AI learns to spot art fakes in five simple tests

Can a machine spot if a drawing really looks like it was made by a specific artist? 

Artificial intelligence can now detect whether a drawing is genuine, or a fake, by analysing how it has been made, using just five key measurements and achieving accuracy of almost 90 per cent. 

The research, conducted by Hassan Ugail of the University of Bradford (UK) and Jan Ritch-Frel, and Irina Matuzava of the Independent Media Institute (US) and David Stork from Stanford University (US) tested the system on 900 authenticated drawings by ten artists spanning five centuries, from Michelangelo and Raphael to Whistler and Constable. 

The paper, published in PLOS One, shows how a new generation of AI can identify subtle stylistic signatures in historical sketches, even when only small amounts of data are available. 

The study sourced materials from the Metropolitan Museum of Art, the Morgan Library, the Ashmolean, the Royal Collection Trust, the Victoria and Albert Museum, and the Casa Buonarroti.  

It comes just weeks after Professor Ugail made international headlines with his groundbreaking uncovering of a possibly new portrait of Tudor Queen Ann Boleyn, wiith historian Karen L Davies.

Above: Life study sketch by Michelangelo, one of the pieces used in the study.

AI art detective  

The system achieved an overall accuracy of 89.8%, correctly accepting genuine works 83% of the time while rejecting impostors at a rate above 90%. 

The system examines features such as line structure, texture, contrast and tonal variation to build a unique “style signature” for each artist. It can then compare new drawings against that signature and flag anything that does not belong. 

In tests across hundreds of drawings by major historical artists, the method achieved strong accuracy while keeping false identifications low. 

Above: Working from a small replica of Michelangelo's statue of "Day" on the tomb of Giuliano de' Medici in the New Sacristy, San Lorenzo, Florence, Jacopo Tintoretto has copied on both recto and verso the back view of the figure.

Visual fingerprint 

Authenticating drawings has long relied on expert judgement, with specialists studying minute stylistic details to attribute works to an artist. 

But historical sketches present a particular challenge, as only small numbers of verified examples are usually available. 

The Bradford team tackled this by developing an AI model that learns what ‘normal’ looks like for a specific artist, rather than trying to compare against thousands of alternatives. 

Professor Ugail said: “Every artist leaves behind a visual fingerprint in how they draw. What we have developed is a way of capturing that fingerprint in a mathematical form and using it to assess whether a new work is likely to be genuine or not. 

“This does not replace expert judgement, but it gives a robust, reproducible layer of evidence that can support decision-making.” 

Smaller data, smarter AI 

Unlike many modern AI systems, which rely on vast datasets, the Bradford approach is designed to work with limited material, reflecting the reality of art history. 

The model can learn from as few as 20 authenticated drawings per artist, making it practical for real-world use, where data is scarce. 

In trials involving 900 authentication decisions, the system achieved a balanced accuracy of nearly 88 per cent, with a very low false acceptance rate. 

This means it is particularly strong at avoiding the most damaging error in art authentication, recognising a fake as genuine. 

Why this matters 

The implications extend beyond the art world. 

Accurate authentication is critical not only for museums and collectors, but also for legal disputes, insurance claims and the global art market, where questions of attribution can involve millions of pounds. 

The research also demonstrates a broader point about artificial intelligence. 

While large, general-purpose AI models often dominate headlines, this study shows that carefully designed systems, tailored to specific problems, can outperform larger models when data is limited. 

In fact, widely used deep learning models performed significantly worse in this task, rejecting many genuine works despite producing fewer false positives. 

Above: Professor Hassan Ugail, University of Bradford.

From Renaissance sketches to modern AI 

The system was tested on works by ten artists spanning several centuries, including Michelangelo and Raphael, using images from major international collections. 

By identifying patterns that align with art historical knowledge, such as similarities between closely related artists, the AI also offers new insights into how styles overlap and evolve. 

What comes next 

Researchers say the technology could be expanded to include additional visual features, as well as other forms of evidence such as material analysis and provenance data. 

Professor Ugail added: “The goal is not to replace connoisseurship, but to strengthen it. By combining human expertise with computational methods, we can move towards more transparent and reliable approaches to authentication.” 

It Starts in Bradford

It starts in Bradford, where ideas become impact, from AI that can detect fake artworks to research that tackles real-world challenges. By turning ambition into discovery and insight into application, our researchers are shaping global conversations and creating tools that make a difference far beyond the lab.

Notes to editors 

• The research is published in PLOS ONE 

• The study analysed drawings by 10 historical artists using images from major open-access collections 

• The system uses five key visual features to assess authenticity, including texture, contrast and structural complexity