Patterns without desires

The art market likes certainty. It prefers painters' names printed in bold type, dates fixed neatly to a decade, values pinned confidently to price estimates. Yet behind this appearance of assurance lies an industry structured around risk. Enormous sums of money change hands on the basis of attribution (who created this work), and attribution itself often rests on a fragile, negotiated consensus or the determined opinion of one or more experts (some of them more self-proclaimed than objectively expert) rather than an established fact.

A single name on a museum label can anchor the value of an artwork, determine its place in art history and shape scholarly narratives for generations. Change that name, and the consequences ripple outward. A painting can lose tens of millions in value. Is the world's most expensive painting, Salvator Mundi ('Saviour of the World'), indeed by Leonardo da Vinci, as many of the top Leonardo scholars confirm? Its last sale price was $450.3 million in 2017 at Christie's. But scholars of equal renown are convinced that it is a derivative work, not by Leonardo at all, in which case it might be worth only $450,000. That's quite a difference for an opinion to make. Enough question marks have been thrown up around Salvator Mundi that the painting has yet to be displayed in public since its sale. It is thought to have been purchased by proxy by the Saudi crown prince Mohammed bin Salman to become the centrepiece of the Louvre Abu Dhabi, but it is not on display and its location is not now publicly known.

Salvator Mundi (c 1500), attributed to Leonardo da Vinci. Courtesy Wikipedia

Calling attribution into question can cause a museum's reputation to wobble. A collector's confidence can evaporate. Governments, insurers, lenders, heirs and institutions all depend on attribution being right, even though they know, at some level, that it might not be. This tension is not accidental. The art market is an overheated system built on partial information, asymmetries of knowledge, and incentives that quietly encourage optimism. Sellers benefit from the highest plausible attribution. Buyers hope that the name on the label will hold. Auction houses rely on inherited scholarly opinions that may be decades old. Museums, once they've committed, are rarely eager to reverse themselves. The result is a system that functions not because certainty is common, but because doubt is carefully managed.

Attribution, in this sense, is not merely a scholarly exercise. It is the keystone of an economic and cultural structure. Without it, prices collapse, catalogues unravel, and historical narratives lose coherence. And yet attribution is also deeply human, shaped by judgment, intuition, training and, inevitably, bias.

Most of art history

For most of art history, attribution rested almost entirely on connoisseurship: the trained eye of experts who compared brushwork, composition and handling across an artist's oeuvre. Connoisseurs developed astonishing sensitivity to visual nuance, often identifying hands and workshops with remarkable precision. But connoisseurship is also subjective. Two equally qualified scholars might disagree profoundly, and have done, repeatedly, down the centuries. In some cases, their disagreements have lasted generations.

The 20th century introduced new tools that seemed to promise firmer ground. Scientific and forensic analysis allowed scholars to test materials, identify anachronistic pigments, examine underdrawings, and date supports – the panel or canvas on which a painting is made. These techniques revolutionised the detection of forgeries and exposed many celebrated fakes. But they did not solve attribution. Forensic tests are excellent at telling us when something cannot be what it claims to be. They are far less effective at telling us who painted it. Forensic testing is also expensive, time-consuming, and sometimes invasive. Paintings must often be transported to specialist laboratories. Microscopic samples may be removed. Insurance premiums soar. As a result, such analysis is applied sparingly, usually only to works considered important or problematic. The vast majority of artworks entering the global market each year are never tested at this level. And so the art market continues to rely heavily on expert opinion, supported by provenance research and, sometimes, selective scientific evidence. It is a system that works well enough to sustain a multi-billion-dollar industry, but poorly enough to allow persistent uncertainty to thrive beneath the surface.

This uncertainty helps explain why the arrival of artificial intelligence in the realm of attribution has provoked such strong reactions. For some, AI looks like a long-awaited corrective: a way to introduce consistency, scale and empirical rigour into a system long governed by authority and trust. For others, it represents a profound misunderstanding of what art is and how it should be judged. Bendor Grosvenor, a British art historian and television presenter whose work has done much to popularise close looking and traditional connoisseurship, and who has himself identified a number of lost Old Master works, is sceptical. Writing in the Financial Times in November 2025, Grosvenor expressed deep reservations about the use of AI in attribution. His concern is not simply that machines might get things wrong, but that the very premise of algorithmic judgment risks flattening the complexity of artistic creation.

