What’s worth more—a Picasso or a painting by a street artist no one has heard of? According to the AI model we built, the answer is the latter.
That surprising result came out of an experiment I ran with a data scientist and an AI expert from Silicon Valley. Our goal was to see whether artificial intelligence could bring more transparency—and, perhaps, greater fairness—to the art market.
The timing is urgent. The art world has been in a recession for 15 years, galleries are closing, young collectors are holding back, and artists trying to make it in the major market centers are living on the brink of poverty. The market is opaque and elitist. More than 50% of auction value in contemporary art comes from just twenty artists. The attention driven by blockbuster exhibitions and record prices is reserved for a handful of artists and galleries—under the pretense that their art is simply “better.” But is it really?
To find out, we built an AI model to decode how artistic value is determined in the art market. We wanted to test whether or not it is possible to evaluate visual quality independent of context – like gender, origin, education, gallery representation, collector influence, pricing history, or museum shows.
At the heart of our project was a multimodal model (LMM) designed to analyze both the visual characteristics and content of artworks, along with metadata such as medium, format, and creation date.
Our starting point: a cleaned and standardized dataset of millions of images, including price information. It featured masterpieces from major museums – from the Mona Lisa to recent works by the likes of Rashid Johnson – as well as the most expensive works ever sold at auction. Using this, we trained a “Fine Art Large Vision Model” (LVM) to predict auction prices based on what it could “see.” Market price became our pragmatic proxy – one of the few widely available, quantifiable indicators of value in the art world, even if it’s heavily skewed by trends, access, speculation, and power.
The early results were promising: in over 50% of cases, the model’s predictions based purely on visual data came surprisingly close to actual prices. But it soon became clear that more reliable predictions required extra metadata – like the artist’s name, provenance, or gallery representation.
After months of training on millions of images, the takeaway was undeniable: our model could not realistically estimate the price of an artwork based on the image alone. In one striking example, the AI valued a Picasso at under $1,000—while assigning a seven-figure price to a work by an unknown street artist I photographed in New York and uploaded into the system.
This revealed two things. First, the AI judged the street artist’s work to be of higher visual quality than the Picasso—challenging market logic at its core. Second, our model failed to produce market-viable results. Technically impressive, yes—but scientifically and commercially useless. Only once we added artist names and gallery affiliations did the model’s predictions align with real auction outcomes.
After extensive testing and optimization, we faced a sobering truth: the problem wasn’t just the AI. It was the training data itself—reflecting a market distorted by social and economic biases. Unlike object detection or medical imaging, visual quality in art can’t be objectively quantified. And because our dataset largely consisted of works already “validated” by the market, we ended up reinforcing circular patterns.
The results were revealing—and frustrating. The market doesn’t reward the artwork itself. It rewards the name. Galleries define what matters.
What is the takeaway for artists? That success is driven less by brushwork than by network. I discovered this years ago in a widely cited study published in Science. Given this, what still surprises me is how rarely art schools teach the business realities of being an artist – and how often artists cling to the belief that their art alone will make their careers.
As for AI, artists probably shouldn’t fear it. No machine can replace a studio visit, a real conversation, or the emotional pull of connecting with an artwork in person. Art is still a human business – built on trust, intimacy, and emotion. As for collectors, they should trust their instincts, even if that means buying an artwork they stumble on in a small gallery or even on the street. For what it’s worth, an AI will probably agree with you.
Maybe the true role of technology in the art market isn’t to price or rank art – but to reveal how it’s valued, and to help people discover what they love. In a market flooded with supply, algorithms can learn your taste, surface artists you’d never find otherwise, and break the grip of trend cycles. In other words, AI isn’t replacing artists – it’s replacing gatekeepers. I imagine a world where your feed shows you art that moves you—not just what’s trending in Basel, on Instagram, or at Sotheby’s. A world with transparent prices, and where artists without elite networks have a shot at being seen.
That’s the promise of AI in art. Not to replace human taste – but to empower it.
