From contested attributions to computational evidence, Dr Carina Popovici examines how artificial intelligence is reshaping the standards of authenticity.
A theoretical physicist by training, Dr Carina Popovici is the CEO and co-founder of Art Recognition AG, a Swiss company applying machine learning to fine art authentication. Based in Adliswil, near Zurich, the firm develops models designed to identify artist-specific visual signatures — micro-textures, brushstroke structures, and stylistic regularities often imperceptible to the human eye.
Popovici holds a PhD in Theoretical Particle Physics from the University of Tübingen (2011), and a Bachelor’s and Master’s in Physics from the University of Bucharest, where she graduated first in her class. Before establishing Art Recognition in 2019 with Christiane Hoppe-Oehl, she worked in quantitative research and risk-analysis roles in Swiss finance and technology. Alongside her role as CEO, she continues to publish research on computer-vision models, including work on vision transformers and synthetic datasets for authorship analysis and forgery detection.
She is a regular contributor to discussions on AI, cultural heritage, and museum standards, speaking at international conferences on art law, technology, and ethics, and advocating for greater transparency and methodological rigour in computational tools used for attribution. She has appeared on Swiss National Television (SRF), delivered a TEDx talk on AI and authorship analysis, and participated in documentaries and institutional programmes, including those hosted by Smithsonian Associates.
Art Recognition’s methodology combines convolutional neural networks, explainability tools, and rigorously curated training datasets developed in collaboration with art historians and technical specialists. The aim is not to replace connoisseurship but to provide reproducible, independent evidence that applies consistent criteria across large bodies of work. The company has analysed thousands of paintings spanning Old Masters, modern, and post-war art, with findings contributing to attribution debates involving Rembrandt, Van Gogh, Picasso, Modigliani, Raphael, and Rubens.
Among the company’s most discussed studies is its analysis of Rembrandt’s The Polish Rider, reported by The Art Newspaper, which demonstrated the system’s ability to distinguish passages executed by different hands within a single painting. Other notable cases include the Flaget Madonna, where the faces of the Virgin and Christ Child were attributed to Raphael with a reported probability exceeding 96 per cent; the de Brécy Tondo Madonna, judged with an 85 per cent probability of not being by Raphael; and Anthony van Dyck’s Portrait of Don Felipe de Guzmán, assessed with a 79 per cent probability of not being by Van Dyck.
de Brécy Tondo Madonna — AI analysis by Art Recognition indicated an 85 per cent probability the work is not by Raphael.
As private treaty sales and cross-border transactions expand, and due diligence expectations increase, Art Recognition’s reports have been used by collectors, museums, insurers, and legal professionals, and have been submitted as evidence in catalogue raisonné processes and attribution disputes.
In a rapidly evolving art market, Popovici’s work reflects the growing use of computational analysis alongside traditional connoisseurship to determine authenticity.
In conversation with The Exclusivist, she reflects on the practical and ethical implications of using artificial intelligence in attribution today.
In Conversation with Dr Popovici
AI vs Connoisseur Authority
When AI analysis diverges from a leading connoisseur or a catalogue raisonné committee, how does the market tend to resolve that tension — and what does that mean for collectors whose works depend on traditional attribution authority?
When AI analysis diverges from the opinion of a leading connoisseur or a catalogue raisonné committee, the first reaction is often to assume that the expert must be right and the AI must be wrong. The immediate impulse is then to ask why the AI has produced a different result and how the system might be adjusted so that the two conclusions align. This reaction shows how deeply traditional expert authority remains embedded in the art market. However, attribution disputes are rarely a simple matter of deciding who is right and who is wrong.
We often encounter cases in which the same painting has been assessed by several experts who have reached different conclusions. In such situations, an AI result will inevitably differ from at least one of those opinions. Rather than replacing expertise, AI can serve as an independent, data-driven source of evidence that helps bring additional clarity to attribution disputes. Unlike traditional connoisseurship, which may vary depending on the expert and the criteria they prioritise, AI analysis applies a consistent methodological framework. For collectors, this additional perspective can provide greater clarity, confidence, and transparency in the decision-making process.