These fears misunderstand what AI image analysis actually does, and what it does not do

Painting, in Grosvenor's view, is not reducible to pattern recognition. It involves intention, context, decision-making and deviation. Artists break their own habits. They experiment. They collaborate. They respond to commissions, materials and circumstances. To treat attribution as an exercise in statistical matching is to misunderstand the nature of artistic practice itself. There is a danger, he says, that once AI is granted authority, it could marginalise human expertise rather than supplement it. Art history has long resisted rigid systems. Its greatest insights often emerge from anomalies, from works that do not fit neatly into existing categories. An algorithm trained on an artist's 'typical' works might struggle with the atypical, the experimental, or the unfinished. Worse, it might discourage scholars from trusting their own eyes when faced with something genuinely new.

Grosvenor's critique gestures toward a broader anxiety: that AI represents a technocratic encroachment on a discipline grounded in human judgment. Art history, after all, is not physics. It does not deal in universal laws. It is interpretive, contingent and deeply historical. To replace disagreement with probability scores risks draining the field of its intellectual vitality. Yet this is only one side of the debate. Others argue that these fears misunderstand what AI image analysis actually does, and what it does not do.

Clovis Whitfield, an art historian and dealer with decades of experience in attribution debates, has responded forcefully to Grosvenor's concerns. Writing in the same newspaper, Whitfield argues that AI does not replace connoisseurship but builds upon it, since algorithms are trained on the accumulated judgments of human experts – catalogues raisonnés, museum collections, peer-reviewed attributions: they learn from connoisseurs before they ever challenge them. AI adds not taste or interpretation, but scale and consistency. It can compare thousands of micro-features across an artist's corpus, identifying statistical regularities that no human could track consciously. It does not pronounce verdicts. It produces probabilities. And those probabilities, Whitfield argues, are simply another form of evidence, one that scholars are free to accept, reject, or debate.

Exchanges such as this

Exchanges such as this crystallise a deeper tension at the heart of attribution. On one side stands the tradition of human judgment, with all its sensitivity and subjectivity. On the other, the promise of empirical pattern-analysis, indifferent to reputation, market pressure or desire. The real question is not which should prevail, but how they might coexist. To answer that, we need to step back and ask why attribution matters so much in the first place.

Part of the answer lies in economics. The difference between a definite 'by' and the shakier 'attributed to' can mean millions. But the obsession with attribution runs deeper than money. It reflects a cultural fixation on authorship that has shaped Western art history since the Renaissance. We are not content to admire a painting; we want to know who made it. We want to embed creativity inside a name, a biography, a singular genius. This fixation is, in many ways, anachronistic. In the early modern period, artists rarely worked alone. Large workshops produced paintings collaboratively, with masters overseeing teams of assistants who executed backgrounds, drapery, architecture or secondary figures. Peter Paul Rubens ran one of the largest studios in 17th-century Europe, delegating substantial portions of his compositions while reserving key elements (faces and hands) for his own hand. Yet the market continues to prefer the simplicity of a single 'author'.

AI has no stake in whether a painting is worth millions or thousands. It simply measures similarity

In 1935, the philosopher Walter Benjamin reflected on the 'aura' of the artwork in an age of mechanical reproduction. Today, we might speak of the aura of authorship in an age of mechanical analysis. Attribution has become a fetish, invested with meanings that extend far beyond historical accuracy. It is bound up with authenticity, originality and value, even when those concepts are only loosely connected. It explains why disputes over attribution can become so heated: they are not merely about brushwork or pigment, but about identity, legacy and power. To accept that a cherished painting is not by the artist it was thought to be is to accept a loss: of value, of prestige, of narrative coherence.

In this context, artificial intelligence is not merely a new tool, but a disruptive force. It threatens to unsettle long-standing agreements and expose the provisional nature of judgments that the market prefers to treat as settled. It does not care about sunk costs or institutional pride. It has no stake in whether a painting is worth millions or thousands. It simply measures similarity.

