Salvator Mundi is not by Leonardo, an AI claims — its 81% verdict has not been peer-reviewed

Few faces in Western painting have been argued over as long as the one at the centre of the Salvator Mundi: a Christ who looks out through a haze of softened shadow, right hand raised in blessing, a crystal orb resting in the left. Whether that haze is Leonardo da Vinci’s own sfumato, a pupil’s patient imitation of it, or the work of the conservators who rebuilt a badly abraded walnut panel has been the real question since the picture resurfaced.

Brazilian researcher Átila Soares da Costa Filho now says software can settle it. In a press release sent to Italian and other European outlets, he reports that his attribution system, Luminari — built on convolutional neural networks trained on an artist’s known work — points to Giovanni Antonio Boltraffio, Leonardo’s Milanese pupil, as the most likely painter.

The Salvator Mundi after cleaning in 2006-07, showing cracks and paint losses — The Salvator Mundi as photographed in 2006–07, after cleaning and before restoration. Image: public domain (Wikimedia Commons)

The method, as the release describes it, works in two stages. To avoid judging a surface repainted by later hands, Soares first had a model generate an “intelligent” reconstruction of the panel as it stood before its successive restorations, using more than 10 billion parameters. A German AI system then compared that face with the one recovered by the restorer Dianne Modestini, finding strong physiognomic compatibility. The attribution test then put Boltraffio’s probability at 81%; the release asserts that a 75% compatibility score is enough to declare a work authentic — a threshold it gives without a source.

Read closely, the claim is less categorical than its headline, which announces flatly that the picture is “not by Leonardo”. The same text says Leonardo’s name came immediately after Boltraffio’s in the system’s ranking of candidates — his score undisclosed — and that the master may have contributed directly or supervised closely. That is not a refutation of the workshop hypothesis; it is the workshop hypothesis. Carmen C. Bambach, curator at the Metropolitan Museum of Art, has argued that much of the painting is Boltraffio’s, with Leonardo responsible for retouching.

Historic photograph of a Salvator Mundi attributed to Giovanni Antonio Boltraffio — A Salvator Mundi attributed to Giovanni Antonio Boltraffio, in a historic photograph. Image: Harvard Fine Arts Library, public domain (Wikimedia Commons)

What is missing is the paper. Soares cites the peer-reviewed journal Conservation Science in Cultural Heritage, published with the University of Bologna, on whose scientific board he sits. The journal did run a Luminari study — on paintings from Rembrandt’s circle, not on the Salvator Mundi. The 81% figure itself has not appeared in a peer-reviewed study, and the French site Unidivers, which received the same claim in July, faulted it for exactly that: no full protocol, no list of training images, no independent replication.

Soares concedes the point in spirit: in the release he calls AI “simply one of the tools” that, combined, bring researchers closer to the truth.

The panel itself, sold at Christie’s in 2017 as the most expensive painting ever auctioned, has not been shown publicly since; the release says it is presumed to be held in Geneva. Its story is also heading to television: Julianne Moore was announced in 2024 to star as Modestini in a drama series adapted from the 2021 documentary The Lost Leonardo.

Until the painting can be seen again — and an algorithm tested in the open — the argument returns to where it began: to that face, and to what a trained eye believes it sees in the shadow around the mouth.

Tags: artificial intelligence, Leonardo da Vinci, art attributions, Átila Soares, Dianne Modestini, Giovanni Antonio Boltraffio