The Tool That Promised Certainty
There’s a particular kind of quiet that falls over an auction house when something changes fundamentally. In late 2025, Sotheby’s introduced its AI-assisted provenance verification system, developed with Art Recognition AG, a Swiss firm whose machine learning model claims to identify artist attribution with 90% accuracy. The announcement felt momentous. Here, finally, was a technological answer to one of the art world’s oldest problems: how do we know what we’re actually looking at?
The system itself represents genuine innovation. Art Recognition AG’s deep learning model trained on more than 80,000 digitized artworks can identify brushwork patterns, pigment composition markers, and stylistic fingerprints that even experienced curators might miss. The algorithm had already been deployed to authenticate disputed works by Rembrandt and Raphael, with results suggesting the machine could see what human eyes had been trained to overlook for centuries. For a moment, it seemed like the art market had finally solved authentication.
But here’s what I’ve learned, standing in front of enough paintings to understand their weight: solutions rarely arrive cleanly. They arrive with consequences we don’t immediately recognize.
The Architecture of Bias
The first real challenges emerged quietly, then loudly. In early 2025, the Association of Art Museum Curators issued a formal position statement urging caution against treating algorithmic authentication as a standalone solution. Their concern was structural, not incidental. When you train a machine on 80,000 artworks, you’re training it on the collection of what has been preserved, documented, digitized, and valued enough to survive. That collection has a face. It has a geography. It has profound historical prejudices built into its very structure.
The numbers confirmed what curators suspected. A Heritage Science study published in 2025 found that AI authentication tools showed a 34% higher error rate when analyzing works by artists from the Global South. The training data gap wasn’t a minor problem waiting to be fixed in version 2.0. It was a feature, not a bug, of how authentication technology actually works. An algorithm trained predominantly on Western European male artists will naturally develop heightened sensitivity to their aesthetic markers. It will see their work more clearly because we’ve shown it more examples to learn from.
This matters viscerally when you think about what authentication means. It’s not just about confirming that Caravaggio actually painted something. It’s about determining whose work gets to enter history as legitimate, whose gets questioned, and whose gets quietly excluded from the market entirely.
What We Lost When We Sought Certainty
I keep returning to something a conservator told me years ago while we examined a small painting together. She said that authentication was never purely about the work itself. It was about the story we tell about the work, the provenance trail, the context, the arguments between experts who disagreed about what their eyes were telling them. That argument, she suggested, was often where truth lived.
When we hand that decision entirely to an algorithm, we’re not eliminating human bias. We’re just obscuring it. We’re moving it from the transparent space of human conversation into the black box of machine learning, where a curator can cite the 90% accuracy rate without having to defend the choice. There’s a dangerous comfort in that.
Meanwhile, Christie’s entered the conversation with its own authentication platform, also launched in 2025, using multispectral imaging combined with machine learning. At $4,500 per artwork assessment, the service positions authentication as a technical service rather than a curatorial practice. Both approaches imagine authentication as something that can be outsourced, standardized, and scaled. Both miss something essential about what expert judgment actually does.
Where Expertise Actually Lives
I want to be clear: I’m not against technological tools. I’m against the fantasy that they can replace the kind of attention expertise requires. You cannot train a machine to have looked at 10,000 paintings over thirty years. You cannot digitize the experience of standing in front of a canvas and watching how the light catches the texture of pigment, how your understanding shifts when you see the painting from different angles in different rooms on different days of your life.
The Art Recognition AG authentication technology represents real achievement. The researchers who built it understood something important about pattern recognition. But here’s what the Association of Art Museum Curators position statements kept emphasizing: algorithms work best when they augment human expertise, not replace it. When they’re one voice in a conversation rather than the final word.
The most rigorous authentication processes now involve human curators, conservators, art historians, and yes, when appropriate, algorithmic analysis working together. The machine catches things. The expert understands context. The expert knows that a painting that doesn’t match the training data might still be authentic. It might be a work that challenges what we thought we knew, and that challenge might be exactly where the most interesting questions live.
The Question We’re Actually Facing
The Sotheby’s moment of late 2025 will be remembered as the point when the art world had to confront something it had been quietly avoiding: authentication has never really been about objectivity. It’s about authority. Who gets to decide? When we hand that decision to a machine, we’re making a choice about whose forms of knowledge we trust.
I think about my grandmother’s quilts sometimes, how a textile expert can date them by looking at thread tension, fabric dye compounds, stitching patterns. That’s algorithmic thinking, in a way. Pattern matching. But it’s also embodied knowledge. It’s about having held enough fabric between your fingers to know what time feels like when it’s stitched into thread.
The art world didn’t fail when it adopted AI authentication. It failed when it started believing that authentication could ever be a purely technical problem. What we’re living through now is the necessary correction to that assumption. I find that more hopeful than the original fantasy, actually. It means we’re learning to ask better questions about what expertise means and why it still matters.
What aspects of this shift concern you most? Have you noticed these changes reflected in how museums and galleries present authentication information? I’d be genuinely interested in your perspective on where you think expertise should retain authority in this space.