Can AI Authenticate Art? What an Algorithm Can and Cannot Verify Before You Bid

Can AI authenticate art well enough to carry a bid? In October 2026 the question moved from the laboratory to the collector's desk. On 8 October, Art Basel and UBS published their Survey of Global Collecting: 22% of 3,100 high net worth collectors now use apps or AI tools for art advice, against 4% in 2024. Six days earlier, a US federal court sentenced a Miami dealer to 46 months for selling fake Warhols backed by forged foundation paperwork, the kind of fraud no image classifier is built to see. The short answer is that an AI model can estimate how closely a work's visible style matches an artist's accepted body of work, and it can flag a problem worth investigating. It cannot verify title, edition records, documents or condition. For a collector weighing a lot in a timed online sale, the useful question is how much weight a probability score can bear, and which checks must sit beside it.

Key takeaways:

• 22% of high net worth collectors used apps or AI tools for art advice in 2026, against 4% in 2024; 20% name fraud, forgery or market manipulation among the risks AI brings.
• Art Recognition, a Swiss AI authentication firm, has built models for more than 200 artists and charges around $2,200 per work.
• The first lot sold on AI authentication alone, a Marianne von Werefkin watercolour, made CHF 15,000 at Germann Auctions in Zurich in November 2024 against a CHF 5,000 to 8,000 estimate.
• The costliest recent contemporary frauds turned on paper and ownership: over $2.5 million in admitted fake Warhol sales in Miami, and more than $86 million in the Inigo Philbrick title fraud.

Why are collectors turning to AI art authentication now?

AI art authentication is the use of machine learning models, trained on photographs of an artist's accepted works, to estimate whether a new work shares that artist's visual signature. Its appeal in 2026 rests on three measurable shifts in collector behaviour.

The first is research. 72% of surveyed collectors carried out moderate or significant independent research before buying in 2026, up from 62% in 2025, and the figure reaches 80% among Gen Z collectors. The second is the move to digital tools: 58% relied on online resources for advice and recommendations, and the share using apps or AI tools rose more than fivefold in two years. The third is expectation. 67% of respondents expect AI to improve the discoverability of artists and artworks, while 20% already name fraud, forgery or market manipulation as the risk AI brings. Noah Horowitz, chief executive of Art Basel, described "a collector base that is increasingly informed, digitally engaged, and active across categories and channels".

The same survey shows collectors becoming more cautious about what they buy. Only 45% bought work by an artist they had discovered for the first time, down from 66% a year earlier and the lowest level in five years, while established artists made up 46% of works in collections against 25% in 2024. A buyer moving towards established names is a buyer moving towards the artists most exposed to forgery, because a market only attracts fakes once its prices reward them. Warhol, Basquiat, Haring and Banksy carry deep secondary markets and long histories of disputed works.

The demand for evidence also has a structural cause. In the early 2010s the market lost two authorities it had relied on for years: the Andy Warhol Art Authentication Board, the Warhol Foundation's panel that issued opinions on disputed works, announced its dissolution in October 2011 after more than $7 million in lawsuits, and the Basquiat estate ended its own authentication service in the same period. The Keith Haring Foundation no longer offers authentication either. Collectors who type "can AI authenticate art" into a search engine are, in effect, asking who replaced those boards. Our earlier analysis of the art authentication crisis traces who fills that gap today. AI vendors present themselves as one answer, and a collector deserves to know exactly what that answer covers.

How does an AI model reach a probability about authorship?

An AI authentication model is a convolutional neural network (a type of image recognition system that learns patterns such as edges, textures and brushwork from pixels), trained separately for each artist. Art Recognition, the Swiss company launched in 2019 and based in Adliswil, describes its analysis as a "fourth pillar" of authentication, set beside connoisseurship, forensics and provenance. By March 2026 it had built models for more than 200 artists.

Each model learns from two photographic datasets. The positive dataset holds images of undisputed or widely accepted works by the artist. The negative dataset holds similar but inauthentic works: known forgeries, copies by students and followers, workshop pieces, and even AI generated images in the artist's style. The two sets are balanced by genre, so that portraits are compared with portraits and landscapes with landscapes. Developers augment each image by changing lighting, flipping and rotating it, then cut it into small squares called patches, so the network studies fine brushstrokes in isolation as well as overall composition and colour. Training a model on one artist takes up to a week; analysing a single work then takes a day or two.

