Somewhere in a fifteenth-century Italian panel attributed to the Master of the Prado Adoration, there is a face that never belonged to the painting at all. In 2025, an MIT engineering student named Alex Kachkine printed a digital reconstruction onto thin polymer film, aligned it by hand and pressed it onto the damaged panel as compensation. Among the 5,612 regions he restored this way, most simply continued colour or pattern surviving nearby. One did something else entirely. The infant's missing face was lifted wholesale from a different painting, merely attributed to the same artist, aligned to the surviving fragments and transferred across. Visually, the join is seamless. Evidentially, it belongs to a different world from the rest of the surface around it.
The press loved the story, understandably. Three and a half hours of mask application against an estimated two hundred and thirty-two hours by hand, a laminate that could in principle be peeled off again, a tidy headline about artificial intelligence rescuing a damaged masterpiece. Nature called it generative AI. MIT's own press office called it AI-generated. IEEE Spectrum introduced the term "style transfer," a phrase that appears nowhere in Kachkine's own technical paper. Everyone reached for the same word.
I spent my MA dissertation in Conservation of Fine Art at Northumbria University taking that word apart, because it turns out to be doing something more than describing a technique. It is deciding, before anyone looks closely, which parts of a treatment deserve scrutiny and which do not. The research behind this piece, From Evidence to Intervention: Rethinking Treatment Authority in Computationally Mediated Paintings Conservation, compares these cases in far more depth than a single article allows, and develops a working framework for assessing them that I hope to take further.
Three cases, one habit of mislabelling
Kachkine's laminate is one of three recent projects I compared, each bringing computation into easel-painting treatment in a different way. A second, built by Mark Jeremy Narag, Julian Baumgartner and Maricor Soriano, trained an actual convolutional neural network, a genuine learned model, to predict how a damaged portrait might look once cleaned. A third, carried out separately by Baumgartner, used an old photograph of a sitter taken before damage occurred, registered and colour-matched by software, as a physical guide for retouching by hand.
Here is the detail the coverage missed entirely. The one genuine learned model in this group, the actual neural network, never touched a painting. It stayed exactly where its authors proposed, a preview for discussion, never used to guide a real treatment. The two methods that did physically enter paintings, Baumgartner's photograph scaffold and the bulk of Kachkine's own workflow, are built from conventional, rule-based image processing, registration, histogram matching, thresholding. Not a neural network in sight for most of it. Baumgartner says so himself, twice, in his own video, while happily describing Kachkine's very similar project as generative AI.
So the one piece of software that actually learned from data stayed furthest from the object. The pieces that went anywhere near a paintbrush were the plain, rule-following kind, the sort nobody bothers labelling at all. If a press release had you worried about the wrong thing, you were worried about the wrong thing.
Why the label is the wrong question
This matters for reasons that have nothing to do with vocabulary. Call something AI and people reach for one of two reactions, dazzled trust or reflexive suspicion, and either one skips straight past the actual question. That question has nothing to do with whether a model was trained on data or told explicitly what to do. It is whether the specific claim being made about this specific patch of paint is something the evidence can actually support.
Baumgartner's old photograph shows how differently that question can land within a single source. It fixed the geometry of the sitter's lost features with real confidence, because he could lay the printed scaffold over the surviving paint and check, by eye, that the proportions lined up before committing to anything. Its colour was a different story. The photograph had recorded a dirty surface under lighting nobody documented, so the colour information it offered was never more than a rough suggestion, something corrected entirely by hand during retouching. Same photograph, same workflow, two completely different standards of evidence sitting one on top of the other.
Kachkine's laminate shows the same split at a larger scale. Local colour, continued across a short gap from paint that still survives right beside it, rests on about as solid ground as reconstruction gets. The infant's transferred face rests on an attribution, a stylistic resemblance between two paintings, and a guess that a shared template drawing might once have connected them. Both ended up fixed permanently onto the same panel, printed by the same process, with nothing in the finished surface to tell a viewer which is which.
Even the neural network's own restraint has a quieter evidential story behind it. Its accuracy was measured against photographs of paintings someone had already cleaned, not against any independently verified record of how the paintings actually looked before damage. Agreement with that target shows the model can reproduce a photographed result. It shows nothing about whether that result, or the model's prediction, recovers anything true about the painting's history. The CNN stayed advisory partly because nobody tried to give it more authority than that. It is a useful reminder that strong-looking numbers and strong historical evidence are not the same thing, learned model or not.
What gets lost along the way
There is a second, quieter problem sitting underneath the labelling one, which is what happens to the record once the treatment is finished. Kachkine documents his mask-selection process in real detail, the thresholds, the tolerances, the code. He says much less about the reconstruction that fed into it, which methods produced which regions, or why the infant's face came from that particular painting rather than another. A future conservator could, in principle, recover exactly what was printed and where. They would have a much harder time recovering why any individual piece of it was considered justified in the first place. Precise records of the wrong stage of the process do not add up to an accountable one.
None of the old questions stop applying
What I kept returning to, working through all three cases, is that none of the arguments conservators have used for a century actually stop applying here. We already ask whether reconstruction is proportionate to its purpose, whether an addition is honestly distinguishable from the original, whether a treatment can be reversed, whether a conservator genuinely understands and controls what they are approving. Computation does not require a new rulebook. It means those same questions now have to be asked about each individual claim inside an output, rather than about the finished image as a whole, because a single coherent-looking surface can smuggle a well-evidenced inch of paint and a complete invention past the same checkpoint together. In the fuller dissertation, I set this out as a Necessity Test followed by an Ethical Gradient, with four principles, Evidential Proportionality, Meaningful Professional Control, Transparency of Uncertainty and Epistemic Reversibility, doing the actual work of deciding how much authority a given claim has earned.
None of this makes Kachkine's laminate bad work, and it is not an argument against using computers in conservation at all. It is an argument against letting a headline word decide, on our behalf, which parts of a treatment get examined closely and which get waved through because a human did not technically draw them freehand. The infant's face will likely outlast this particular news cycle by a few centuries. It would be nice if, by then, somebody other than the conservator who applied it still knew where it actually came from.
This article covers one strand of a much longer piece of research. If the questions raised here are ones your own institution or practice is grappling with, I would welcome the conversation, and I am actively looking at how to bring the full dissertation to publication.
Sources
- Piper, Rhiannon, 'From Evidence to Intervention: Rethinking Treatment Authority in Computationally Mediated Paintings Conservation' (unpublished MA dissertation, Northumbria University, 2026) — the full research behind this article
- Kachkine, Alex, 'Physical Restoration of a Painting with a Digitally Constructed Mask', Nature, 642.8067 (2025), pp. 343–50
- Narag, Mark Jeremy G., Julian Baumgartner, and Maricor Soriano, 'Digital Restoration of Paintings Using Convolutional Neural Network Trained on Cleaned Edges and Local Patches', Optics for Arts, Architecture, and Archaeology (O3A) X, 13569 (2025)
- Baumgartner Restoration, Retouching Made Easy. No, Really! (2025), YouTube
- Kutzke, Hartmut, 'Digitally Constructed Mask Brings a Damaged Painting Back to Life', Nature, 642.8067 (2025), pp. 305–06
- Chu, Jennifer, 'Have a Damaged Painting? Restore It in Just Hours with an AI-Generated "Mask"', MIT News, 11 June 2025
- Dolgin, Elie, 'Art Restoration Meets Digital Innovation', IEEE Spectrum (2025)