Its indifference makes people uneasy. A system built on trust and reputation is naturally wary of a method that bypasses both. Yet its impartiality also makes AI potentially valuable. In a market where incentives quietly push toward optimism, a tool that is immune to desire may offer a corrective.

For now, the debate remains largely theoretical. Critics worry about what AI might do to art history. Proponents point to what it already achieves. To move beyond abstraction, we need to look closely at cases where attribution is genuinely contested, where experts disagree in good faith, and where AI has entered not as judge but as an unexpected participant in an ongoing argument. Such cases show how AI can sharpen disputes rather than resolve them, forcing scholars to articulate assumptions that might otherwise remain implicit. They also expose the real problem facing the art market, which is not whether machines can judge art but whether the system is willing to confront how uncertain its judgments have always been.

One such case

One such case centres on Caravaggio's The Lute Player, a composition known in several versions and long contested by scholars. Caravaggio's oeuvre is relatively small, his career compressed into a tumultuous decade, and his influence enormous. Every potential addition or subtraction is a high-stakes reckoning that reverberates through art history and the market alike.

Art historians have argued for decades over which versions of The Lute Player can plausibly be attributed to Caravaggio himself, and which should be understood as workshop products, later copies or imitations. Their disagreements hinge on subtle matters: the handling of light on skin, the confidence of contour, the modelling of hands, the psychological presence of the figure – the same visual evidence yielding different conclusions.

As reported in The Guardian in September 2025, three versions of The Lute Player were tested by the Swiss firm Art Recognition, using their AI image-analysis system: one in the Hermitage in St Petersburg (not in doubt, and believed by almost all scholars to be by Caravaggio); another once owned by Badminton House in Gloucestershire (now called ex-Badminton, and previously dismissed as a derivative); and a third associated with the Wildenstein collection in Paris (which long divided scholars too). The AI testing was conclusive: it said the Hermitage was authentic, as was the ex-Badminton, but the Wildenstein was not. This conclusion matched the convictions of some scholars perfectly. It contradicted others. This debate continues, but when you have firm, objective, science-driven results, it makes a strong argument for one side, and one that is far more difficult to argue against if you're of the other opinion. Scholars in all fields – institutions, as well – tend not to like to have to change their mind publicly once an opinion has been declared, and so some get entrenched, even in the face of new evidence that contradicts them.

AI didn't 'prove' the painting was by Vermeer. It provided a probability assessment based on measurable similarity

This dynamic unfolded around Johannes Vermeer's Girl with a Flute. Vermeer's surviving oeuvre numbers fewer than 40 works, making each attribution enormously consequential. The painting in question has divided opinion for years. In 2022, the National Gallery of Art in Washington, DC concluded that it was not by Vermeer, describing its execution as heavy-handed and lacking the master's intuitive control. The Rijksmuseum in Amsterdam disagreed, asserting that the doubts dissipated upon close viewing. This was not a casual disagreement. Both institutions are among the most respected in the world. Both employed teams of specialists. Both knew their conclusions carried enormous financial and scholarly implications. To complicate matters further, the painting belonged to the museum in Washington, DC, meaning that accepting a non-Vermeer attribution would mean voluntarily devaluing one of its own holdings.

Girl with a Flute (c 1669-75), from the studio of Johannes Vermeer. Courtesy the National Gallery of Art, Washington, DC

AI stepped into the standoff, once again in the form of Art Recognition. As reported in The Times, Art Recognition examined Girl with a Flute, comparing it with verified Vermeers at multiple levels of granularity. The system analysed brushwork, colour variation, compositional structure, and micro-patterns invisible to the naked eye. Its conclusion aligned with the Rijksmuseum's position: the painting fell squarely within Vermeer's statistical signature.

Carina Popovici, the theoretical physicist who founded Art Recognition after witnessing the failures of traditional authentication, was careful not to overstate the result. The AI did not 'prove' the painting was by Vermeer. It provided a probability assessment based on measurable similarity. Still, that assessment carried weight. Popovici has been clear about what her technology can and cannot do. 'The experts were all wrong [in the 2011 Beltracchi forgery scandal],' she told The Times, referring to the German forger who fooled leading scholars and sold fakes worth tens of millions. 'There was clearly something wrong with the traditional process of authenticating art.' Popovici's aim was to give experts a new kind of evidence, one that could challenge entrenched assumptions, without regard to authority.

