Training AI Models Under KVKK: Legal Requirements for Using Datasets

⏱️ Reading time: 13 min · 📁 Category: KVKK & Artificial Intelligence

Training AI Models Under KVKK: Legal Requirements for Using Datasets

The training of AI models is not just a technical process;
It is directly linked to the processing, transfer, conversion and reproduction of personal data.

Since 2025, KVKK and global data protection authorities have examined the legal bases, anonymity levels and output risks of model-training datasets more closely.

Organisations developing AI, using external AI models or providing datasets must recognise KVKK risks specific to training and establish appropriate legal mechanisms.

DPA decisions in 2025 penalised numerous companies for KVKK/GDPR violations because training data was not anonymous. 1. What is Personal Data in the Model Training Process? Where Does It Begin?

Every set of data used in AI models is not automatically anonymous. In terms of KVKK, "personal data" can occur at three levels in model education:

1) Direct Identifier Data

• Name, surname
• Phone number
• TC ID no
• IP address

2) Indirect Identifier Data

• Location data
• Cookie ID
• Behaviour logs
• Social media activities

3) Reversible Data from the Model (Model Inversion Risk Zone)

• Inferring patterns relating to individuals represented in the model’s training data
• Membership inference: understanding whether a person is included in the training set

Anonymous data ≠ masked data. True anonymity requires irreversibility. 2. Anonymous Data - Personal Data Separation: Critical Points for Model Education

Although data sets used for AI models are often referred to as "anonymized", anonymity is only possible under the following conditions, according to the KVKK:

• Data must no longer be associated with anyone
• It must be resistant to reversing techniques
• The data owner should not be reached from model outputs
• Re-identification must remain unlikely despite technological advances.

If these conditions are not met, the data is still considered "personal data" and subject to the provisions of the KVKK.

Model inversion and membership inference mean many datasets cannot be considered anonymous in practice. 3. Legal Basis Required for AI Model Training According to KVKK

The training of AI models must be done with one of the legal reasons under KVKK m.5 and m.6.

A) explicit consent (not safe for all situations)

• Users need to explicitly approve model training
• Withdrawal of consent affects the model → retraining is required
• Sustainability in practice is difficult

B) Necessity for Contract Performance

• May apply where model training is necessary to provide the service
• Not every model enters this coverage

C) legitimate interest (most commonly applied but most risky basis)

• Balance testing (LIA - Legitimate Interest Assessment) is mandatory
• Must not overly limit user rights

D) Data Manifestly Made Public

• Social media data does not enter this coverage -> still in KVKK protection

E) Processing Special Categories of Data

• Can not be processed without explicit consent
• The use of biometric or health data in model training is very high risk

Lack of LIA and DPIA reports was a major issue raised in KVKK audits in 2025. 4. Model Training Risks: Assessment in terms of KVKK

The main risks that arise when training AI models are:

1) Model Inversion Attack

• Inferring personal data used in training from the model

2) Membership Inference Attack

• Understanding whether a particular person is included in the training set

3) Overlearning (Data “Etched” into the Model)

• The model's data cannot be "forgotten" -> KVKK m.4 violation (linking to their intended purpose)

4) Cross-Border Data Transfer

• Processing of data abroad for model education
• EU/US servers → no adequacy decision

5) Provision of the Data Set from Third Parties

• Unclear dataset origins → liability for unlawfully processed data

Every data used in AI training is under the responsibility of the data controller. "The data set has arrived" is no excuse. 5. What Should an AI Training Process Be Like for KVKK?

The technical and legal processes for KVKK compliance must be conducted in an integrated form:

A) Data Inventory and Source Verification

• It must be documented where the data set comes from
• Contract + technical control for third-party data providers

B) DPIA - Data Protection Impact Analysis

• Risks before model training should be analyzed
• If there is a high risk, consent may become mandatory

C) LIA – Legitimate Interest Assessment

• Test the model’s necessity and proportionality

D) Real Anonymity and Synthetic Data Use

• Synthetic data is recommended in high-risk sectors
• Anonymity must be irreversible

E) Audit Trail (Control Trail) Installation

• It should be recorded when the model is trained with what data

6. Special Risks for Companies in Turkey

Many models developed in Turkey use EU/US cloud infrastructure, creating serious risks for **cross-border data transfers**.

• Data flow must be checked on AWS / GCP / Azure models
• AI data processing frequently came up in the KVKK Board's 2025 decisions
• Cross-integration SaaS solutions can create secret data transfer

In 2025, the KVKK Board imposed sanctions on companies that "can't explain the source of the data set used in AI training." Conclusion: AI Training Is High-Risk Data Processing under KVKK

The nature of the data used in model education, its legal basis, storage time and the possibility of reversibility from output; it forms the heart of KVKK compliance.

Therefore, AI training processes from the KVKK perspective should be based on the following principles:

• Minimum data -> maximum anonymity
• Open and accountable process
• Auditable model training records
• Strong legal basis
• Continuous risk analysis

The key issue in KVKK compliance after 2025: “Model transparency + data source management.”

KVKK and AI Training Consultancy