top of page

Where the Signature Sits

Nov 13
8 min read

Updated: 4 days ago


AI and the art market's old method — a machine price taken back at the top, two vendors speaking in opposite voices, and a lender's value that will not hold still.


A print by Camille Pissarro, L'île Lacroix, à Rouen, was listed on 1stDibs at $28,000. Beside the dealer's price, the platform placed a number of its own, generated by a machine-learning model, and a line telling the buyer that the asking price was $27,470 above the recommendation. The machine thought the print was worth about five hundred dollars. The dealer Edward Pollack explained what had gone wrong: the tool "makes no distinction between posthumous and lifetime prints." Harris Schrank, another dealer, said the policy "demeans both the site and its dealers." The platform, after a while, stopped showing its number on expensive works.


The episode dates from 2024 and 2025, and it surfaced in our monitoring only this autumn. But it belongs at the front of this piece, because it is the only recorded case in five months of watching in which an AI price met real money in the art market. And what happened was a retreat.


That retreat says something about the art market that does not change quickly. The market is traditional — not in its taste, which can move in a season, but in how it makes a deal. A price is set by a dealer who can defend it. An estimate is given by a house that stands behind it. A loan is made against a value that a lender has checked in person. Our own long observation of this field is that it prefers to keep that method rather than trade it for a faster one. This fortnight's record shows the preference at work, and, more usefully, shows exactly where it holds and where it does not.


The number that came down

Look again at what the 1stDibs tool missed. The difference between a lifetime impression and a posthumous one is not a detail. It is precisely the kind of distinction a print specialist is paid to make, and it moves value by an order of magnitude. A model trained on prices can see that two sheets carry the same image; it cannot see which one the artist saw. The number it produced was confident and wrong in the one respect that mattered.


The platform's response was telling in its order. In 2024 it removed its recommendations above $50,000, and wherever the gap to the dealer's price was large. In 2025 it went further, removing the "1stDibs Estimate" from new seller-direct pieces and keeping it only on vintage items. A seller survey of 500 respondents, reported at the time, found three in four describing major problems, among them what sellers called "buyer miseducation." Who ran that survey is not established.


So the machine number left the market from the top down. It disappeared first where prices were highest, where a mistake costs most and where a named person most needs to stand behind the figure. It stayed longest where the stakes were lowest. That is not a technical outcome. It is the market's old method reasserting itself at exactly the point where it matters.


Two voices, two rooms

The same fortnight brought two art-data vendors into view, and their language moved in opposite directions.


Artsignal, the valuation platform backed by Christie's Ventures, spoke through a single interview in Cultured on 10 November, its co-founder and chief executive Sam Glatman the only person quoted. For the first time it named art-backed lenders among its customers, alongside auction houses, galleries, museums and institutional collectors — though no organisation was named. More striking was the change in tone. Earlier in the year this company spoke of being "adviser-grade" and a "dominant intelligence layer." Now its authentication was "not a definitive stamp", its advice "not as detailed as a top art advisor", and its answers, it conceded, could "look completely different each time" the same question was asked.


The day before, Artprice announced Artprice Visual Search®, inside its third-quarter revenue release. The beta is due in early December, the commercial launch in January 2027. The release promised to recognise a work "in a complete and absolute manner", to reduce uncertainty "to a statistically marginal level", and to give subscribers "investigative and valuation power." People appear in it only as the source of the annotated images the system learned from; no one checks the result. No benchmark, test set or error rate accompanied the claims. A year earlier, Artprice's own news service had said that "human editorial experience remains the definitive filter." By August 2026 the company was describing a "deterministic oracle" with "zero margin for error." November kept the August voice.


Set side by side, the two do not contradict each other so much as reveal the rule they both follow. The strength of the claim tracks the audience. Artsignal is turning toward lenders, who will ask whether a figure can be defended later, and its claims have narrowed. Artprice is speaking inside a revenue release, to subscribers and shareholders, and its claims have widened. The quieter voice is the one addressed to the people who would have to rely on the number. Neither vendor names a single user who relies on its tool for a decision.


A value that will not hold still

The sharpest line of the fortnight sits inside the Artsignal interview, and it sits there twice. Lenders, we are told, use the tool "as an independent assessment of value." And the same tool's reasoning can look different every time it is asked.