This distinction is crucial. AI image-analysis does not claim certainty. It might conclude that a painting is 93 per cent similar to an artist's known works, or 97 per cent dissimilar (a 3 per cent match). Those numbers do not settle debates on their own. But they change the burden of proof. A scholar arguing against a high-probability match must explain why. A seller insisting on an attribution that is contradicted by analysis must justify their optimism.

Understanding how these systems work helps explain why they provoke such strong reactions. Despite popular caricatures, AI image analysis does not 'learn style' in any holistic or aesthetic sense. It does not recognise beauty, emotion or meaning. Instead, it decomposes images into quantifiable features. Stroke direction, pressure patterns, pigment distribution, spatial relationships, and compositional ratios are translated into data. Over hundreds of works, statistical regularities emerge. The systems are trained on corpora assembled through decades of human scholarship. They also incorporate known forgeries and misattributions, allowing the algorithm to learn what does not belong. Crucially, reputable firms maintain transparency about their training sets and methodologies, precisely to avoid the circularity that critics fear.

The concern that AI merely reproduces existing opinions is understandable but misplaced. While the training data reflects scholarly consensus, the analysis itself operates independently. It does not know which works are prestigious or valuable. It does not care whether an attribution has stood for a century. It simply measures similarity. When its conclusions diverge from the consensus, they do so without deference. This independence is both AI's strength and its provocation. It strips away many of the social and economic pressures that shape human judgment and exposes how often attribution rests on habit rather than evidence.

For all its promise

For all its promise, AI image-analysis remains reactive. Think of it as a detective called in after disputes have already arisen, and tasked with untangling the past. But a parallel development points toward a more proactive future in which attribution and authentication are anchored at the moment of creation. Craig Follett, co-founder of Peggy, an AI-powered online marketplace and social platform for contemporary art, describes this shift as a move 'from an era of "Trust Me" to an era of "Verify Me".' Peggy's technology enables living artists to scan the unique physical topography of their paintings at multiple focal lengths while verifying their identity. The result is a digital fingerprint of the artwork, capturing minute surface features that are impossible to replicate exactly.

As Follett explained it to me: 'Forensic AI tells you how similar a work is to an artist's broader corpus. Our technology confirms that this is that specific one-of-one canvas the artist held in their hands.' In other words, while Art Recognition and similar firms ask whether a painting looks like it belongs to an artist, Peggy asks whether it is the exact object it claims to be.

The implications are profound. If such technology were widely adopted, future disputes like those surrounding Vermeer or Caravaggio would be far less likely to arise for contemporary artists. Forgery rings would struggle to gain traction. Provenance gaps would shrink. Markets that operate at lower price points could gain the kind of trust historically reserved for blue-chip works. This does not render traditional methods obsolete. Provenance research remains essential, particularly in cases involving looted or stolen art. Material forensics will always be necessary to establish age and composition. Human connoisseurship is still irreplaceable when interpreting meaning, context and intention. What AI offers is an additional axis of information, one that can signal where deeper investigation is warranted and where it may be unnecessary. What it excels in, however, is attribution.

The market stands to benefit from publicly showing that the works they have for sale are by who they claim they are

One of the benefits of properly prepared AI image-analysis systems is that the algorithm is taught via a closed dataset. This is a key aspect that those in the art world who fear AI tend not to understand. The AI is not drawing information from everywhere on the internet. It is carefully curated by a team of thoughtful humans, and the basis of the data set is the catalogue raisonné of the artist against whom the work in question is being tested. The catalogue raisonné is the definitive catalogue of all known works by an artist, assembled by a team of leading human art historians specialising in the artist in question, and including, whenever available, information from forensic tests, as well as listed provenance. It should contain the best information gathered by humans in terms of connoisseurship, provenance and forensics. That's how the AI learns how to recognise an artist's hand – but it looks with a level of microscopic detail and cool objectivity that human experts cannot attain. It then offers up a probability, not a yes or a no, that is interpreted by humans. It is simply the newest and most effective complementary tool to the three traditional branches of art authentication and attribution.

