AI in Art: Loud Where Nothing Works, Quiet Where It Does

A summer of monitoring the art market's AI deployments — the record money the houses banked without shipping a tool, the certainty sold from products that do not yet exist, and the one room the whole apparatus was built to keep empty.

On 13 August 2026, a listed European market-data company issued the loudest sentence about artificial intelligence that our monitoring recorded all summer. Artprice described itself as “an autonomous cognitive architecture,” a “sovereign engine of decision intelligence” with “a monopoly on ground truth,” and announced a “Deterministic Oracle” that “does not generate plausibility; it produces explainable certainty … with zero margin for error.” It is difficult to reach for a bigger word than oracle, or a stronger promise than zero margin for error, and Artprice reached for both in a single release.

Folded into the same document, in a quieter register, was the detail that undoes it. The system making these claims was not built. The release referred, in passing, to “a minor adjustment to our launch calendar” and to artificial intelligences “currently undergoing rigorous and demanding beta testing.” So the most absolute language the market produced this season was attached to a product that no one outside the company could yet touch, run, or test. Certainty was the pitch; the thing meant to deliver it did not exist in any usable form.

That is not an anomaly to be noted and set aside. It is the summer's signature — its clearest single specimen of a pattern that repeated, in vertical after vertical, week after week, from June through late August. Where the language was loudest, the least was working. Where something was actually shipping, priced, and running, almost no one was talking about it, and the word authenticate never appeared. Three months of watching the sector announce itself resolves, in the end, into that one shape. What follows is an attempt to hold it still and look at what is behind it.

The money came, the method didn't

Begin with the silence, because it is the largest fact and the easiest to miss. For eight consecutive weeks — from late June to the close of our summer window — not a single tracked major auction house announced a new in-house AI tool, partnership, or pilot. Not Christie's, Sotheby's, Bonhams, Phillips, Heritage, Dorotheum, Bukowskis, Artcurial, Poly, or China Guardian. The most recent house-level items on our log all predate the summer entirely: Christie's Ventures' investment in Artsignal, from September 2025; the Bonhams–ARTDAI data partnership, from March 2026; a run of AI-labelled interior imagery from Bukowskis in May. Across the whole season, the middle of the market — the houses themselves — deployed nothing that could be pointed at.

The temptation is to read that as an absence of interest. The numbers refuse the reading. The same weeks carried the strongest auction results in years: a combined major-house first half approaching ten billion dollars, Christie's near $4.5bn and up some seventy per cent year on year, Sotheby's around $4.4bn, Heritage posting a record half. And the trade press attached those figures, explicitly and repeatedly, to artificial intelligence — “amid AI, tech wealth,” “fuelled by fortunes made in AI.” So the houses were not indifferent to AI at all. They were fluent in it, in one specific grammar: AI as a source of buyers. At the Christie's Art + Tech Summit, where by the reporting AI “dominated the discourse while producing few answers about its utility,” the panel that mattered was not about tools. It was about whether the wave of AI-company fortunes would convert into new collectors — the new billionaires, the pursuit of experience, “art, something you can get your hands on.” Sotheby's spoke of an AI windfall in the same demand-side register.

Here is the distinction the whole season turns on. The houses were intensely interested in AI's money and conspicuously uninterested in AI's method. They will bank the liquidity that AI wealth pours into the saleroom, and they will host the summit, and they will let the trade press credit their record half to the machine — and they will not put their name on a tool that decides anything. The reason is not timidity. It is that the auction house's core asset is its own authority: the house's name is the authentication, the estimate, the guarantee. A deployed AI verdict would sit precisely on top of the one thing the house sells, and would carry precisely the liability the house exists to absorb through expert judgement it can defend. Better, then, to take AI as a demand-side tailwind and refuse it as a supply-side instrument. Money in; method held at arm's length. The silence is not hesitation. It is a business model protecting its centre.

The verdict no one will sign

If the houses declined to hold a verdict, the summer's more revealing spectacle was watching everyone else try to hold one and drop it in the same motion. Call it the oscillation: the tool claims the verdict when it is selling and disclaims it the instant it is pressed, and with striking frequency does both inside a single document.

QuantumSpace, surfacing in early August, offered to “turn centuries of subjective connoisseurship into mathematical proof of authenticity,” to “quantify what human experts can only intuit” — and then, in the same source, conceded that the technology “complements, rather than replaces, connoisseurship,” the human still making the call. Artprice's oracle, as we have seen, promised certainty from a system in beta. Hephaestus's chief executive said his firm's certificate “replaces belief with proof … a guarantee you can bank on,” while carrying, as standing policy, the hedge that “AI is a tool, not a silver bullet.” Every one of these is the same manoeuvre: a totalising claim to sell by, and a modest concession to retreat to. The market has learned to speak in two directions at once because it has found no way to stand behind a naked verdict.