I would not say that collectors depend only on traditional attribution authority. In some cases, traditional expertise can leave a work effectively blocked, especially when competing expert opinions prevent a clear market position from emerging. For collectors, the essential need is a clear, objective, and well-documented assessment. AI authenticity analysis can contribute to this by offering a transparent and structured evaluation. The results are data-driven and accompanied by comprehensive reports explaining how the datasets were created, which art-historical literature was used, how the trained AI models performed, and which visual features contributed to the AI’s assessment.
Market Repricing Risk
If AI begins to reveal systemic misattributions within certain artists’ markets — for example, distinguishing between an artist’s hand, studio production, or later additions — which segments of the market do you believe are most vulnerable to repricing, and how should collectors think about that risk when acquiring today?
I would not necessarily describe this as a risk. In many cases, it should be seen as an advantage. If AI can help distinguish between the hand of the artist, studio participation, later additions or restoration, it can bring a level of clarity that was previously unavailable. This is valuable not only from an art-historical perspective, but also from a market perspective.
A good example is our work on Rembrandt’s The Polish Rider, discussed by Noah Charney in an article published by The Art Newspaper. This was a painting around which questions had existed for decades: which parts could be associated with Rembrandt’s own hand, and which parts might have been executed by someone else? AI analysis allowed us to approach that question in a new and more precise way.
For collectors, this kind of information should not automatically be understood as diminishing value. On the contrary, it can increase confidence by making the attribution more transparent. If a painting is described as “artist and workshop” or “studio of”, the market often has to rely on broad categories. But if one can quantify which passages are by the master, which belong to the workshop, and which may be later interventions, the work can be assessed more precisely. Greater knowledge should ultimately be beneficial and help establish a more accurate and defensible market position.
When acquiring works of art today, collectors should view AI as an integral part of the due diligence process. Alongside human connoisseurship, provenance research and technical analysis, AI offers an additional source of evidence that can help collectors make more informed decisions and navigate attribution uncertainty with greater confidence.
Invisible Fraud Channels
Beyond forged provenance and AI-generated documentation, what emerging fraud methods concern you most? In particular, are there vulnerabilities within private treaty sales, estate dispersals, or cross-border transactions that the wider market has yet to fully recognise?
Beyond forged provenance and AI-generated documentation, I am particularly concerned about the growing sophistication with which questionable works are introduced into the market. Fraud today is often less about a single forged document and more about the construction of an apparently credible narrative around an artwork.
Private treaty sales, estate dispersals and cross-border transactions are an integral part of the art market, but they are often less transparent than public auction sales. They may involve less public scrutiny and fewer opportunities for independent verification, which creates vulnerabilities that the wider market should take seriously.
One risk that I believe remains underappreciated is the tendency to rely on reputation and trust alone when transactions occur outside highly visible market channels. The more private a transaction is, the more important independent due diligence becomes.
In my view, the solution is not to treat these segments of the market with suspicion but to introduce stronger and more consistent due diligence standards across the industry, regardless of the type of transaction. Transparency, objective analysis, provenance research, technical examination and independent verification should be standard practice. As the market becomes increasingly global and digital, collectors will benefit from relying on evidence rather than assumptions.
AI-Premium Works
As technical analysis becomes more sophisticated — from AI-driven pattern recognition to advanced imaging — do you anticipate a future in which works supported by robust, data-based verification command a pricing premium? If so, how should collectors think about positioning their holdings for a market that increasingly rewards evidentiary certainty?
I think there are several aspects to consider. On the one hand, AI makes the authentication process more accessible: the client can begin by submitting a high-resolution photograph, and the AI model itself is highly scalable while maintaining rigorous scientific and scholarly standards.
On the other hand, the process is not purely automated. It also depends on a highly skilled team. Art historians carry out the background research needed to compile and curate the training datasets; AI developers continuously improve the AI models and process the results; and our experienced client relationship team guides clients through the process and ensures that the findings are communicated clearly. In that sense, our work at Art Recognition combines the power of technology with a high-quality human service. Even though we are an AI company, the value we deliver lies not only in the algorithm but also in the expertise, research, interpretation and support that surround it. The premium should therefore not be attached simply to the use of AI itself but to the quality of the evidence and service delivered to the client.
Collectors should think about positioning their holdings through strong documentation, clear provenance, material study where appropriate and independent AI-based analysis. In my opinion, the market will increasingly value not only the attribution itself but how convincingly that attribution can be supported.
