Artificial intelligence processes data in dimensions that were barely visible at the time the GDPR was created. As a recent Dispute with the Baden-Württemberg Ministry of the Interior shows, AI models use huge data sets and can learn from personal information - and the legal situation often remains unclear. Six years after the European General Data Protection Regulation came into force, companies are still struggling with how to implement AI projects in a legally compliant manner.
The German Data Protection Conference has now presented a guideline to provide clarity. But is that enough? While start-ups fail due to complicated legal issues and corporations invest millions in compliance departments, technology is developing rapidly. ChatGPT, image generators and autonomous systems pose new questions every day - for which there are often no clear answers.
The most Important in a Nutshell
- As the conference of independent federal and state data protection authorities, the DSK calls for comprehensive consideration of data protection principles in the development and use of AI in its guidance.
- Data protection by design - data protection requirements in accordance with the GDPR must therefore be integrated into planning and development right from the start.
- The minimization of personal data is central: Only data that is absolutely necessary for the purpose may be used.
- Transparency - Sensitive decisions made by AI systems must be traceable and documented.
- AI systems must be protected against data protection risks such as memorization or unauthorized disclosure of data.
- Privacy reserving techniques (e.g. differential privacy) and mechanisms for intervenability are strongly recommended.
- Data sources must be carefully checked and documented, particularly with regard to their trustworthiness and representativeness.
- Pilot studies and continuous evaluation of data quality and fairness help to avoid possible discrimination and distortions.
Challenges and risks in AI development
Published in June 2025 Guidance from the DSK addresses some important data protection risks associated with the use of AI systems, especially in the case of improper or negligent development. In many cases, the processing of large amounts of data is a prerequisite - often including personal data. However, some companies run the risk of over-interpreting the importance of scope for the success of the project. This can conflict with the principle of data minimization prescribed in the GDPR. One example of this is the creation of massive training databases without a clear analysis of whether all data points are necessary for the planned modeling. This not only exposes companies to an increased risk of regulatory penalties, but can also lead to inefficient AI modeling. Technical risks such as data memorization, i.e. the inadvertent inclusion of sensitive data in decision-making processes, are also critical.
This is because AI models, especially deep neural networks, not only store abstracted parameters, but in extreme cases can also reproduce fragments of the training entries. This is particularly problematic if insufficiently anonymized personal data is used. Unintentional disclosure of this information could not only trigger data protection crises, but also reputational crises. Furthermore, the ethical aspect of bias avoidance is often insufficiently considered. AI models systematically adopt biases from the underlying data. This means that there is a risk of technical reproduction of discrimination if there is insufficient control. The DSK's guidance emphasizes the need to use fair and representative data sources, carry out regular bias analyses and enable continuous improvements through feedback loops.
Seven guarantee targets for high data security
To minimize data protection risks, the DSK defines a comprehensive set of assurance objectives as integral guidelines for all development phases of AI systems: Design, development, implementation, operation. The objectives are based on the basic principles of the General Data Protection Regulation (GDPR) and build an essential bridge between the partly abstract legal standardization and its practical technical implementation.
Data minimization
Probably the most important principle of the GDPR requires that AI systems only process the data necessary for their specific purpose. Specialized mechanisms should already be used in the design phase to control and reduce data collection and processing. For example, AI models can preferably be trained with synthetic, pseudonymized or anonymized data in order to minimize the processing of personal information. In addition, it makes sense to apply validation and filtering guidelines that exclude superfluous or potentially problematic data categories. Technologies such as Federated Learning make it possible to create a global AI model without personal data leaving its place of origin. This not only contributes to data minimization, but also to the protection of sensitive information.
Availability
Robust availability management ensures that AI systems and the underlying data are always ready for use, especially for business-critical applications. This can be ensured by using scalable data management systems such as big data databases. Such systems offer the possibility of storing data without redundancy and keeping it ready for retrieval at all times. Predictive maintenance approaches can also be used to detect potential failures of important system components at an early stage. Those responsible should ensure a suitable infrastructure that is geo-redundant and enables real-time access to the required databases.
