Artificial intelligence. Taste-based image recognition. Data science.
I’ve never really thought of any of the above as particularly significant for the art market, or the auction houses that have been working in much the same way for centuries. As an inherently subjective undertaking, art — and its sale — has always felt resistant to this kind of number crunching and analysis.
According to Sotheby’s though, who recently acquired the AI startup Thread Genius for an undisclosed sum, I’ve been missing a trick.
Thread Genius: mapping the world’s visual taste
The company was founded in 2015 by engineers Andrew Shum and Ahmad Qamar, whose previous experience includes working on Spotify’s renowned music recommendation functionality.
Their new AI platform — originally designed for e-commerce, style-based enterprises — is based on a set of algorithms that instantly identifies objects and recommends images of similar objects; in other words, it aims to understand and map a user’s ‘tastes’ based on visual recognition.
You can check it out for yourself on the Thread Genius website.
The acquisition reflects Sotheby’s continuing efforts to use technology and data to improve its operations and business.
With Thread Genius’s platform, they hope to quickly and accurately present clients with art that matches their personal tastes. And by combining this tech with a 2016 acquisition — the Mei Moses Art Indices, a database of 50,000 works across eight collecting categories that tracks repeat sales at auction, and so arguably offers a more objective measure of value — Sotheby’s will gain additional insights for streamlining sales of available art.
Yet while this does sound impressive, I think it’s important to take a step back and consider some intriguing questions about AI.
Artificial intelligence will — by its very nature — have a mind of its own and so move in unexpected directions, but its core functionality and potential are still laid down by a set of discrete rules and inputs. For example, coders at Thread Genius will have defined what makes a match ‘good’ on the basis of a combination of input variables such as shape, colour, contrast, or perhaps other art-specific categories like style or medium.
Who’s to say that ‘taste’ is captured exhaustively by this finite and programmatically-defined list of variables? What about a mood or feeling that a work evokes, which is tied to a buyer’s or indeed artist’s unique personal history — say, features that link Picasso’s Guernica to Tracey Emin’s My Bed — which wouldn’t be picked up by image recognition alone?
Of course, Sotheby’s might have roadmap plans to codify these divergent factors, but it remains that an AI system is only as good as its inputs. An interesting consequence is that the AI could in principle start to define artistic tastes and values, if parties on either side of the sale are convinced its selections are generally ‘correct’ — especially if these are linked to historical sales data.
What’s more, no matter how good the AI gets, the reality is that the value of a work of art comes mostly from its originator. Buyers want a work by Picasso, not by a Picasso-like artist — or indeed by a perfect Picasso forger! So there appears to be an intrinsic limit to a system attempting to facilitate art sales based purely on visual recognition.
Before picking too many holes in Sotheby’s acquisition though, it’s important to note they’ll very likely use the technology as just one element alongside existing, staff-based processes, i.e. the long-standing method of checks and recommendations from art experts. And as one of many tools, its usefulness will come from speeding up aspects of this traditional work.
So while I don’t think it’s an unequivocally genius move that will disrupt a whole industry (Skynet won’t be forcing art upon us just yet!), I also don’t think it’s mere gimmick and applaud Sotheby’s for innovating and experimenting.
Whatever happens, I’m certainly interested to see where they take it.
Written by Thomas Stimson