Up until recently, the main objective of the artworks digitalization was to expand their accessibility and exploration whatever the place and time. Now, the use of machine learning has brought new ways of analysing art through typical methods such as close reading and distant viewing. While close reading approaches like artist authentication or art valuation involve the analysis of one art piece, distant viewing approaches like gallery curation or art style identification include large collections.

Art Valuation

Since art appraisal is a self-regulated field, there is no public formula to evaluate an artwork. While auction houses rely on human expertise to provide an estimate, some companies are using financial data and market trends to provide technology-driven appraisals. They use a combination of image recognition and textual analysis to examine artworks, comparing them against multiple historical data to appraise their value and significance. For instance, Appraisal Bureau – an analytics company specialized in the valuation of art pieces, has built an AI platform that provides neutral appraisals. They have fostered relationships with galleries, getting access to a large volume of private sales data which helps their algorithms appraise works more quickly and accurately. However, their data model is not meant to replace human judgment but to enhance it since valuing artworks takes into account a wide variety of factors, some indeed in conflict with each other.

Artwork Restoration

In 2019, a team of scientists from the Rijksmuseum in Amsterdam used machine learning to recreate missing pieces of Rembrandt’s original form of The Night Watch (1642). The painting was trimmed on all four sides to fit its new position when it was moved to the town hall in 1715. The scientists used a copy of the original artwork made by Gerrit Lundens around the same period as the basis for the digital reconstruction. Using three distinct neural networks, they mapped out visual matching points across both paintings, removed spatial distortion on the resulting digital overlay, and restored the added parts in the style of Rembrandt. Although the reconstruction was successful, some might think that the technology couldn’t mimic the genius of Rembrandt into the extended painting. Indeed, the scientists decided to display temporarily the printouts of the extended composition, mounting them in front of the original painting instead of flushing them with it.

Artist Authentication

Verifying the authenticity of artists using machine learning is one of the most popular use cases for the technology in the art world. Indeed, algorithms can provide reliable and verifiable assessments, reducing the risk of human errors and biases. Some statistical algorithms such as the ones built by Hephaestus Analytical – a company that uses artificial neural networks to analyze paintings, improve the accuracy of authentication based on a variety of parameters such as the spacing of strokes, the thickness of a paint layer, and the recurrence of objects or patterns in the artist’s oeuvre. They extract these characteristics for each artist, assigning each one a value that measures its importance in the artist’s style. This creates a complex polygon to which its overlapping with the statistics of another painting can determine the likelihood that it may be attributed to a particular artist. That process can also help actively addressing art forgers who are well-versed in their techniques, challenging traditional art authentication methods.

Art Style Identification

Art style is a set of specific elements which can be associated with a particular artistic movement, school or time period. Although that concept is essential to classify artworks, it’s still a challenging problem to correctly identify a style in a fully automatic way. Art experts typically spend years developing the skills required to recognize artistic styles. Some of the challenges include ambiguous interpretability of abstract elements, nuances separating different artistic categories, blurred lines between art periods, existence of artistic attributes that belong to multiple styles or do not belong to any style, and variations in the style of the same artist. In recent years, many data scientists took up the challenge by developing several resolution approaches based on deep learning. They trained convolutional neural network models on significant datasets, which contain millions of annotated images. The pre-trained models learned to automatically identify visual patterns, enabling them to capture complex style characteristics. Alternatively, these models extracted relevant features directly from the artwork images, eliminating the need for handcrafted labels, and improving the accuracy and precision of the process.

Gallery Curation

Curating an exhibition often involves deliberately adopting a particular concept or perspective to shine a new light on a set of artworks. Since museums and art galleries have extensive collections, machine learning can help curators identify similarities between artworks, art styles, and art movements based on artwork parameters. Last year, the Nasher Museum of Art located in Durham, North Carolina, used ChatGPT to set up an AI-generated exhibition. Their team of curators built a dataset of 14,000 art pieces from the museum’s permanent collection. They trained that dataset using a custom version of ChatGPT and developed a series of prompts allowing the system to select and group the artworks by themes. They generated the exhibition texts, artwork labels, and a sequence in which to present the works although the tool was unable to create a floor plan of the gallery. The curator also acknowledged that the technology was not neutral and eventually filled with inherent biases, which left room for improvement. In the same way, other curators have experimented how machine learning might offer new agential perspectives in the organisation of large scale exhibitions such as biennials.