A loan secured on art rests on a value that can be written down, checked, and defended if the loan goes wrong. A figure that shifts when the question is repeated cannot carry that weight. There are only two readings. Either lenders are not relying on the tool in the way the phrase implies — using it, perhaps, as one more input beside their own judgement — or no one has yet asked them to reproduce the figure under pressure. No lender is named. No loan is shown written on the number.


The only lender named anywhere in our monitoring tells the old story plainly. Citi Art Finance values artwork in-house, lends at around half of that value, and inspects the work in person at the start of the loan and every year after. Its art advisory "does not advise clients on purchase or sale prices." The vendor says lenders rely on AI. The one lender with a name relies on its own people.


Use without a signature

If the signed decisions stay with people, where does the market use AI at all? The answer, this fortnight as all autumn, is: privately, and below the level of announcement.


The First Thursday report on AI in galleries — 103 gallery professionals, mainly in the UK, Europe and the US — came round again, this time through the Observer. In this telling, 84 per cent use AI, four in ten regularly, eighty per cent through personal accounts, eight per cent under a written policy, and 63 per cent with no guidance from management at all. The figures drift a little from one retelling to the next, and the original report has not been opened in any entry of our monitoring, so they stand here as reported rather than confirmed. But the shape is clear enough. Gallery staff use AI as individuals, on their own subscriptions, for their own tasks, outside any system the gallery has sanctioned. It is use without a signature attached — which is precisely why it never surfaces as a deployment.

At the auction houses the pattern is the same at a different scale. For the fourteenth consecutive count, no major house announced an in-house AI tool. Beneath that silence, the cataloguing vendors keep multiplying — Bidsquare's AI Auto Cataloging is new to our record, joining Estimint, Gavelist, AScribe and Circuit Auction — while the houses that use them go unnamed. The one internal trace, a Heritage Auctions posting for an AI Solutions Specialist on an Enterprise AI team, circulated again on job boards and went no further. The work gets done. No one puts the house's name on it.


This is the line the market draws, and it draws it consistently. Where nothing is signed, AI moves forward without ceremony: a gallery assistant drafting a press text, a mid-market house writing lot descriptions from photographs. Where a value must bear a name — a dealer's price, a specialist's attribution, a lender's collateral — the old method holds, and the one machine number that reached that seat was taken back.


What the bridge carries

There is a second thing this fortnight shows, and it concerns not who decides but what travels. When art crosses into finance's systems — into a lender's file, a wealth platform, an insurer's risk model — what crosses is value. Artsignal's pitch to lenders carries comparables and "valuation evidence and logic." When provenance appears, in its Lalanne example, it appears as something that raises the price. The 1stDibs model lost the difference between a print the artist made and one made after his death. In both cases the history of an object matters only as far as it moves a number.


The absences complete the picture. No insurer, lender or court is shown relying on an AI art valuation. No collector gave a first-person account of using AI in a judgement. Family offices, museums and estate data produced nothing this fortnight, and estate data has produced nothing since our monitoring of these lenses began. Meanwhile, sentiment runs well ahead of practice: in the Art Basel & UBS survey this year, nearly two thirds of respondents expected AI to improve pricing and valuation, and about the same share expected it to help with provenance and authenticity. Yet the only named people this fortnight in contact with an AI price were dealers objecting to one.


What does not cross is everything a family keeps a picture for: who owned it and why, what it meant in the house, what would be lost with it. That part of an object has no field in a lender's file. It is also the part the market's traditional method — the dealer who knows the sheet, the specialist who knows the hand — was always partly built to protect.


Backwards, or holding

It is tempting to call the art market backward: slow to adopt, attached to handshakes and inspections and names. Our own reading for a long time was close to that, and the record still supports the observation that the market prefers to stay with its method rather than move away from it.

But this fortnight suggests a sharper way to put it. The market is not refusing AI. It is sorting it. It lets the machine in wherever no one has to answer for the result, and it keeps the machine out wherever someone does. That is not lag. It is a judgement about what a valuation in this field actually is: not a number, but a number that a named person will stand behind when it is tested. AI has not yet produced that person, and until it does, the traditional method is not falling behind the technology. It is holding the one thing the technology cannot yet carry.


The vendors have noticed. One is lowering its voice to speak to lenders. The other is raising it to speak to shareholders. Neither has yet brought forward the person who will sign beneath the number.


Founder & CEO of SMA Crown Confidential


Bespoke AI Intelligence for Private Decision-Makers


Comments


bottom of page