Imagine Raphael in his Roman studio, 1518: light from a high window illuminates the canvas, he mixes pigments with a student hunched beside him. The brush changes hands, a face hastily sketched. Four centuries later, a server in England buzzes: that face doesn't add up. It's not infamy: it's just mathematics meeting history.
The server hums
The screen in Bradford's lab lights up slowly, lines of code flowing like veins under the skin of a digital body. The algorithm, a modified ResNet50, swallows dozens of images: the Athens school, Fornarina, each brushstroke broken down into layers of color and shadow. Then comes the Our Lady of the Rose, canvas from the Prado, 1518-1520. The faces align: Mary serene, the Child holding out his hand, Saint John with a lost gaze. Everything flows, fluid curves, reds fading into the flesh like a held breath.
But the fourth face… Saint Joseph, top left, leaning downward, with a sparse beard and distracted-looking eyes. The algorithm slows down there, processing: transitions are too sharp, shadows don't blend. It's not a scanning error. It's a crack.
The study, led by Hassan Ugail from the University of Bradford, trained the model on authenticated works by Raphael: 98 percent accuracy, a number that sounds precise, almost too precise for art. They used a support vector machine to classify, pixel by pixel, the brush strokes and palette. The result: the entire painting wobbles, but that face doesn't. It's not his. Heritage Science, December 2023, puts it in black and white.
The crack in the Renaissance
The debate is not new. Since 1800, experts have been grumbling about Our Lady of the Rose: a canvas arrived at the Prado in 1813, attributed to Raphael but with a shadow of doubt. Saint Joseph, again, seemed less refined, a rough draft that clashed with the rest. Critics like Giovanni Morelli, in the late nineteenth century, noted jarring anatomical details: overly stiff fingers, wobbly proportions. But it was intuition, naked-eye comparisons, discussions in Roman cafés.
Now the crack is widening. The algorithm doesn't hesitate: different brushstrokes, perhaps by Giulio Romano, a pupil of Raphael who died young in 1530, with a drier, less lyrical style. Or another, unknown. Is the painting, valued at millions, losing a bit of its aura? Not exactly. But the story is cracking, like a canvas absorbing moisture.
Ugail makes it clear, in interviews in 2023: the computer sees through a microscope, beyond the human eye. Yet, it's no substitute. It's used for the provenance, the pigments, the condition of the canvas. A tool, not a judge. What if the face were really Romano's? A bit like a cook passing the ladle to his assistant: the dish comes out the same, but the flavor has a different note.
The men behind the machine
The Anglo-American team began in 2022, with a grant from the British Academy. Ugail, a mathematician with a penchant for art, combined computer science and history: ResNet50, pre-trained on Microsoft, then fine-tuned on Raphael. High-resolution images from the Prado, the Louvre, and private collections. They tested entire faces, not just the painting: the Sistine pass, the Transfiguration pure. Only there, in that distracted saint, does the system stumble.
Reactions? Cautious. The Prado confirms: analysis underway, but AI is adding data. Experts like carmen garrido, a museum restorer, talk about integration: don't throw away centuries of study for an algorithm. And in Italy, on Near future, ethics are being debated: AI rewrites history, but who decides the canon? A laboratory paradox, where humans program machines to judge humans.
"These signals could serve as a universal sign of intelligent life." No, sorry: that was for alien radar. For art: "AI sees details that escape, but art remains human." Ugail, again, in a 2023 paper.
The method in brief: Training on 50 well-known works by Raphael, focusing on texture and composition. Test on Our Lady of the Rose: 92% match for the rest, 65% for San Giuseppe. Low, suspicious. Hypothesis: intervention by an apprentice to speed things up, common in Renaissance workshops.
The face paradox
Here's the crux: AI, created to imitate humans, unmasks humans in their chaos. Raphael, a multitasking genius, delegated: Romano finished faces, others backgrounds. That clumsy Saint Joseph? Perhaps tired, after hours of posing. Or hasty, a student who copies but doesn't capture the soul. Irony: the machine, effortless, notices the tiredness of others.
And the value? A "pure" Raphael painting would be worth more, but this hybrid tells a more compelling story: workshops like pre-modern assembly lines, masters and apprentices waltzing together. AI highlights this, but doesn't explain why: was it urgency, or economics? A human detail, in an age of digital perfection, that jars like a misplaced brushstroke.
I think this is the problem: we look at art as an ideal, but it has always been imperfect, collaborative. The algorithm reminds us of this, mercilessly. Or perhaps too mercilessly: it only sees data, not the painter's sweat.
A rule for the future
In 2025, with an update on ScienceAlert, the issue resurfaces: AI doesn't stop debates, it fuels them. Hybrid experts, historians with laptops, use tools like this for digital catalogs. But the rule remains: art isn't a file, it's an accumulation of hands, doubts, and eras.
What if Saint Joseph were the silent hero? The one who stares away, knowing he's not the center of attention. Better a cracked painting than one as smooth as a screen. At least he can breathe.
The canvas remains at the Prado, its faces intact. The algorithm sleeps, awaiting the next mystery. Four hundred years later, we still stumble upon a distracted face.
