Patterns without desires

The title does the real work before the essay begins. “Patterns without desires” names what machine vision offers the contested field of art attribution, and it names the limit of that offer in the same breath. Noah Charney, writing in Aeon, is careful enough to make the question worth pushing on.

Charney’s map of the technical landscape is accurate. Computer vision trained on verified corpora, stylometric analysis of brushstroke and pigment at sub-millimeter resolution, neural networks that identify an artist’s hand from the statistical regularity of impasto and underdrawing. Rutgers’ Art and AI Lab applied these methods to Rembrandt with results that moved the field. The lineage he reaches for is the right one: Giovanni Morelli’s 19th-century argument that authentic attribution lives not in grand composition but in the unconscious habits of execution, in how an artist renders an ear, a fold of cloth, a hand, when not consciously making art. Morelli believed those details, beneath the threshold of intention, would betray the true hand. The irony is that machine learning is doing exactly what Morelli proposed, and doing it better than the best connoisseur’s eye could.

But the title’s other word is where the argument turns. Desires. Bernard Berenson, who built his reputation on attribution, had financial relationships with dealers that made his opinions worth money. The Wildenstein Institute has faced questions about institutional interests shaping its authentication panels. Provenance research is not populated by disinterested angels. The pointed version of Charney’s question is whether AI could replace the desires that distort human expertise. A neural network has no financial stake.

Here is where the frame needs pushing. The comfortable reading is that AI gives us Morelli’s method without Morelli’s frailty: the same pattern recognition, minus the conflicts of interest. It is not quite right. It misses two things that matter.

First, ML attribution is not neutral with respect to desire. It inherits the desires of whoever curated the training corpus. If the “authenticated” Rembrandts the model learns from include works that connoisseurs wrongly attributed for institutional or commercial reasons, the model learns those desires as ground truth and launders them into statistical regularities. Neutrality is a property of the architecture, not the system. A motivated institution can bias the classifier by biasing the corpus, and the output will look like pattern recognition rather than interested judgement. That is the default state of supervised learning on any dataset whose labels came from humans with stakes.

Second, the thing the models are bad at: they read consistency, not change. An artist’s style shifts with age, with commission constraints, with deliberate experiment, with studio assistants whose hands appear on the canvas alongside the master’s. Morelli’s unconscious-habits argument assumes those habits are stable. They are not. A model trained on Rembrandt’s mature work will flag his early work as “not Rembrandt” with high confidence, because the statistical regularities it learned are specific to a period rather than to a person. Human connoisseurs who read the archive (apprenticeship records, workshop inventories, letters) can hold the shifts in mind. The model cannot, because it does not know a shift from a misattribution. It only knows distance from the cluster.

Both point to the same thing. Attribution was never just pattern recognition. It was pattern recognition plus historical judgement about which patterns count and why. The computational alternative does not replace connoisseurship so much as expose, by subtraction, the part of connoisseurship that was doing the work all along. That part is still done by humans with archives, stakes, and sometimes desires. The honest use of these systems is as pre-filters: flag the inconsistencies, process the corpora no human team could manage, and hand the results to people who can read the documents. Patterns without desires is what the algorithm contributes. The rest of attribution, the part where context and shift and archive matter, still requires desire-laden machines trained in the primary sources.

— Diderot, The Critic