AI Raphael painting analysis has identified something unusual inside one of the Renaissance master’s best-known religious paintings: the face of Saint Joseph may have been painted by somebody else. Researchers examining the Madonna della Rosa found that most of the work resembled Raphael’s established style, while Joseph’s face appeared sufficiently different to raise questions about another artist’s involvement.
The finding does not mean the painting is a forgery, and it does not prove who produced the disputed face. It adds computational evidence to a debate art historians were already having about how much of the composition came directly from Raphael and how much may have involved his workshop.
For anyone standing in front of an old masterpiece, the question is irresistible. When a painting carries one famous name, how many different people actually helped create what you are seeing?
What AI Raphael painting analysis found
The research, published in npj Heritage Science, examined whether a machine-learning system could distinguish authenticated Raphael paintings from work by other artists. The researchers then applied their method to different sections of the Madonna della Rosa.
When the complete painting was assessed, the system produced a Raphael probability of 0.57. The authors treated that relatively uncertain result as a reason to examine individual areas more closely rather than as a final verdict.
The closer inspection suggested that the Madonna, Christ Child and Saint John were consistent with Raphael’s style. Joseph’s face stood apart, prompting the researchers to suggest that another artist may have contributed that section.
The University of Bradford’s account of the research attributes the work to a team including Hassan Ugail, David G. Stork, Howell Edwards, Steven C. Seward and Christopher Brooke. Their conclusion concerns a probable difference in artistic style, not a proven act of deception.
Why the face of Joseph matters
The Madonna della Rosa depicts a familiar religious grouping centred on the Virgin Mary and the infant Christ. Joseph appears as a secondary figure, and questions about the consistency of his execution have existed within art-historical discussion before the AI analysis.
That history matters because it would be misleading to suggest the computer discovered a mystery nobody had ever suspected. The researchers themselves acknowledge earlier scholarly doubts and describe their findings as supporting an existing discussion.
What changes is the method. Rather than depending only on a specialist’s visual judgement, the system compares patterns across many image features, including characteristics associated with edges, texture and brushwork.
The result is best understood as an additional witness, not the judge. It can tell historians that one area appears unusual compared with the examples used to train the model. It cannot independently reconstruct the full circumstances in which a Renaissance painting was made.
LiveAIWire’s earlier examination of AI archaeology and cultural heritage identified a related pattern: computational tools can reveal otherwise difficult-to-spot evidence, but its meaning still depends on historians, conservators and the physical record.
How a computer learns to recognise an artist’s style
The researchers adapted a computer-vision model called ResNet50, which had already learned to identify visual features across a large and varied image collection. They then combined those image features with a classification system designed to separate Raphael works from paintings by other artists.
This approach, known as transfer learning, allows researchers to reuse capabilities developed for general image recognition. The model does not start with an understanding of Renaissance history. It learns which visual patterns are associated with examples labelled Raphael and which appear in the comparison group.
The paper reports that the researchers used 49 authenticated Raphael images and 49 images by other artists. That is an understandable restriction when studying a historical painter with a limited body of surviving work, but it is also an important limitation.
A system trained on a relatively small set can learn useful stylistic distinctions without becoming an infallible authority. Lighting, image quality, restoration, reproduction methods and the selection of comparison artists can all affect what a computer appears to recognise.
The researchers also used edge-analysis techniques to study visual details linked to brushwork and other artistic marks. Such methods can highlight differences that deserve investigation, but the system is still working with images rather than independently examining every layer of paint or historical document.
What the 98% accuracy claim really means
The study reports 98% accuracy on its image-classification validation task. That figure describes performance within the particular dataset and comparison exercise created by the researchers.
It does not mean there is a 98% probability that Joseph’s face was painted by a specific person. Nor does it establish that the model would identify any Renaissance forgery with the same accuracy under every lighting condition, restoration history or photographic setup.
Those distinctions matter because classification accuracy can sound far more general than it is. A model may perform strongly when distinguishing selected Raphael images from selected comparison artists, yet face a much harder challenge when comparing closely related painters working in the same studio.
The researchers acknowledge that their work concerns a restricted class of paintings and contributes only one part of a complete authentication process. Provenance, physical materials, historical context, conservation records and expert interpretation remain essential.
That caution reflects a broader issue LiveAIWire has explored in relation to the difference between algorithmic accuracy claims and real-world reliability. An impressive percentage means little unless the underlying comparison and its limitations are clearly understood.
Could the second artist have been Giulio Romano?
One possible explanation involves Giulio Romano, an important member of Raphael’s workshop. The National Gallery identifies Romano as Raphael’s principal pupil and assistant, placing him firmly within the historical circle relevant to the debate.
The researchers note that some art historians had already suggested Romano may have contributed to the painting. Their model, however, does not prove that Romano produced Joseph’s face.
Identifying a section as different from established Raphael examples is not the same as matching it conclusively to a named alternative painter. To establish that, researchers would need appropriately validated comparisons against Romano and other plausible workshop contributors.
The responsible conclusion is therefore narrower. Joseph’s face appears less consistent with the Raphael examples analysed, and workshop participation is a historically plausible explanation. The precise identity of any second artist remains unresolved.
Renaissance workshops were not one-person businesses
The idea that another artist contributed to a Raphael painting can sound scandalous only if we assume that historical masterpieces were always produced by one isolated genius. Renaissance workshops were collaborative environments involving assistants, pupils and specialist contributors.
The Metropolitan Museum of Art’s discussion of Raphael and his workshop describes major projects completed with the help of assistants. It also identifies Giulio Romano as a leading figure within Raphael’s studio.
Collaboration could involve preparing surfaces, developing background details, transferring designs or completing secondary figures. A work might still be legitimately associated with a master whose composition, supervision and principal figures defined the finished piece.
That is why the suggestion of another hand should not automatically be described as fraud. Understanding who contributed what can enrich the history of an artwork without cancelling its connection to the artist named on the museum label.
It also complicates modern assumptions about authorship. As LiveAIWire has examined in its coverage of technology and the creative industries, the idea of a single, easily identifiable creator does not always capture how important cultural works are actually produced.
What AI can reveal, and what it cannot settle
Machine-learning tools may help researchers compare visual features across large collections, identify sections that deserve closer inspection and test whether particular patterns recur across attributed works. They could also help prioritise paintings for more detailed scientific or historical analysis.
However, a photograph cannot provide the complete physical history of a painting. Later repairs, changing varnish, damaged surfaces and previous restoration work may alter the appearance of areas that were originally created by the same person.
A model can also inherit assumptions from its training material. If the paintings labelled authentic are misattributed, unusually photographed or unrepresentative of an artist’s development over time, its outputs may reproduce those limitations.
The same concern appears in LiveAIWire’s reporting on AI-generated novels and the hidden patterns beneath convincing creative work. A computer can detect regularities across examples, but interpreting those regularities requires a clear understanding of what was measured.
There is also a difference between recognising a visual style and understanding artistic intention. As explored in LiveAIWire’s analysis of whether machines can demonstrate genuinely open-ended creativity, pattern recognition does not automatically amount to the human judgement required to interpret an artwork’s cultural significance.
The most interesting outcome is not that AI has replaced art historians. It is that a familiar masterpiece can be examined from another angle, exposing questions that sit between technology, historical scholarship and the realities of collaborative artistic production.
Joseph’s face may indeed reflect another artist’s work. The machine cannot settle that question by itself, but it gives researchers a specific place to look and a clearer reason to keep asking who held the brush.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.