Integrity
Ensuring the Data integrity requires measures that guarantee that processed data is correct, consistent and protected against unauthorized manipulation. Developers can apply cryptographic hashing methods to raw data in order to reliably detect subsequent changes. In addition, comprehensive validation measures should be implemented to ensure that the data remains consistent and plausible. Statistical analyses can also be used to evaluate the quality and trustworthiness of data sources, for example by checking the data distribution and identifying anomalies at an early stage. While these measures primarily concern the integrity and quality of the data, bias checks should also be taken into account during processing in order to identify distortions in data sets and minimize ethical risks such as reproduced discrimination.
Non-linking
This objective requires that personal data is not merged into new information by linking separate data sets. Data controllers should ensure that highly correlated surrogate variables do not inadvertently enable concatenation. Safeguards such as differential privacy can be used to minimize the likelihood of traceability or re-identification to individuals. In addition, semantically isolated data architectures or alternative data coding methods should be used to ensure a clear separation of individual data strings.
Confidentiality
Appropriate protection mechanisms must also be integrated into the design of the AI system so that confidential or personal data is not unintentionally disclosed or extracted by attacks. These include privacy-preserving techniques such as Differential Privacywhich ensure that sensitive data is protected by suitable mathematical methods. Measures such as regularization techniques can also be used to improve the generalization of the AI model and avoid unwanted memorization of personal data. This reduces the likelihood of sensitive information being leaked or reconstructed by attackers.
Intervenability
To enable users to enforce their rights, it should be ensured that personal data can be managed individually. This includes the technical ability to delete data and then quickly update the AI system through retraining or selective adaptation. Technologies such as Machine Unlearning could be helpful here to retrain models without the previously deleted data. Automated feedback loops can speed up these processes without affecting the overall performance of the AI system.
Transparency
The transparency of an AI system must be ensured for sensitive decisions by Explainable models (Explainable AI) and audit-proof documentation in order to ensure the traceability of system decisions. This also applies to the parameterization of the models and the individual processing steps. Those responsible should carry out regular tests to check that the system is functioning properly and record the results. Visualization tools can help to present the decision paths of an AI model in an understandable way in order to make the system clearer and more accessible for users and technical auditors.
Guidance for each development phase
The following overview is suitable for looking at the recommendations for action from the perspective of the individual development phase:
Design phase
A data protection-oriented approach should already be pursued in the planning phase of an AI system. This includes the definition of clear objectives and the careful selection of necessary data sources, whereby anonymized or synthetic data should be preferred in order to maximize data protection. At the same time, a precise specification can prevent unnecessary or potentially conflicting data from being used.
Development phase
In the development phase, thorough data preparation and verification is required to minimize distortions or risks of discrimination. Particular focus should be placed on the quality and representativeness of the data. At the same time, systems should be implemented in such a way that they can react flexibly to changes, such as the removal of personal data without affecting the model.
Introductory phase
When introducing the AI system, its functionality must be seamlessly integrated into existing processes without jeopardizing the rights of the data subjects. To this end, the system's data usage and output mechanisms should remain verifiable so that adjustments, such as the deletion or modification of personal information, can be easily implemented.
Operating and monitoring phase
During operation, continuous monitoring of the system is necessary to avoid data protection violations and errors. Regular validation of system decisions can ensure that no undesirable results occur. In addition, feedback from use should be used to optimize the system and make it more flexible with regard to new data protection requirements
Conclusion - data protection by design
The development and use of AI systems not only requires technological excellence, but also the early and well thought-out integration of data protection requirements. The new guidance from the German Data Protection Conference (DSK) clearly shows that data protection is not an afterthought, but must be an integral part of every AI project - from the initial planning phase through to ongoing operation.
Key principles such as data minimization, transparency, intervenability and confidentiality form the basis for the responsible handling of personal data. Techniques such as federated learning, differential privacy or machine unlearning offer concrete approaches to meet these requirements. Anyone who develops or operates AI systems should use the seven guarantee objectives of the Standard Data Protection Model (SDM) as a compass - not only to comply with legal requirements, but also to ensure trust and quality.
Are you planning an AI project in your company and are concerned about data protection? Feel free to contact us to benefit from our experience in this field.
EN 
DE 
ES 
FR