Going broke underestimating the intelligence of forgers.
We note with sadness the passing, on August 11th, of our esteemed colleague, Ernst van de Wetering, admired globally as the “final word in Rembrandt authentication.” De Wetering’s eminence as a connoisseur sometimes distracts from our memory of the beginnings of his career as a scholar who, in 1968, set out to use scientific technologies like X-radiography and examination under ultraviolet light, combined with connoisseurship, to determine a work’s authenticity with greater certainty.
Earlier in the summer, Hephaestus was honored by the University of Bologna to give a keynote address at its conference on “Il falso: fake in the art market.” Alongside some of the world’s leading art experts, our CEO, Denis Moiseev, discussed the detection of art forgery in his talk entitled, “Developing rigorous protocols of authentication: the future of science and AI.”
A major theme of Moiseev’s talk was how current art market participants underestimate the skills of enterprising forgers, and that authenticators are falling behind them. Some market commentators have gone so far as to describe this situation as an “arms race” pitting art emulation against detection. But it’s not really a race. Even though the level of precision and sophistication of scientific tools has increased exponentially since 1968, scientific analysis is rarely used unless a concern about authenticity has already been raised. The tools in current use? Various forms of imaging technologies and pigment analysis. The same tools that were pioneered by de Wetering five decades ago. In light of his passing, it now seems prescient that Moiseev remarked back in June that, “The number of connoisseurs is diminishing, and detection methodologies remain largely unchanged.”
Forgers are quite capable of duping experts and breaking through the defences of the art market. The Ruffini scandal is perhaps the best example of this, for it revealed that despite the scrutiny of connoisseurs and what Moiseev calls the “unscientific application of science”, a series of Old Masters exhibited and sold by some of the most reputable institutions in the world were, in fact, modern forgeries. The Louvre, National Gallery in London, New York’s Metropolitan Museum of Art, the Galleria Nazionale in Parma and Vienna’s Kunsthistorisches Museum were embroiled in the saga, having between them authenticated and borrowed paintings that were sold by Ruffini and later attributed to Orazio Gentileschi, Parmigianino and Frans Hals. They are all forgeries.
The question is, how can we develop protocols that detect forgery today? An answer may arise from comparing how our arms race with forgers compares to the way computers were taught to outwit human intelligence at board games. Some games are harder to crack than others. Why is that? Machines began by beating humans at simple, complete information games: Noughts and Crosses, then checkers, chess and GO. The games became more difficult because of complexity in strategy, moves and rules. Although computational power became so great that it overcame these challenges, the addition of extra moves made things more difficult. Likewise, if additional forms of analysis are applied in a well-thought-out sequence, the scientific protocol becomes more robust.
However, ALL the rules and possible moves of complete information games can be calculated by computers. Beating humans at incomplete information games, such as poker, proved far more difficult. Other than the additional complexity of the unknown, the computer had to compute the possibility of a player bluffing. At present, forgers have access to the same information as authenticators: the power is balanced. By keeping certain information secret or by creating a protocol that has steps that are difficult to decipher the balance of power tips to the authenticator.
Developments in machine learning and artificial intelligence extend the reach of connoisseurs by revealing an artist’s invariant characteristics: the distinctive pressure applied to brushstrokes; the habitual spacing of compositional elements; the proximity of certain pigments to one another. Hephaestus has trained its machine learning tool to identify nuanced differences between strongly associated artists, such as Canaletto and Bellotto, and between autograph works and copies. This kind of digital tool may not only reduce the costs and length of time involved in traditional analysis (by eliminating possibilities) but may also provide connoisseurs with the scientific vocabulary to explain their intuitions.
Solving the problem of forgery is complex, but new technologies and their scientific application may help us eradicate forgery from the art markets once and for all. For more information about our machine learning tools please write email@example.com