Can a Machine Spot a Fake? AI, the "Visual Fingerprint," and the Fight Over Authentication

For as long as there has been an art market, there have been forgeries, and for almost as long the job of catching them has fallen to a small priesthood of connoisseurs — experts who could look at a drawing and simply know, from the weight of a line or the confidence of a shadow, whether it belonged to the hand it claimed. That instinct has always been part of what makes the field feel closed to newcomers. Now a run of recent research and start-up activity is asking an uncomfortable question: could a computer do the same job, and might it do it more transparently? For anyone who buys art, insures it, or simply enjoys looking at it, the answer matters more than it first appears.

A British study puts a number on connoisseurship

The most substantial development came this summer from the University of Bradford, where Professor Hassan Ugail and colleagues published a study in the journal PLOS ONE showing that an AI system could tell a genuine drawing from a fake with close to 90 per cent accuracy, using just five visual measurements. The team tested the method on 900 authenticated drawings by ten artists spanning five centuries, from Michelangelo and Raphael to Whistler and Constable, drawing on images from the Ashmolean, the Victoria and Albert Museum, the Royal Collection Trust and the Metropolitan Museum of Art.

What is striking is not only the headline figure of 89.8 per cent, but the shape of the errors. The system correctly accepted genuine works around 83 per cent of the time while rejecting impostors at a rate above 90 per cent. In plain terms, it is especially good at avoiding the single most damaging mistake in authentication — waving through a fake as the real thing. It does this by learning what “normal” looks like for one specific artist, building what Ugail calls a “visual fingerprint” from features such as line structure, texture, contrast and tonal variation, and then flagging anything that does not fit.

Small data, sharp tools

The Bradford approach is quietly radical for another reason. Most headline-grabbing AI systems are hungry for enormous datasets, but art history rarely obliges. Verified drawings by a given Old Master might number in the dozens, not the thousands. The model was designed to work from as few as twenty authenticated sheets per artist — a deliberate concession to the reality of the archive rather than the laboratory. Notably, the researchers found that larger, general-purpose deep-learning models performed worse at this narrow task, rejecting many genuine works. It is a useful corrective to the assumption that bigger is always better, and a reminder that the most valuable tools are often the ones built for a single, specific problem.

Turning brushstrokes into a searchable web

Where Bradford focuses on drawings, a parallel effort is taking on paintings from a different angle. The New York start-up QuantumSpace — founded in Italy, with an office in Milan, and dreamed up by a founder who developed the idea as a graduate student in London — is using knowledge-graph technology to map the relationships between artworks, artists, pigments, locations and provenance. The company says it can extract up to 60,000 data points from a single painting, and that its archive now covers some 68,000 works described by roughly 3.5 billion individual data points.

The practical payoff is less about grand pronouncements of authenticity and more about spotting anomalies a human might miss. The software has been credited with identifying an undocumented restoration in a work attributed to Caravaggio after inconsistencies surfaced in its brushstroke analysis, and it appears well suited to reading craquelure, the fine web of hairline cracks that forms as an oil painting ages. For galleries, one of the most immediately useful tricks is mundane but expensive to solve by other means: confirming that a work returned from loan is physically the same object that left, with no fresh humidity damage or surface change.

Why this matters beyond the saleroom

None of this is a parlour trick. The global fine-art market returned to growth in 2025, with total sales estimated at around $60 billion, and forgery — hard to quantify precisely, but estimated by some at up to $1.6 billion a year — remains a genuine drag on confidence. A convincing fake does not only cheat a buyer; it distorts valuations and muddies an artist’s legacy for generations. Anything that adds a reproducible, evidence-based layer to attribution has value for museums, insurers and the courts, not just dealers.

The researchers and entrepreneurs alike are careful to frame their tools as support, not replacement. Ugail is explicit that the aim is to strengthen connoisseurship rather than retire it, combining human expertise with computational evidence. That framing is wise, because the same AI advances that help catch forgers are also helping produce them. The honest position, for now, is that a machine can offer a second opinion that is fast, consistent and open to scrutiny — and in a field long governed by the unexplained certainty of the expert eye, a little transparency is no bad thing. For the new collector standing in front of a drawing they cannot quite read, that shift is worth watching.