas, addictions, relationships, fears). This trust is the heart of the profession and once damaged it cannot be restored. Artificial intelligence (AI) tools add a whole new risk to this trust: because a sentence you type into the cloud can go out of your control and be stored on another company's server, or even be used in training the model. In this unit, you will learn how to protect client data in the AI ​​age; You will learn what is allowed and what is prohibited within the framework of KVKK (Personal Data Protection Law; Law No. 6698, which regulates the processing of personal data in Türkiye).

Why is client data "special quality"?

KVKK divides data into two: ordinary personal data (name, telephone) and special personal data. Any information regarding mental health (even the fact that a person is receiving therapy) falls into this second, most protected group. Health data is a separately protected category in the law; As a rule, processing requires explicit consent or strict exceptions. In other words, typing the content of a client's session into a random AI tool could be not only an ethical but also a legal violation.

A professional secrecy obligation is also added to this. Professional ethics prohibit sharing client information with third parties without permission. A cloud-based AI tool is legally a “third party.” So the question is: Am I really authorized to give this data to this tool?

Caution: "No one will see it" or "I'll just summarize" is not protection. The moment data enters the vehicle, it is out of your control. Maintaining confidentiality is a decision made BEFORE sending data; cannot be taken back later.

Three states of data: which data, to which medium

When working with client data, divide each information into three categories:

Data type

example

Can it be given to AI?

rule

direct identification

Name-surname, TR ID, telephone, address, workplace

no

Never given in its raw form

indirect identification

Rare profession, small settlement, unique event

No (opens ID alone)

Generalized or removed

Clinical content (anonymous)

Symptoms, themes, process (identity extracted)

Conditionally yes

Anonymized and in a secure vehicle

The goal is de-identification: making the data such that neither you nor anyone else can link it to a specific person. Just deleting the name is not enough; The sentence "the only dentist in town" gives the person away even if the name is deleted.

Step by step: secure anonymization

1. Extract direct identifiers. Name, surname, range instead of age ("30s"), location, workplace, school, date, relatives' names.
2. Generalize indirect identifiers. "Hazelnut exporter in Trabzon" rather than "working in the agricultural sector in a city". Abstract unique events.
3. Blur the numbers. "about 6 months ago" instead of the exact date.
4. Inspect the vehicle. Does the AI ​​tool you use use data in training? Is there an "enterprise/privacy mode"? Where is the data stored?
5. Minimum data policy. Provide the minimum information needed to perform the task; don't give too much.
6. Consent and transparency. Educate the client that you are using AI for administrative/drafting purposes in the process.

Tip: After anonymizing a text, ask yourself this question: "If this client read this text, would he recognize himself? Would anyone else recognize him?" If you cannot say "no" to both, anonymization is not complete.

three mini cases

Case 1 — Small mistake, major violation. While a specialist has the AI ​​summarize the session notes, the client's name is "M.Y." and leaves the rest as follows: "Nurse at Hospital X in Kadıköy, daughter's name is Elif, divorce case..." Even though the name is shortened, these details easily identify the person. Anonymization failed; violation is deemed to have occurred. The right way: generalize the role as "worker in the healthcare sector", the child as "has a child", the case as "a legal process".

Case 2 — Vehicle selection makes a difference. Two psychologists do the same job. One types the raw data into a personal, free chat tool; The tool's contract says "inputs may be used to improve the service." The other uses an enterprise tool with a data processing agreement, the setting of which is open "using my input in education", and with anonymous data. They both say "we used AI", but one protected the data and the other leaked it.

Case 3 — Right to erasure. A client leaves therapy and requests the deletion of his data (his right under KVKK). The expert deletes his own recordings, but remembers the session transcript he pasted into an AI tool 3 months ago; he cannot recall the data or ensure that he has deleted it. This shows why working anonymously is essential from the start: no identification data is placed where it cannot be deleted.

Copiable prompts and templates

Anonymize the text below. REMOVE or generalize the following: first name, last name, age (convert to range), city/county/neighborhood, workplace, school, date, relative names, unique events. Then I was asked "can identity be inferred from the remaining text?" Also write your evaluation.Text: [text]

List each phrase in this text that could directly OR indirectly describe the person (including rare occupation, small place, unique event). Suggest how I can generalize for each. Text: [text]

Draft a simple, understandable paragraph to be added to the client information text for a psychology practice, containing the information that "artificial intelligence tools can be used for administrative and draft purposes during the service process; client data will be anonymized and diagnosis/treatment decisions belong to the specialist."

It provides a list of control questions that I should ask in terms of the privacy policy of an artificial intelligence tool that I am considering using: data storage, use in education, data location, deletion, data processing agreement, corporate mode.

Weak prompt / Strong prompt

Weak prompt: "Summarize the following session recording: [raw text with real name and full details]"

This prompt transfers sensitive data in its raw form to the third party; It is a violation of KVKK and professional secrecy.

Powerful prompt: "First anonymize the following text (generalize name, place, workplace, date, relatives' names), then summarize the anonymous version in SOAP format. Also evaluate whether anonymization is sufficient: [text]"

This prompt first protects, then produces; He finishes the job before his identity is revealed.

Common mistakes

• Just delete the name. Indirect descriptors (rare occupation, small location) give the person away even if you delete the name.
• Writing raw data to free/personal tool. Providing data without checking whether the inputs are used in education or not.
• Bypassing consent. Not informing the client at all about the use of AI; Transparency is an ethical obligation.
• Thinking it's "temporary". Forgetting that data entering the cloud cannot be retrieved; To be helpless when the right to delete arises.
• Violating the least data principle. Pasting dozens of unnecessary details for the task; Every extra data is an extra risk.

In summary

Client data is in the special quality, most protected group under KVKK and is within the scope of professional secret. A cloud-based AI tool is legally a third party; raw client data cannot be given to him. Separate data into direct/indirect identification and anonymous clinical content; Use only truly anonymized content in a privacy-controlled tool with a minimal data policy. Educate the client about the process. Confidentiality is a decision made before sending data that cannot be reversed later.

Application task

Take a hand (or imaginary) session note. First mark all direct and indirect identifiers in yellow. Then replace each with a generalized expression and write an anonymous version. Finally ask yourself, "Does this client know himself?" Ask the question and note your answer. You can also open the privacy settings of an AI tool you use and ask “are the inputs used in training?” Find the answer to the question.

checklist

• [ ] I stripped out direct identifiers (name, location, date, workplace).
• [ ] I generalized indirect descriptors (rare occupation, unique event).
• [ ] I have verified that the tool I am using does not use data for training.
• [ ] I followed the principle of least data; I just gave what was necessary.
• [ ] I informed the client about the use of AI / observed the consent framework.
• [ ] "Does the client recognize himself?" I passed the test.