An information system based on artificial intelligence needs data. Data is necessary for training, evaluation and use, and AI itself also produces data. The AI Act, Regulation (EU) 2024/1689 of the European Parliament and of the Council, defines several distinct concepts relating to data. But what lies behind these legal definitions?
An edited English translation of Dr Szabolcs Gyurkovics’s article published in Jogi Fórum on 24 September 2026 (original in Hungarian).
Does the content of the data or its purpose matter?
The GDPR defines personal data as any information relating to an identified or identifiable natural person. It also establishes special categories of personal data. Understanding which personal data a system processes remains essential when assessing the data protection compliance of AI.
The AI Act concepts discussed here highlight a different perspective. File type or content alone does not determine whether information is training, validation, testing or input data. The decisive question is what the data is used for in developing and operating the AI system.
The same kind of photograph might support training in one context, evaluation in another and the use of an operational system in a third. Whether it constitutes personal data must be assessed separately.
The example of a cat recognition system
The 2012 Google Brain experiment was a well-known milestone in image recognition. Researchers trained a large neural network on unlabelled frames sampled from ten million YouTube videos. Among other features, the network learned patterns sensitive to cat faces, without being told during training which images contained cats.
To illustrate the concepts, imagine an AI system that recognises cats. The following examples trace a simplified development and use process; they do not reconstruct every step of the historical experiment.
Training data: learning the patterns
Under Article 3(29) of the AI Act, training data is data used to train an AI system by fitting its learnable parameters.
Take images that show cats and images that do not. If the system learns from them and adjusts its learnable parameters in the process, the images serve as training data. This is the material from which it learns patterns relevant to its task.
The classification follows from using the images for training, rather than from the fact that they are photographs.
Validation data: evaluation during development
Article 3(30) relates validation data to evaluating the trained system and tuning its non-learnable parameters and learning process. Among other purposes, this helps prevent underfitting or overfitting.
In our example, these are the images that help developers evaluate the model and choose development settings. Training and validation may alternate several times during development.
The Regulation also defines a validation data set separately: it may be an independent data set or a portion of the training data set, using a fixed or variable split. Here too, the role of the data is central.
Testing data: independent evaluation of the finished system
Under Article 3(32), testing data provides an independent evaluation of an AI system to confirm its expected performance before it is placed on the market or put into service.
In the cat recognition example, a set of images kept separate from training and validation is used to evaluate the finished model. This supports an assessment of performance independent of the development decisions.
This is testing before deployment. It is different from users uploading images to a service that is already operational.
Input data: the basis for the system’s output
Article 3(33) defines input data as data provided to, or directly acquired by, an AI system on the basis of which the system produces an output.
When a user uploads a photograph to the finished system to find out whether it contains a cat, that photograph is input data. A text instruction, or prompt, can also be input data, but the concept is broader: it can cover images and other data received by the system.
In this example, the result is the system’s answer as to whether a cat can be recognised in the image. Article 3 of the AI Act does not contain a separate definition of “output data”.
The four categories at a glance
| Data category | What is it used for? | Cat recognition example |
|---|---|---|
| Training data | Fitting learnable parameters | Images from which the model learns patterns |
| Validation data | Evaluating the system and choosing development settings | Images that inform decisions during development |
| Testing data | Independently assessing performance before deployment | A separate image set used to evaluate the finished model |
| Input data | Producing an output | A photograph uploaded by a user for recognition |
What does this mean for data protection compliance?
The AI Act concepts discussed here describe the role data plays throughout an AI system’s lifecycle. The question is both what data we have and what we do with it.
The same kind of data can play different roles in training, evaluation and actual use. Its role as training, validation, testing or input data does not itself determine whether it is personal data or whether processing it is lawful.
When designing an AI system, it is therefore useful to examine the content of the data together with its uses. Combining these perspectives helps explain how the system works and provides a basis for assessing data protection compliance.
Dr Szabolcs Gyurkovics LL.M.
Lawyer and cybersecurity law specialist; data protection and AI compliance expert
Lawyer and cybersecurity law specialist. More than eight years of data protection experience, supporting AI and digitalisation projects.
Szabolcs and the zemITis team help review AI use at work, clarify data protection issues and establish internal policies.
Sources and further reading
- Dr Szabolcs Gyurkovics: AI Act – data categories in practice (Jogi Fórum, 24 September 2026; Hungarian)
- AI Act – Regulation (EU) 2024/1689, Article 3(29)–(33)
- GDPR – personal data and special categories of data (Articles 4 and 9)
- Le et al.: Building high-level features using large scale unsupervised learning (2012)