The clues were hiding in plain sight for four hundred years - microscopic ridges of dried oil paint, each one a fingerprint left at a crime scene nobody knew they were investigating. The suspect: El Greco's own workshop, long accused of finishing The Baptism of Christ after the master died in 1614. The detective: a deep learning algorithm called PATCH that nobody trained to recognize the suspects, because nobody could. And the reveal? Well, it turns out the art world may have been pointing fingers at the wrong crew this whole time.
The Cold Case File
Here's the backstory. Doménikos Theotokópoulos - you know him as El Greco - was working on the biggest commission of his career when death rudely interrupted. The Hospital de Tavera in Toledo had contracted him in 1608 to paint altarpieces for their chapel, including a massive Baptism of Christ measuring over ten feet tall. When El Greco died six years later, inventories described the painting as "sketched, not completed." It didn't get installed until 1624, a full decade after his death (Wikipedia).
So art historians did the reasonable thing: they assumed his son Jorge Manuel and the workshop apprentices grabbed their brushes and finished the job. For centuries, that was basically settled law.
Except maybe it wasn't.
Enter the Algorithm That Taught Itself
A team of researchers from Case Western Reserve, Purdue, and several other institutions built something genuinely clever. PATCH - Pairwise Assignment Training for Classifying Heterogeneity - is a machine learning method that can detect whether different "hands" worked on the same painting, and it does this without any labeled examples of who painted what (Van Horn et al., 2025).
Think of it like this: most AI art detectives need mugshots. You feed the network verified examples of Artist A's brushwork and Artist B's brushwork, and it learns to tell them apart. That's great for Raphael or Rubens, where authenticated works exist in abundance - researchers have hit 98% accuracy attributing Raphael's paintings with deep transfer learning (Cappa et al., 2023). But for anonymous workshop painters from 16th-century Toledo? There are no mugshots. These people didn't sign their work or leave behind convenient portfolios.
PATCH sidesteps this entirely. It takes a painting, chops it into tiny patches at microscopic resolution - we're talking individual bristle marks in the paint - and asks a deceptively simple question: do these two patches look like they were made by the same process, or different ones? It's achieving unsupervised results through supervised means, which is the kind of jazz-logic contradiction that actually works beautifully once you hear the whole riff play out.
The Plot Twist Nobody Expected
When the team turned PATCH loose on The Baptism of Christ, the algorithm found something that made art historians do a double-take. Those regions previously attributed to Jorge Manuel and the workshop? They showed underlying connections to the "El Greco" sections. The supposed evidence of multiple hands might just be one painter varying his technique - different brushes, different days, aging hands riffing on familiar themes with slightly altered phrasing (Scientific American).
As a control, they also analyzed Christ on the Cross with Landscape, where PATCH correctly identified heterogeneity consistent with multiple contributors - exactly what art historians expected for that piece. So the method isn't just telling everyone what they want to hear. It's improvising honestly.
The Fine Print (Because Science Has Fine Print)
Before anyone rewrites the textbooks, some caveats are in order. MIT researcher Mark Hamilton pointed out that "any AI system needs to be robustly tested on real unseen data" before we trust its predictions on priceless Renaissance art. The sample size is small. The training data came from paintings by nine contemporary art students at the Cleveland Institute of Art - a creative workaround, but a far cry from 16th-century workshop conditions.
PATCH outperformed both statistical methods and unsupervised machine learning approaches for this kind of pairwise comparison, which is encouraging. But "outperformed" and "definitively solved" are playing in very different keys.
Why This Groove Matters Beyond the Gallery
Here's where the melody gets interesting. PATCH isn't just an art history tool. The same problem - figuring out whether different regions of an image came from the same process without ground truth labels - shows up everywhere. Remote sensing images where you need to distinguish land use patterns. Manufacturing quality control where you're hunting for anomalies. Medical imaging where the "artist" is a disease process and the "canvas" is tissue.
The researchers essentially built a general-purpose heterogeneity detector that happens to have debuted on El Greco. That's like inventing the saxophone and choosing to premiere it at a four-hundred-year-old concert.
The Investigation Continues
The case isn't closed - it rarely is in art history or machine learning. But PATCH has introduced a new kind of witness to the stand: one that examines evidence invisible to human eyes, carries no art-historical biases, and lets the brushstrokes speak for themselves. Whether El Greco's ghost is vindicated or the workshop theory survives further scrutiny, the real breakthrough is the method itself - a dissonant, surprising chord that somehow resolves into something harmonious.
References
-
Van Horn, A., Smith, L., Mahmoud, M.S., et al. (2025). PATCH: A deep learning method to assess heterogeneity of artistic practice in historical paintings. Science Advances. DOI: 10.1126/sciadv.aea0485. arXiv: 2502.01912
-
Cappa, C., et al. (2023). Deep transfer learning for visual analysis and attribution of paintings by Raphael. npj Heritage Science. DOI: 10.1038/s40494-023-01094-0
-
Luo, H., et al. (2021). Discerning the painter's hand: machine learning on surface topography. npj Heritage Science. DOI: 10.1038/s40494-021-00618-w
-
Hincz Lab. Machine Learning for Art Attribution. GitHub: hincz-lab/machine-learning-for-art-attribution
-
"El Greco." Wikipedia. https://en.wikipedia.org/wiki/El_Greco
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.