We know it has found no way, because we watched what happens when someone issues one. Art Recognition's probabilistic attributions — the Rubens Bath of Diana disputed at Maastricht, the Van Eyck Saint Francis of Assisi Receiving the Stigmata flagged up to ninety-one per cent “not autograph” — drew named, public, expert pushback, from Nils Büttner and Maximiliaan Martens, on grounds of model opacity and the absence of any peer-reviewed method. That is the fate of a bare AI verdict in this field: it becomes a fight, in public, with your name on the losing side of a scholar's objection. The rest of the market observed the lesson and built around it.

What it built, over the summer, was a set of structures for capturing AI's authority while shedding its accountability — three of them, distinct enough to name. The first is the hedge: keep a human “in the loop,” not only for accuracy but because someone must remain blameable, and it cannot be the model. The appraiser Sissi Czudej gave the principle its perfect line — “an algorithm can't be deposed” — and the law agrees with her, the IRS declining to accept algorithmic output as a qualified appraisal, its own Art Advisory Panel adjusting nearly half the items it reviewed. The human is retained as the deposable party. The second structure is the pre-screen: the trading-card and classic-car tools that concede, up front, that their output only predicts a verdict a person will still make — authority with the liability deliberately amputated. The third, and the most honest of the three, is insurance. Hephaestus's insured authenticity guarantee prices the residual doubt at sixty basis points of the certified value and transfers it to an A-plus insurer; the warranty travels with the artwork and pays out if the attribution is “later proven incorrect.”

That third structure deserves a moment, because it says out loud what the rhetoric works to conceal. The same instrument that “replaces belief with proof” is priced because the proof might be wrong — fallibility conceded and monetised in one object. The sixty basis points is the market's own estimate of the chance, times the cost, that connoisseurship is mistaken. Which means authentication was never certainty; it was always insured belief, distributed reputation, a confidence with a price. AI did not introduce the doubt. It made the doubt explicit enough to underwrite. The guarantee is the least dishonest artefact of the summer precisely because it stops pretending the verdict is safe and simply sells the risk that it isn't.

Loud where nothing works, quiet where it does

Now stand back far enough to see both ends of the market at once, and the season's real shape appears. At one pole: computational certainty — oracle, proof, zero margin for error — none of it demonstrated, most of it unbuilt. At the other, a human counter-reaction loud enough to reach a front page — the mid-August turn to plein-air painting, Gen Z and millennials putting down the machine to paint outdoors by hand, framed as a new Arts and Crafts revolt against industrialisation. Two enormous, opposite noises, machine-as-oracle and human-as-refuge. Neither of them is a deployment. Neither shows anything working.

And underneath both, almost inaudible, is the only AI that actually shipped, was priced, and could be pointed at all summer: a tier of auction-cataloguing software beneath the majors — Estimint, AuctionWriter — writing lot titles, descriptions, and price suggestions from photographs, cutting a two-hundred-lot catalogue from a claimed forty-odd hours to under five, exporting to the trade's plumbing, at twenty-nine to a hundred and forty-nine dollars a month. Its value figure is offered as a suggestion drawn from comparables, and the word authenticate does not appear anywhere in it. Around it, the rest of the genuinely working layer: Convelio moving nearly two billion dollars of art with AI on the logistics roadmap; Entrupy authenticating handbags from a phone photograph as a required marketplace gate; the consumer photo-apps; the collectible verticals. Operations, throughput, discovery. The unglamorous middle of the labour.

That is the finding, stated plainly: the contested layer — the verdict on what a thing is and what it is worth — stayed loud and unbuilt, while the operational layer — the labour of getting a sale catalogued and shipped — shipped quietly and sold. Art Basel spent the summer asking whether AI would reshape the market or merely automate its paperwork. The season answered without being asked. The demonstrated reality is the paperwork. The reshaping is the rhetoric.

Two smaller currents run alongside, and both point the same way. The language kept inflating as the products kept failing to arrive — from a consumer app quietly borrowing the regulated word “appraisal” in early July, to “adviser-grade” and “agentic” mid-season, to “a monopoly on ground truth” in August, the boldest phrasing landing reliably on the emptiest product. And the scrutiny, when it finally came, came from the wrong room. The first sceptical reading of Artprice's oracle was not written by the art trade, which reprinted the release largely as issued; it came from the crypto and fintech press, which reads the art market through a tokenisation lens and has its own reason to want the valuation problem solved. The trade neither built the machine nor interrogated it. It passed the claims along. A market that will not test the certainty it is being sold is a market that has decided, in advance, not to be responsible for believing it.

