Júlia Škovierová: AI can help teach artistic skills, but it should not judge creativity
Can artificial intelligence help teach drawing, painting, or sculpture? Researcher Júlia Škovierová from the Robotic Perception Group at the Czech Institute of Informatics, Robotics and Cybernetics at the Czech Technical University in Prague (CIIRC CTU) and her team are exploring this question through the EDUARDIS project, an intelligent assistant that provides students with continuous feedback during the creative process. It does not assess talent or originality. Instead, it focuses on the technical aspects of the work, such as composition, perspective, proportions, and the correct workflow, helping students master the fundamentals of artistic practice.
The project builds on the many years of teaching experience of Assoc. Prof. Milan Salák, who developed a set of 48 educational tasks for teaching drawing, painting, and sculpture. In 2024, he approached the CIIRC team with the idea of using artificial intelligence to support education. This led to the EDUARDIS project supported by the Technology Agency of the Czech Republic (TA CR), which was launched in September 2025. Since then, researchers have been developing an assessment methodology, preparing a unique dataset, and building a system based on computer vision, machine learning methods, and small language models. EDUARDIS is scheduled for its first pilot deployment in teaching during the winter semester of the 2026/2027 academic year, primarily at the Faculty of Education of the University of Hradec Králové (UHK). In the future, the system is expected to expand to other institutions, including those outside academia. It could also be used in short-term art courses for different age groups and in other forms of art education.
An interdisciplinary team of experts in computer vision, artificial intelligence, image data processing, and web application development is working on the project. Students are also involved in the research and development of the system, gaining their first experience with scientific work. Professor Václav Hlaváč serves as an expert advisor to the project, Ihor Shabash contributes to the development of the web application and machine learning methods, and Tomáš Staudek is responsible for the web application design and the project’s visual identity.
How does artificial intelligence learn to identify mistakes in students’ work? What can it assess objectively, and where does the role of the teacher remain irreplaceable? And how difficult is it to translate the experience of art educators into the language of algorithms? We discussed these questions with Júlia Škovierová.
1)You describe EDUARDIS as an AI-powered teacher of artistic skills. Could you introduce the project? What AI technologies and methods are behind the concept?
The goal of the EDUARDIS project is to create an intelligent AI-powered teaching assistant that provides students with real-time feedback based on photographs of the individual stages of their work on art assignments. After completing a particular stage, the student takes a photograph of their work, uploads it to the system, and receives both textual and visual feedback.
To analyse image data and assess whether individual steps have been carried out correctly, we use a combination of “classical” computer vision methods, neural networks, and small, locally running language models (SLMs). This hybrid approach allows us to combine the strengths of the different methods: precise image analysis, recognition of visual structures, and the generation of clear and understandable feedback for students.
2) Your research focuses on computer vision. How does the system “look” at a student’s work? What characteristics of a drawing or artistic work can it analyse?
The system assesses only objectively measurable technical aspects of an artwork. These include, for example, the correct arrangement of the composition, the use of perspective, the proportions of individual objects, the use of the image format, the presence of details at the appropriate stage of the creative process, and the correct use of colour relationships according to the assignment.
For each educational task and each of its individual stages, a set of precisely defined assessment criteria is established based on the teaching methodology. Using these criteria, the system evaluates whether the student is following the correct process and provides continuous textual and visual feedback.
The system does not assess creativity or the aesthetic value of the finished work. It focuses exclusively on the technical correctness of the execution and adherence to the prescribed methodology.
3) What types of data do you need to train the system, and how difficult is it to obtain a high-quality dataset for art education?
To train the system, we need highly specific data that are not available in commonly accessible databases. Photographs of finished artworks alone are not enough. We need to capture the entire process of creating them, including the individual stages accompanied by expert annotations identifying both correct procedures and typical mistakes.
We are therefore creating our own database containing photographs of the individual stages of the creative process, together with both positive and negative annotations. For each task, objectively assessable criteria are defined and used to train and validate our algorithms.
Collecting these data is very demanding in terms of both time and organisation. Teachers need to document individual stages of students’ work during lessons and subsequently provide expert annotations. Although a large number of student works are created every year, the amount of usable data is considerably smaller because not every work includes complete documentation of the entire process.
We expect a significant improvement with the pilot deployment of EDUARDIS, during which students themselves will upload photographs of the individual stages. This will reduce the administrative burden on teachers while allowing us to collect a substantially larger amount of complete data.
4) How does assessing a technically correct solution differ from evaluating a creative outcome? How do you address this challenge within the EDUARDIS project?
From the perspective of computer vision and artificial intelligence, assessing a technically correct solution is considerably easier than evaluating a creative outcome. Technical aspects can be described using clearly defined rules and assessment criteria, such as the correct use of perspective and composition or adherence to the individual stages of the workflow.
