There's a stranger in this Raphael

Look at the rose. It sits on the parapet in the corner of the canvas, the detail this painting is named for. And for a century and a half, whenever anyone doubted this picture, this is where they pointed. The rose, and the ground beneath it. Hold on to that. Then put it down, because it is not the rose. Move your eyes up, into the shadow at the top left, to a face nobody puts on the postcards. An old man, half turned, barely lit. In December 2023, a computer scientist, who is not an art historian, pointed a neural network at this canvas. He did not ask it about brushwork, or provenance, or the rose. He asked it, face by face, one simple question. And it answered something that 150 years of trained eyes suspected, but could never prove. Three of these four faces are Raphael. The painting is called the Madonna della Rosa, painted somewhere between 1518 and 1520, though one of the researchers who studied it, Howell Edwards, dates it slightly earlier, around 1517 or 1518. Even the year is contested. It hangs in the Museo del Prado in Madrid, about 103 by 84 centimetres, roughly the size of a kitchen tabletop stood on end, painted not on wood, but on canvas. Four figures fill it. Mary, the infant Jesus, a young John the Baptist, all lit, all forward, close enough to touch. And behind them, pushed into the top left corner, half in shadow, an old man, Joseph. Now look for the paperwork, because that is usually where a painting's identity gets settled. Here, the trail starts late and starts abruptly. In 1667, the painting is recorded hanging in a monastery in Spain. It stays in Spanish hands until it enters the National Collection in 1857. Before 1667, there is nothing. No sale, no commission, no inventory listing it, no letter mentioning it, no name attached to its making. Whatever happened to this canvas in the two centuries after Raphael's death simply was which leaves attribution with exactly one tool to work with. Not a document. And the eye, it turns out, is not enough. At first, nobody doubted it at all. Early connoisseurs regarded the Madonna della Rosa as a Raphael autograph, meaning he painted every inch of it himself. Then, in the mid-1800s, the art historian Johann David Passavant looked again and started asking questions. He and others began attributing the execution partly to Raphael and partly to his workshop. And over time, suspicion settled on one name in particular, his pupil Giulio Romano, with Gianfrancesco Penni floated as a second possibility. The specific complaint was always the same. The Madonna, the child, and the young John the Baptist were judged far superior to Joseph, who felt, to more than one trained eye, like an afterthought bolted on in the workshop. And yet, in Spain, where the painting actually lived, nobody ever questioned it at all. Enter someone who had never spent a career staring at brushwork. Hassan Ugail directs the Centre for Visual Computing and Intelligent Systems at the University of Bradford. A mathematician, a computer scientist. He and his team built an algorithm and fed it paintings we already know came from Raphael's own hand, training it to recognise his style down to the brushstrokes, the colour palette, the shading, every aspect of the work. There was a problem immediately. One dead painter does not leave enough pictures to train a network from scratch. So they borrowed one already built for something else entirely. A Microsoft architecture called ResNet50 and paired it with an older machine learning method, a support vector machine, until it could read roughly 4000 separate features in a single canvas. Tested against known Raphael's, it hit 98% accuracy. It does not know who Joseph is. It only knows what Raphael's hand does. Normally the process is simple. Feed the machine a whole picture and it returns one answer. Authentic Raphael or not? So they fed it the Madonna della Rosa, whole, the way it had always been judged. And the machine gave them nothing. When we tested the Della Rosa as a whole, Ugail said, the results were not conclusive. A 98% instrument going silent is not a failure. It is a symptom. It means the canvas is not speaking with one voice and something inside it is dragging the average down. So the team changed the question. Instead of judging the painting as a single object, they cut it apart and asked the machine to look at one face at a time. Four faces went in separately, one after another. The Madonna passed. The Christ child passed. The young John the Baptist passed. Joseph did not. The rest of the picture was confirmed as Raphael, Ugail said. Joseph's face came up as most likely not Raphael. Howell Edwards put it in stronger terms still. The analysis demonstrated that the Madonna, the Christ child and St John the Baptist are unequivocally painted by Raphael and that Joseph is not and was painted by someone else. The name that keeps surfacing is the same one connoisseurs had already reached for. Giulio Romano. Nobody can say that for certain, but notice what just happened. A neural network with no knowledge of art history landed on the exact same face that trained human eyes had been uneasy about for 150 years. Now go back to the rose. For decades that was the standing theory, that the rose and the lower portion of the canvas were the parts painted by someone else, a claim still repeated in reference entries about the work today. The algorithm checked it anyway. The lower part of the painting it found is most likely by Raphael, which means art history had the right painting and the right suspicion for 150 years. It simply had the wrong corner. It was looking down when it should have been looking up. The reception was not warm. Hugail has said art historians were a bit dismissive of the findings and this same team's earlier claim about another painting, the De Bressi Tondo, was already rejected by old masters experts elsewhere. Hugail himself draws the line carefully. This is not a case of AI taking people's jobs, he says. Authentication still runs on provenance, pigments, condition, everything a document can tell you that a network cannot. His collaborator, David Stork, calls it a step toward better protocols, not a verdict. Hugail released the code anyway for anyone to test. Go back to Madrid. The museum's own page for this painting still credits it simply to Raphael. No asterisk, no footnote about the corner. And that is the shape this leaves you with. A face that three centuries of documents cannot account for. That 150 years of trained eyes suspected without ever proving. And that a neural network flagged in a fraction of a second without knowing his name, his story, or that anyone had ever doubted him at all. This algorithm was built for one painter. It has already looked at a handful of others. Somewhere in the world's museums are workshop paintings with one figure that has always felt slightly wrong. Next time, don't look at the rose, look at the corner, look closer.