When Silicon Valley Met the Old Masters: The Promise and Peril of Automated Authentication
Late last year, Sotheby’s announced what felt like a seismic shift in how the art world verifies authenticity. The auction house rolled out an AI-assisted provenance verification system developed with Art Recognition AG, a Swiss firm whose deep learning algorithm claims to identify artists with 90% accuracy. On the surface, this seemed inevitable: machine learning has disrupted everything from radiology to legal research, so why not art authentication? The promise was genuinely intoxicating. No more fuzzy attributions. No more decades-long scholarly debates resolved by committee consensus. Just clean algorithmic certainty.
But here’s where the dinner party conversation gets interesting. The tool was trained on over 80,000 digitized artworks and has already been deployed to authenticate disputed Rembrandts and Raphaels. Real works by real masters, settling real financial stakes. Yet within months of its launch, the art world’s institutional guardrails began to creak. The Association of Art Museum Curators position statements issued a formal caution against treating algorithmic authentication as a standalone authority. This wasn’t gatekeeping from threatened experts. It was a structural warning about what happens when you let machines make judgments they’re fundamentally unprepared to make.
The Training Data Problem: Why Your AI’s Blind Spot Is Also Mine
Here’s the uncomfortable truth that separates authentic innovation from technological theater: those 80,000 training images didn’t materialize from nowhere. They represent centuries of collecting, cataloging, and institutional privilege. They overwhelmingly feature works by Western European male artists whose careers were meticulously documented, photographed, and archived while contemporary artists from the Global South were fighting for basic exhibition space. Your algorithm is only as intelligent as the data it learned from, and art history’s training data is profoundly skewed.
A 2025 study in Heritage Science revealed that AI authentication systems showed 34% higher error rates when analyzing works by artists from the Global South. Not 5% higher. Not 12%. Thirty-four percent. This isn’t a rounding error or a temporary calibration problem. It’s a structural flaw baked into how these systems think. An AI trained primarily on Caravaggio can’t reliably authenticate work by contemporary Nigerian painters or 19th-century Brazilian sculptors because those artists simply weren’t adequately represented in the training set. The algorithm learns racism embedded in art history itself.
Think about the philosophical difference between a machine making a judgment call and a human expert making one. When a curator says “I’m uncertain about this attribution,” they’re making a professional confession. They’re saying their knowledge has limits, and those limits are part of their credibility. When an algorithm says “90% confident,” it’s performing certainty. It sounds like precision. But if that precision is built on asymmetrical training data, you’ve just automated bias and called it progress.
Christie’s Alternative: Multispectral Imaging and the Illusion of Neutrality
Christie’s didn’t wait around to watch Sotheby’s navigate this minefield. They launched their own authentication platform in 2025, combining multispectral imaging technology with machine learning. The approach feels different, more scientific, less algorithmic, but it’s worth interrogating that assumption. At $4,500 per artwork assessment, they’re positioning themselves as offering deeper technical rigor than their rival’s system. Multispectral imaging reveals layers invisible to the human eye: paint composition, aging patterns, underlying sketches. It feels objective in a way that Art Recognition AG authentication technology sometimes doesn’t.
But comparative criticism gets instructive here. We’ve seen this before in other fields. Digital forensics experts once claimed their tools were neutral arbiters of truth. Facial recognition systems promised to eliminate human bias. Each time, we discovered that tools designed to transcend human judgment instead encoded human assumptions more permanently. At least a flawed human expert can change their mind, learn something new, or admit they were wrong. An algorithm just propagates its training data’s limitations at scale.
The multispectral imaging approach has real advantages. It’s genuinely illuminating for technical analysis of materials and execution. But pairing it with machine learning whispers the same seductive lie: that technology makes judgment unnecessary. These tools should be augmenting expert analysis, not replacing it. The price point suggests Christie’s knows this. Four grand per assessment isn’t cheap, but it’s structured so that institutions and galleries still need to hire human experts to interpret what the machines reveal.
What Authentication Actually Means, and What We Risk Losing
Here’s what bothers me most about the Sotheby’s scandal: it forced the art world to explicitly articulate something we’ve always known but rarely stated clearly. Authentication isn’t actually about certainty. It never was. Authentication is about establishing a chain of reasoning, historical, material, stylistic, contextual, that convinces a community of experts and stakeholders that a work is what we claim it to be. It’s consensual knowledge production. Messy. Provisional. Revisable. This messiness is a feature, not a bug.
The moment we treat authentication as a technical problem to be solved by algorithms, we’ve misunderstood what authentication actually is. A 15th-century painting is authentic not because an AI says so with 90% confidence. It’s authentic because scholars have traced its ownership, analyzed its materials against period-appropriate techniques, compared it to documented works by the same hand, and collectively decided the evidence supports the attribution. That process takes time. It tolerates uncertainty. It allows for revision when new information emerges.
The caution issued by the Association of Art Museum Curators wasn’t anti-technology sentiment from threatened gatekeepers. It was a professional body essentially saying: we need these tools to help us see better, but we cannot outsource the actual work of making professional judgment to a black box trained on unexamined assumptions. That’s not conservative thinking. That’s intellectual rigor.
The Productive Tension We Should Lean Into
So where does this leave us? Not rejecting AI-assisted authentication. Not abandoning multispectral imaging or machine learning models. But recognizing them as what they actually are: powerful tools that augment human expertise rather than replace it. The scandal isn’t that these technologies exist. It’s that they were presented as if they could transcend the human work of authentication.
The real conversation we need to have concerns the future construction of training data. What does it look like if we deliberately archive and digitize work by underrepresented artists? What happens to AI authentication when it’s trained on genuinely global collections rather than Western European hierarchies? These questions matter more than the present capabilities of any single algorithm. They ask whether we’ll use these tools to reproduce historical inequalities or to interrupt them.
The Sotheby’s AI provenance scandal isn’t really about artificial intelligence. It’s about what we’re willing to automate and what we insist must remain human decisions. It’s about whether convenience is worth the cost of delegating judgment. What’s your position on where that line should be drawn?