Firms that hop on the AI bandwagon and use open data sets, or don't use the catalogue raisonné, or don't supplement the catalogue raisonné with additional material (like negative data sets – for example, images of derivate works or known forgeries – to teach the AI what not to be fooled by), risk providing partial or misleading information and eroding confidence in AI as a tool. That is why I helped co-author an open paper in association with Art Recognition and the Center for Art Law, laying out the need to establish protocols for using AI responsibly.

At this point in time, museums already approximate the four-pronged model of assessment that constitutes best practice: connoisseurship, forensic science, provenance research and AI analysis, operating in concert. Four strong legs to stand on. AI is still so new that it is only sporadically included – yet, when it has been used, groundbreaking results have often come to the fore, as with the Lute Player and Girl with a Flute. Buyers frequently assume that exhaustive checks have been performed when they have not. The truth is that the market applies such scrutiny selectively. This is ironic because the market stands to benefit most of all from publicly showing that the works they have for sale are by who they claim they are. It's just a matter of time before the market realises this and AI analysis becomes a standard expectation for any expensive work put up for sale. The sooner buyers demand it, the sooner the market will provide it. This will help curb forgeries and make attributions feel more certain, therefore reassuring buyers that the price they are paying is appropriate to the work they're acquiring.

A market that resists accuracy undermines its own credibility. One that embraces transparency builds trust

What this vision leaves unspoken is a deeper anxiety. The art market has always thrived on a productive imbalance between knowledge and ignorance. Risk is not an unfortunate side-effect, it may rather be a subconscious part of the appeal. Like the stock market, the art market is fuelled by volatility, speculation and the tantalising possibility of being right when others are wrong. You see a painting at a flea market. It looks intriguing, perhaps you recognise the hand of a famous artist. You buy it, approaching the seller casually, hoping he doesn't see how interested you really are. You take it home, study it further – it turns out it is by a great artist and there's huge potential profit to be made. To remove too much uncertainty is to threaten the drama that animates the system. A fully verified market may be more transparent, but it may also feel flatter, less electric – though, speaking as a professor specialising in art forgery, I would prefer a verified market.

One might assume that any resistance to AI would come from the sellers. But Art Recognition tells me that 80 per cent of their clients are sellers – gallerists, auction houses, or individuals with art they wish to sell. In fact, having spoken to many people about AI, I can see that those most resistant tend to be experts (who fear that their long monopoly on attribution, and therefore their jobs, will be replaced by technology) and sellers who worry that AI will downgrade attributions as well as upgrade them. But resistance is also short-sighted, since a market that resists accuracy undermines its own credibility, while a market that embraces transparency builds trust.

The first auction house to announce that every painting above a certain price threshold has undergone AI image analysis would change expectations overnight. Buyers will demand it. Sellers with legitimate works will welcome it. Those with questionable ones will not. Over time, the deterrent effect alone may prove transformative.

The deeper significance of AI, however, lies beyond fraud prevention alone. It forces the art world to confront a long-standing contradiction at its core: its simultaneous desire for certainty and risk. Attribution has always been provisional, shaped by human fallibility, incentive and hope. AI does not eliminate that tension, but it shifts it. By introducing a form of evidence immune to reputation and desire, it makes uncertainty harder to disguise and easier to quantify: it compels the market to decide which kind of uncertainty it is willing to live with.

Ultimately, the question is not whether machines can judge art. They cannot, not in any human sense, and not alone. The question is whether we are willing to renegotiate the role that uncertainty plays in a system built on belief. AI can't abolish risk; it merely redistributes it. It asks who benefits from not knowing, who benefits from greater clarity, and how much ambiguity the market truly needs to function. In a world where myths can thrive simply because no one thought to ask the right question, AI makes those questions unavoidable. What emerges is not absolute certainty, but a different basis for trust, one that acknowledges uncertainty openly rather than quietly trading on it.

Disclosure: Noah Charney is an art history advisor to Art Recognition.