Sotheby’s and machine learning for the arts

Can machines appraise art? How about creating it?

How much is that painting worth? Why is your painting not worth a dime whilst Leonardo’s Salvator Mundi sells for $450 million? [9] It can depend on the reputation of the artist, on the age of the painting, on its subject, but also on many many many other factors. Appraising artwork is a tough and very subjective job, where art experts examine and balance all of these factors. The expert eye of the art connoisseur is, so far, the only way that we have to value artworks that can fetch millions and millions of dollars on auctions.

Some startups have tried to come up with algorithms to predict potential auction prices, but predictions rates haven’t been much better than guessing [1]. Valuation of art is an art (pun intended), and hasn’t changed very much in a very long time. The market is ripe for disruption.

Auction houses are not standing by. Recently Sotheby’s acquired Thread Genius [3] a startup that provides algorithms to identify objects and recommend similar objects to the user. Thread Genius is fundamentally a visual search engine that relies on convolutional neural networks [4] to analyze patterns, shapes, colors, strokes that can be trained on specific purposes: even if it started with clothes recognition for the fashion industry, its deep learning capabilities were soon translated to paintings.
Thread Genius is the first step in systematizing the appraisal method: by finding comparable artworks, appraisers can better benchmark their estimates, but it’s a short-term initiative. The lack of long-term disruptive usage of machine learning creates an opportunity for large auction houses and art dealers/galleries to innovate and disrupt.

My recommendation to Sotheby’s is to stay ahead of the game in deep learning. Competitors are most likely going to jump on the bandwagon of machine learning and, despite its smart move with the Thread Genius acquisition, Sotheby’s can’t afford being an old fashioned auction house in the world of machine learning. Sotheby’s has access to an enormous database of past transaction onto which it can train the most advanced feed-forward machine learning algorithms. Not only it has the potential to become the Netflix of the arts, suggesting art pieces that you might like based on your past preferences, but it needs to strive to understand the potential value of artwork and eventually maximize the capture of value by optimizing its pricing.

The next episode for artists, galleries and dealers will come when deep learning will get so good to actually generate the art itself, to come up with something entirely novel and revolutionary, surprising humans and triggering new trends and new possibilities.

It seems so futuristic but it is happening already. Sotheby’s competitor Christie’s recently sold the “Portrait of Edmond Belamy” for $432,500 [2]. The portrait was generated using generative adversarial neural networks, a class of algorithms that can create so realistic looking images that can be difficult to discern from a human’s work or so realistic even realistic videos for events that have never actually happened (deepfakes).

Portrait of Edmond Belamy, 2018, created by GAN (Generative Adversarial Network). Sold for $432,500 on 25 October at Christie’s in New York. [2]

Artistic product development as a whole will be impacted by the change, Google is already experimenting with computer generated music and went as far as creating an entirely new and AI-designed musical instrument [8] whose sound doesn’t resemble anything that we have build before. IBM is betting that AI can generate new scents and fragrances [7], an art that so far had been a privilege of few good noses.

Everything we consider art can be impacted by the tide of machine learning, and incumbents need to be ready to embrace it as an opportunity to create, appraise and trade new art, maybe in conjunction with human judgement, but maybe, who knows, even without it!

Machine learning and AI is reshaping every industry and art might be one of the toughest to disrupt but eventually the wave of change will get there too, and even in this whole push towards a more digital art world some questions remain unanswered.

Will a computer ever be able to appreciate an value the artistic breakthrough of Marcel Duchamp’s fountain? Or Lucio Fontana’s cut canvas?
Will machine learning bring increased price transparency and subsequently increased liquidity into the art market?
If machine learning were to become very good at pricing art pieces, will we still need auction houses? Or will they transform into stock exchanges?
As deepfakes become more and more realistic and widespread, will the value of real authentic art become jeopardized?

(731 words)


[1] “Artificial Intelligence For Art Valuation – A Review By Artmarketguru”. 2018. Artmarketguru.

[2] “Is Artificial Intelligence Set To Become Art’S Next Medium? | Christie’s”. 2018. Christies.Com.

[3] “Sotheby’S Acquires Thread Genius To Build Its Image Recognition And Recommendation Tech”. 2018. Techcrunch.

[4] “Art Genius – Sotheby’s – Medium”. 2018. Medium.

[5] “How Will Algorithms Change The Art Market? | Financial Times”. 2018. Ft.Com.

