For centuries, deciding whether a painting is a genuine Rembrandt or the work of a skilled imitator has rested on the trained eyes of experts, later helped by X-rays and infrared imaging. Researchers at Carnegie Mellon University have now added another layer: an explainable AI framework that looks for visual patterns linked to Rembrandt’s work and shows experts which parts of a painting drove its verdict. The work is published in the Proceedings of the National Academy of Sciences.
Building an explainable AI for Rembrandt
The project was led by mechanical engineering professors Phil LeDuc and Jon Cagan, with alumnus Andrés Arias-Rosales and Rembrandt scholar Arthur Wheelock Jr. Wheelock spent more than 40 years as a curator at the National Gallery of Art in Washington, D.C., and is an internationally recognised expert on Rembrandt and Dutch art.
Before building the tools, the team visited the National Gallery to study Rembrandt paintings and discuss his style, technique and the open questions around his work. Attribution is hard: experts weigh historical documents, subject matter, Rembrandt’s changing styles, brushwork, paints and supports. They also have to allow for pupils and assistants who may have painted alongside him, and later restorers who may have altered his compositions.
“We wanted to understand Rembrandt through the eyes of a world-class expert, then understand it through the eyes of ‘the machine,’ our AI approach, and bring those together,” said Arias-Rosales.
How the explainable AI model works
The researchers curated a dataset of more than 1,300 high-resolution digitised images of paintings, including 311 Rembrandts. The model learned statistical patterns across features such as edges, colour, light distribution and visible brushstrokes. It then produces a score estimating how closely a painting matches the patterns seen in Rembrandt’s work.
According to CMU, when the model was highly confident, it classified paintings with 95.8% accuracy.
The model classifies a painting as Rembrandt or not, gives calibrated probabilities and confidence scores, and creates visual maps highlighting the areas that most influenced its decision. The main explainability method in the paper is called Unified GradCAM.
“Experts need to understand how the AI model/algorithm came to its decision,” said Cagan. “That is critical in successful human-AI collaboration.”
These maps can point curators to details they might otherwise miss. CMU gives an example: if the AI repeatedly highlights a figure’s hands with a low score, researchers can digitally remove the hands and ask the model to classify the painting again. If confidence in a Rembrandt attribution then rises, a curator might consider whether a student painted the hands.
Testing the explainable AI on a real mystery
The team has already tried the tool on an open case. In 2024, a painting titled Portrait of a Girl was found in an attic in Maine and later sold for US$1.41 million because of its possible link to Rembrandt. The CMU model gave it an 84.7% confidence of being a Rembrandt.
“These results aren’t an authentication, they are a statistical observation meant to be considered alongside historical research, material analysis, X-rays, and other tools with the final decision made by human experts,” said LeDuc.
The team hopes the tool could become an interactive application for museums, curators and other art professionals. They believe it could be adapted to study other artists as well.
As a writer, I appreciate that the researchers frame this as a tool for raising better questions rather than giving final answers. I think that framing matters, because the stakes in art attribution are high for collectors, insurers and museums alike.
FAQ
What is CMU’s explainable AI for Rembrandt?
It is an AI framework from Carnegie Mellon University that scores how closely a painting matches Rembrandt’s visual patterns and highlights the areas of the painting that most influenced that score.
How accurate is the model?
According to CMU, it classified paintings with 95.8% accuracy when it was highly confident. It was trained and tested on more than 1,300 paintings, including 311 Rembrandts.
Can the AI authenticate a painting on its own?
No. The researchers say its results are statistical observations to be weighed with historical research, material analysis and X-rays, with human experts making the final decision.