Raphael: “The De Brecy Tondo”
In early 2023, the art world was abuzz with the revelation that the de Brecy Tondo Madonna had been authenticated as a masterpiece by Raphael through an AI system. The announcement from Bradford University in the UK reported a 95 percent resemblance between the Tondo Madonna and Raphael’s Sistine Madonna at the Gemäldegalerie Alte Meister in Dresden. At Art Recognition, we found these revelations intriguing and decided to conduct our own analysis of the painting. Our AI model returned an 85% probability that the artwork was not created by Raphael. This unforeseen development piqued our curiosity, motivating us to delve deeper into this matter.
Operational Principles of AI Systems. The discrepancy between the two AI systems, developed at Bradford University and Art Recognition respectively, stems from their distinct operational principles. Bradford’s AI focuses on facial recognition, trained on datasets of faces, which can produce high similarity scores between images depicting the same individual, regardless of variations in face orientation, lighting, or quality. This program, while proficient in recognizing similar faces, isn’t in fact made for artwork authentication.
Conversely, at Art Recognition, we employ a different approach, considering various artistic elements like brushstrokes, chromatics, and object placements. Our training datasets, curated by a team of art historians and AI developers, encompass authentic artworks and a wide range of negative examples, helping the AI model to distinguish genuine art from counterfeits and providing a probability assessment for authenticity determination. The divergent results can be attributed to these different approaches.
A New Model. As the face recognition method proved inadequate for art authentication, the Bradford group developed a new technology more akin to the approach used by Art Recognition. Notably, a crucial difference emerged in the datasets used to train the respective AI systems. While the Bradford group’s AI was trained using 49 images, Art Recognition utilized a substantially larger dataset of over 100 images. Moreover, while the negative examples used by the Bradford scientists did not include any imitations of Madonna paintings, at Art Recognition we created a dataset of negative examples that includes fake Madonna paintings, but also imitations of other motifs such as allegories and religious scenes. This difference in the size and composition of the training datasets underscored the significant impact that data selection has on the outcomes of AI-driven art analysis. Following this controversy, we decided to release our full Raphael training dataset. The images and documentation can be downloaded following this link: https://github.com/Art-Recognition/Raphael_Dataset
The growing influence of AI in the art world demands transparency and accountability. The complexity of AI models necessitates rigorous scrutiny by peers and validation from the scientific community before deployment as a market tool. This ongoing debate sheds light on the intricate factors that influence the use of AI in determining art authenticity and explores how to effectively employ this tool within the art community.
