How our AI
authenticates art
Art Recognition’s AI is built by our team for one purpose: to provide a transparent, documented assessment of stylistic consistency. The method is documented, the reasoning is traceable, and the limits are named.
A bespoke AI, built
for one question
Our AI is engineered by our team for art authentication alone. Not a general-purpose model adapted to art, but a dedicated system that learns what makes an artist’s work unique and distinguishes it from collaborators, followers, and imitators to authenticate submitted artworks.
Public AI models train on data scraped from the web, which inevitably contain errors and unverified claims. We don’t. Our training data is scholarly, curated by art historians, and documented end to end.
The early development of our technology was supported by funding from the European Union’s Horizon 2020 research and innovation program and carried out in collaboration with academic partners. Its methodology has since been published in peer-reviewed journals together with scientists from Tilburg University in the Netherlands and the University of Liverpool in the UK.
This is the fourth pillar of authentication, alongside connoisseurship, provenance, and forensic analysis: systematic visual analysis at a scale no eye can hold at once.
Our analysis process, in four steps
Every analysis follows the same steps, in order: data collection and curation, AI model training, AI model testing, and results. Built and carried out by our team.
Data collection and curation
For each artist, our art historians build a training corpus from the scholarship the field already recognizes: catalogue raisonnés, exhibition records, accepted attributions.
Alongside the authentic works, we add a contrast set: documented forgeries, imitators, followers, atelier works, and digital forgeries created by generative AI. The model learns not only what the artist’s work looks like, but what distinguishes it from closely related works.
Images are augmented for contrast and color, helping the AI models to recognize stylistic characteristics regardless of camera or lighting condition. Our coverage spans from old masters to modern artists. Some of our datasets are shared publicly, including the Raphael dataset, widely used by researchers across the field.
Training the AI
Training takes place in two stages. First, the model is pre-trained on a broad dataset spanning the history of art, giving it a general understanding of paintings. It is then trained on a specific artist’s body of work and a carefully curated contrast set to learn what makes that artist’s style unique.
Each artist has a dedicated AI model, built on the same underlying architecture. The architecture combines two types of neural networks — a Convolutional Neural Network and a Vision Transformer — which are trained together to learn the artist’s distinctive style. Working in combination, they learn visual characteristics ranging from fine details such as brushwork, texture, and edge patterns to broader compositional relationships, including motif placement and proportions.
During the training, the model learns from the artworks at different scales at once: from the full images to understand the overall composition, and from smaller image patches to capture fine stylistic details. Where forensic data is available, we incorporate it into the analysis. Restored or repainted sections are accounted for and treated separately. We pioneered the use of Vision Transformers in art authentication.
Testing the AI model
Before the model is used on any new work, it must demonstrate consistent performance. A separate test set, never seen during training, is used to evaluate the model’s performance.
The standard metrics — precision, accuracy, and recall — measure how reliably it distinguishes authentic works from negative examples.
Output
Once the AI model has been trained and tested, it is ready to authenticate a new work. The submitted image is compared against everything the model has learned about the artist. The model returns a probability of stylistic consistency, supported by documented reasoning.
The output flows into one of two deliverables: an AI Evaluation (the result of the analysis, with a Certificate of Authenticity when confirmed), or a comprehensive AI Authenticity Report, documenting the methodology, art historical scholarly background, probability score, features learned, and visualizations.
Detecting AI-generated forgeries
Generative AI can now produce realistic imitations of any artist’s style. Our training data includes such imitations alongside authentic works and human-made forgeries. As a result, our models can identify these works when they are submitted for authentication. Confirmed by our recent published research.
Our scope
The technology works on paintings, drawings, prints, and other works where the artist’s hand is visible in the image. Sculpture, drip painting, street art, and artists with limited visual documentation remain outside what the technology can support. We name these limits in the initial conversation, before any work begins.
Read our academic publications →