Not all AI is Equal: A Response to the de Brécy Tondo

September 2023

AI has been used to try and authenticate the so-called de Brécy Tondo. This article explores the issues associated with the work that has been done.

The so-called de Brécy Tondo is on display at the Cartwright Hall Art Gallery in Bradford
Giacomo, Researcher & Business Development Manager

Interested in authenticating your painting?

Key Takeaways:

  • Research teams from the universities of Nottingham and Bradford have used facial recognition technology to compare the faces of Raphael’s Sistine Madonna (1512) (Gemäldegalerie Alte Meister, Dresden) with the de Brécy Tondo at Bradford Council’s Cartwright Hall Art Gallery. Facial recognition is not an effective means of authenticating an artwork, not least given that any forger or copyist would focus on facial features and structure.
  • Not all AI is equal. Instead of using an unfiltered dataset derived from publicly available, low-resolution images that invariably include forgeries, Hephaestus integrates AI analyses trained from carefully sampled data sets. Using a rigorous suite of scientific tests, provenance research and connoisseurial expertise, Hephaestus produces the most conclusive authenticity results. An artwork cannot be considered authentic unless it is widely accepted by the market.
  • Connoisseurs will always agree and disagree on the attribution of specific artworks, a logic that similarly applies to AI. Without integrating AI into a scientific protocol complemented by provenance research and connoisseurship, the ‘Battle of the AIs’ is a futile project that will never lead to a reliable, academically and market-accepted attribution.

In a recent article discussing the so-called de Brécy Tondo in The Observer entitled ‘Battle of the AIs: rival tech teams clash over who painted ‘Raphael’ in UK Gallery’, the results of research teams in Nottingham and Bradford in machine learning facial recognition technologies were set against that of Art Recognition, the former asserting the tondo’s authenticity and the latter denying it. Facial recognition is not a reliable, accurate or widely accepted means of art authentication, not least because the practice of digitally testing only an isolated region of a picture is a misleading basis from which to draw an artwork’s attribution. Equally problematic is the use of largely unfiltered data sets of published images from the internet to teach machine learning algorithms. While we appreciate that journalists need to attract readers’ attention with provocative headlines, ‘Battle of the AIs’ neglects both the fact that some machine learning architectures may be better at attributing Old Masters and others are better at dealing with Expressionist work and that AI can be used as a tool to aid conclusivity, as opposed to mere speculation.

British art historian Sir Timothy Clifford, director of the National Galleries of Scotland from 1984 to 2006, expressed extreme doubt in the efficacy of AI models as a standalone means of authentication, stating that “I do feel rather strongly that mechanical means of recognising paintings by major artists are incredibly dangerous… I’ve never contemplated the idea of using these AI things. I think they’re terribly unlikely to be remotely accurate. But how fascinating” (1).  Today’s practice of testing only select areas of a work, ascertaining results from an unfiltered dataset and not appropriately substantiating the results of AI analyses with rigorous analogue testing is what drives market scepticism. In the 2018 Wiesbaden forgery scandal involving major examples of Russian avant-garde art, the judge, Ingeborg Baümer-Kurandt aptly stated: “Ask 10 different art historians the same question and you get 10 different answers… behind the experts, there are diverse vested interests influencing how these paintings are evaluated” (2).  Just as a single expert’s opinion does not ensure a watertight attribution, neither would a single AI screening prove sufficient for an attribution.

An artwork cannot be considered authentic unless it is widely accepted by the market. In this vein, we believe that AI technology, although a powerful tool for pre-screening and conclusive purposes is unlikely to be able to sway leading art market participants without being integrated into scientific analyses. As illustrated by both the Observer article ‘Battle of the AIs’ and Art Recognition’s unrequested AI testing of Rubens’s Samson and Delilah at the National Gallery, solely testing with AI is more likely to lead to further disputes than consensus and acceptance in the art market.

Machine learning technology is developing rapidly, especially in the context of art authentication, and, therefore, it is vital to recognise the key role that traditional testing methodologies play in the authentication process, not least substantiating the results of digital AI analyses. At Hephaestus, results of analogue scientific tests, provenance research, connoisseurial expertise and AI analyses all have to line up in order for an authentication to be made. This is a task almost impossible to be reverse engineered, even by the most sophisticated forgers. For that reason, our expertise has been sought out for by law enforcement agencies and museums.

The Observer headline ‘Battle of the AIs’ suggests that all AI art authentication models are of equal merit but, as in any other context, AI models are only as good as the data sets on which they are taught. Given Hephaestus’ ongoing collaborations with major institutions to sample accurate data sets on bona fide authenticated work, we utilise machine learning technology as a means of achieving the highest levels of conclusivity in attribution, not as a standalone gimmick as Clifford’s critique suggests.

(1) Quoted in Dalya Alberge, ‘Battle of the AIs’, The Observer, 9 September 2023 <available at:>.

(2) Quoted in Catherine Hickley, ‘Unknown or Unreal? The Shadow on Some Russian Avant-Garde Art’, The New York Times, April 6, 2018 <available at:>.

Other Posts