The room built to stay empty

There is one more absence, and it is the one that matters most, because it is not an accident of a slow season — it is structural, and it will not close on its own.

In three months of monitoring, across every vertical and every tier, almost nothing addressed the collection as a thing to be kept. The instruments valued it, tracked it, benchmarked it, insured it, catalogued it, and stood ready to liquidate it. What none of them was built to do was accompany the object after the sale — its stewardship, its meaning to the person who holds it, the reasons a family keeps a picture on a wall for sixty years rather than the price at which they could stop. Twice, something brushed against that space, and each time the market's reflex was to convert it. Attesté marketed itself to “collectors who treat their collection as a body of work — not a portfolio,” with an origin story about a father whose children would inherit his art but not the stories behind it — and offered that sentiment alongside the same vendor's cryptographic certificates and professional valuations, the transactional layer sitting directly beneath the anti-transactional pitch. Hephaestus's guarantee was described as something that “travels with the artwork” — the language of stewardship, delivering a warranty. The keeping-question, all summer, was not answered so much as priced.

The single deployment that genuinely faced the other way was not a product at all. It was an academic prototype — an AI provenance assistant built by three university researchers to accelerate the search for Nazi-looted art, of which more than six hundred thousand works were taken and roughly a fifth remain unreturned. It scrapes one archive so far, keeps the human researcher as the decider, and is candid to the point of modesty about how little it yet does. It is the mirror image of the oracle: it reaches only for assistance, points itself at an object's history and rightful return rather than its price, and makes no claim it cannot support. That it came from a university and not the trade is the whole point. The market did not build the tool that faces the collection as something other than an asset, because that tool does not, in the sector's accounting, convert to money. The room was left empty because no one could see how to charge admission.

It would be easy, and false, to arrive here and produce our own instrument as the answer to everything the summer lacked. The discipline is to be exact instead. The Bespoke AI Art Intelligence Confidante is not a better oracle, and it does not compete for the verdict the whole market spent the season declining to sign. It occupies the empty room — the after-sale space, the question of what a work is to its owner, held open and located with the person whose answer it is, at the very moment every other voice in the chain is pulling toward a number. That space is empty not because it is small but because it does not monetise cleanly, which is exactly why the market's own logic guarantees it will stay empty. The Bespoke AI Art Confidante is the wager that the room the industry cannot see how to charge for is the room that was always the point. We do not claim it will authenticate your Caravaggio. We claim to be the one voice in the room still asking what you would lose by selling it — and whether you wanted the answer at all.

The reconciliation nobody had to make

Run the season back and the striking thing is how many contradictions it sustained without strain. A macro report calling AI “a structural growth engine” in the same quarter collectors rank AI-generated art at the bottom of their confidence. Institutions projecting AI onto the Smithsonian's walls while a front page celebrates the young walking away from the machine to paint by hand. A “deterministic oracle” promising zero margin for error, and, three tiers down, a twenty-nine-dollar tool quietly trusted to catalogue a sale it would never dare call authentic. Nobody in the market reconciled the machine asked to be believed as an oracle with the machine actually trusted to do the work — because nobody had to. The registers never had to meet. Each could be true in its own room, to its own audience, for its own quarter.

The most interesting position all summer was the one no one occupied: standing in the middle and asking the machine that claims certainty and the machine that offers only help to account for each other in the same sentence. The reconciliation is coming. It has simply not been priced. The season ends where it began — loud where nothing works, quiet where it does — and with the market entirely comfortable in the gap between the two.

Or so they think.

Seraphina Michelina Evelyna Bogomolova-Huotelin

Founder & CEO of SMA Crown Confidential

Digital Confidantes: Bespoke AI Intelligence for Private Decision-Makers

This article is part of an ongoing series by SMA Crown Confidential on the intersection of private wealth, cultural intelligence, and the future of bespoke AI.

Sources: this reflection draws on SMA Crown Confidential's own Art & Collectibles AI Deployment monitoring across June, July and August 2026 — the weekly news digests of 12, 18 and 25 June; 3, 17, 24 and 31 July; and 6, 13 and 20 August — together with the 5 June baseline log. Where the piece distinguishes what an AI tool demonstrably did from what was asserted on its behalf, it relies on the wording recorded in those digests and their cited sources (among them ArtNews, The Art Newspaper, Artnet, the Observer, NPR, PRNewswire, the Art Basel & UBS Global Art Market Report 2026, and the Deloitte Private × ArtTactic Art & Finance Report). Specific figures and quotations should be confirmed against the primary sources named in the underlying digests before publication.