[6] “Art Market Ripe For Disruption By Algorithms | Financial Times”. 2018. Ft.Com.

[7] “Forget About Chanel No. 5. IBM Is Now Making Perfume Using AI.”. 2018. Vox.

[8] “Nsynth Super”. 2018. Nsynthsuper.Withgoogle.Com.

[9] “Value Soars For Leonardo Da Vinci Drawing After ‘Salvator Mundi’”. 2018. Nytimes.Com.


Under Pressure: How machine learning can help tackle leakage in European water distribution


Walking on the Sidewalk: Building the future of cities in Toronto

Student comments on Sotheby’s and machine learning for the arts

  1. Thank you for challenging my understanding of the arts. I found this topic extremely interesting as it questions what it means to be creative.

    I appreciate that Sotheby acquired Thread Genius to stay ahead of a disruptive trend, but I remain skeptical about the long-term viability of machine learning in the arts. At the heart of it, art pieces are often limited in number with valuations subjected to shifting preferences and economic cycles. Art is only worth what someone is willing to pay for it in the moment, as seen in the significant price variation that occurs between pieces in an artist’s collection.

    With limited number of sales and thus data, I’m concerned that the data will not be relevant enough or representative enough for machine learning. The machine risks valuing pieces on outdated sales or drawing comparisons between dissimilar pieces of art that share similar features. There’s also a subjective element to the valuation, as part of the price stems from an emotional connection to the work. These subjective elements are hard to control.

    Additionally, as someone who has choreographed some of their best dance pieces randomly walking on the sidewalk or dealing with a crisis, creativity can’t always be timed. It tells a story and provokes emotion. Without these human experiences, I question the machine’s ability to showcase the human condition, the key ingredient to timeless art.

  2. Thank you DigitalGansta! Much of how art is valued is tied to the identity of the artist himself/herself. We have historically seen how a prominent art collector buying up work from a promising young artist can increase prices, as can the death of an accomplished artist. How will technologies that evaluate a work of art purely on its aesthetic change the way the art industry values pieces? Similarly to your reference of Duchamp, how will pieces that change the history of art be valued?

    In response to your research on machines creating art, I find it difficult to imagine such works really being of value. We as emotional beings have the greatest experiences with art when a painting or a melody evokes an emotion that connects on a soul level. Machines create output based on inputs – but some of the greatest pieces of art (music, poetry, paintings, etc.) come from the heart of their creators. Machines are getting better at mimicking the human brain, but how could a machine match the emotional depth of a true artist? Would we want to live in a world where that exists?

  3. I agree that the art market presents a fascinating application of machine learning, particularly as it relates to pricing and valuation. The comparison to a stock exchange is apt — we already use auctions as pricing mechanisms for art, and it does not seem unreasonable that this could be evolved. I echo Jaclyn’s point that this advancement is predicated on open access to (historical) sales data in the first place. While there is currently a general norm of transparency in the commercial art world, there is no centralized database, especially for the primary market. How can we harness information that is right now distributed almost exclusively through word of mouth?

    In terms of predicting consumer preference, I do think that it will be important for machines to work with humans. Taste in art can be both nebulous and surprising. While a retailer like Stitchfix has amassed data points that ladder up to both fit and style, I’m curious what kinds of “tags” you envision to build an algorithm for art.

    For the creation of art, I think music and painting are fundamentally different. Music is generally more accessible, consumed at a much higher volume, and is arguable more commercialized. Are you envisioning a world that is generally more filled with art? Where art is accessible to everyone? What about the joy created by making art yourself?

  4. I definitely had not thought about machine learning for the art industry. Very interesting to learn a bi tmore about the trend this industry is reading. So I would beginning stating that I can definetly see the machine learninr business transforming the art industry, in a manner that I actually had never thought of. It makes completely sense that event his industry starts moving towards the machine learning, specially towards predicting and suggesting new paintings for consumer based on their historical preferences, and the market preferences.
    However, I would say that art is a controversial industry, where innovation and disruption is crucial for the business. Therefore, if Sotheby uses the input of the historical choices to predict the future, it will be probably be increasing its probability to underestimate the power of disruption in the industry, and be suggesting for consumers arts that to not portray the reality of the new trends of the art industry. In this sense, I would agree that for the more conservative buyers it does make sense the algorithm, but I would be more skeptical in utilizing this algorithm to suggest new art collections for consumers and clients that are more innovative in their art style collection.

Leave a comment