By contrast, the creative process cannot be constrained by unambiguous rules. Every student works with a different artistic style, colour palette, or means of expression, and this diversity is desirable in art education. After all, evaluating a work from a creative perspective is a complex task even for humans and is more a matter of subjective opinion than objective assessment.
The accompanying images show that students can create very different works while working from the same assignment and following the same technical rules. This clearly illustrates that the system assesses the correctness of the working process rather than an individual artistic style.
5) What role do art teachers themselves play in the project? Are they involved in developing the methodology or creating the training data?
Art teachers are among the key partners in the EDUARDIS project. They contribute to developing the methodology, defining assessment criteria, collecting data, and providing expert annotations. Their many years of teaching experience form the basis of the rules according to which the system assesses the individual stages of the creative process.
The project is based on a set of 48 educational tasks developed from the many years of teaching experience of Assoc. Prof. Milan Salák. We are currently focusing on six key tasks in drawing, painting, and sculpture, which serve as a validation set for the development of the system.
For data preparation, we work primarily with teachers from the University of Hradec Králové (UHK). During lessons, they document the individual stages of students’ work and use an annotation tool developed by our team to identify correct procedures as well as typical mistakes. This process is creating a unique dataset for training and validating the algorithms.
6) How does the system generate feedback for students? Does it only assess the result, or can it also monitor and comment on the creative process itself?
The current version of the system provides students with feedback in two forms: textual and visual. The textual component describes the identified issues and recommends how to correct them, while the visual component directly highlights problematic areas in the uploaded photograph. In the future, we plan to extend the system so that it can track the continuity between individual stages of the work. It will therefore assess not only the correctness of individual steps but also the entire creative process, the continuity of the workflow, and the student’s gradual improvement.
7) Are there aspects of art education that are particularly difficult for current AI models to grasp?
One of the greatest challenges is translating the pedagogical principles of art education into a formal representation that can be processed by computer vision algorithms and artificial intelligence models. Many of these principles are based on teachers’ many years of experience and are expressed intuitively. Translating them into clearly defined criteria therefore requires close cooperation between experts in art education and computer science.
8) How do you verify during development that the recommendations provided by the system are genuinely useful and help students improve their skills?
The recommendations provided by the system are based on teachers’ many years of educational experience and on the methodology developed for the individual learning tasks. Throughout the development process, they are continuously discussed with educators, who verify their accuracy and suitability for the different stages of the creative process.
A typical example is landscape painting, where students are advised not to focus on adding details during the initial stages. They should first establish the composition, perspective, and basic colour relationships, while detailed work should come only in the final stages. The system can therefore identify details that have been added too early and alert the student to them.
The actual benefits of the system will be evaluated during its pilot deployment in teaching. We will examine whether continuous feedback helps students learn correct working procedures more quickly, reduces the number of repeated mistakes, and improves the quality of their work.
9) Could the technology developed within the EDUARDIS project also be used in other fields, such as teaching design, architecture, or technical drawing?
Yes. The entire EDUARDIS system is designed to be modular and extensible, allowing it to be adapted to different types of educational tasks and other fields of study.
We see the greatest potential in technically oriented disciplines where workflows and assessment criteria are clearly defined. Compared with artistic creation, it is easier in these areas to formalise the correct procedure and automatically evaluate the individual stages of a task.
With appropriate training data and expert annotations, the same methodology could be applied, for example, to teaching technical drawing, architecture, design, or other disciplines in which both the correctness of the final solution and adherence to the proper working process are important.
10) What are the biggest research questions you currently need to address for EDUARDIS to become a fully functional and widely applicable educational tool?
One of the greatest research challenges is translating the specialist language of art education into a formal representation that can be used by artificial intelligence algorithms. We therefore analyse individual tasks in detail together with educators and look for ways to translate their experience into clear and unambiguous assessment rules.
In some cases, compromises must be found between pedagogical requirements and the technical capabilities of the system. To make certain aspects of the creative process automatically assessable, it is sometimes necessary to define the assessment criteria more precisely or adjust the way an assignment is formulated without changing the fundamental nature of the teaching process.
The research also involves continuously testing different computer vision methods, neural networks, and language models. For each task, we look for the combination of algorithms that provides the best balance between accuracy, robustness, and processing speed.
The EDUARDIS project builds on the many years of teaching experience of Assoc. Prof. Milan Salák, who developed a set of 48 educational tasks for teaching drawing, painting, and sculpture. In 2024, he approached us with the idea of using artificial intelligence to support distance learning. This led to the TA CR project TQ23000335, which was launched in September 2025.