The output is a probability, a figure such as 86% or 91%, expressing how far the submitted image falls inside the pattern the model learned. That figure describes visual similarity to a training set. It says nothing on who owned the work, when it left the studio, or whether the paper and pigments belong to the period. Carina Popovici, Art Recognition's co-founder and chief executive, acknowledged the limits of interpretation herself: "We can speculate, but we don't really, really know for sure." She has also stressed that the datasets are "the product of scholarly expertise", assembled by in-house art historians from catalogues, biographies and academic literature.

Three technical constraints follow directly from this design, and each one matters to a buyer:

• Image quality. The model sees a photograph. Lighting, resolution and colour calibration change what it learns and what it judges. Jane Kallir, president of the Kallir Research Institute and author of the Egon Schiele catalogue raisonné, notes that proper training can require thousands of images, a corpus few artists ever produced.
• Style drift. An artist's manner changes across a career, which shrinks the comparable dataset for any given period.
• Opacity. The models and datasets are commercially sensitive. Sharon Hecker, art historian and curator, put the scientific objection plainly: "Any science has to be independently replicable."

The price of a report is modest against a serious lot. Art Recognition charges around $2,200 per work, depending on the detail requested, and needs only a photograph, which removes the shipping and insurance costs of sending a work to a specialist.

Can AI authenticate art? Three test cases and the precedents they set

The public record of AI attribution now covers enough cases to judge its strengths and its limits. Three of them define the debate for a collector.

The Werefkin watercolour at Germann Auctions, November 2024

Germann Auctions in Zurich offered three works from the collection of the late curator Martin Kunz in November 2024, each with an authenticity certificate issued by Art Recognition. Two of them, a 1945 drawing by Louise Bourgeois and a 1981 collage by Mimmo Paladino, were already backed by the artists' archives; they sold for CHF 28,000 and CHF 19,000, both below their low estimates. The third, an untitled and undated watercolour attributed to Marianne von Werefkin, had no prior expert certification. Offered on 25 November 2024 with an estimate of CHF 5,000 to 8,000, it sold for CHF 15,000, about $17,000.

Fabio Sidler, chief executive of Germann Auctions, framed the economics: "Since A.I. only requires a photograph, there are no further costs associated with transportation and insurance." The sale is a genuine precedent, the first lot sold with AI as its only authentication. Its limits are equally clear. The price sits in the low five figures, a single regional house chose which works to submit, and the AI did not publish a percentage for any of the three lots on that occasion.

The Caravaggio "Lute Player" and the probability that contradicts the consensus

Art historians have long agreed that, of three known versions of Caravaggio's "The Lute Player", the Hermitage and Wildenstein versions are autograph and the Badminton House version is a copy. In September 2025 Art Recognition reported an almost 86% probability that the Badminton House version is authentic, and judged, with less statistical certainty, that the Wildenstein version is likely a copy. Keith Christiansen, former curator at the Metropolitan Museum of Art, disputed both conclusions. Angelamaria Aceto, senior researcher at the University of Oxford's Ashmolean Museum, set out the specialist's position: "connoisseurship is about contextualizing things. It's about thinking critically."

Rubens, Raphael and the battle of the AIs

In 2021 Art Recognition calculated a 91% probability that "Samson and Delilah" at the National Gallery in London was not painted by Peter Paul Rubens. The museum rejected the finding, stating that "Not one single Rubens specialist has doubted that the picture is by Rubens." In 2023 two AI analyses of the disputed "de Brécy Tondo" reached opposite conclusions: researchers at the University of Bradford, using facial recognition, found it was by Raphael, while Art Recognition found an 85% chance it was not. Art Recognition then published its Raphael dataset, calling transparency a "prerequisite for credibility". The model has also matched established scholarship, assigning a 97% probability to a long disputed Van Gogh self-portrait at the National Museum in Oslo, an attribution later confirmed by technical analysis and the artist's letters.

Across these five cases, AI findings added confidence when they agreed with documentary and technical evidence, while findings that stood alone against the consensus moved neither the attribution nor the market. Jane Kallir warns that AI "potentially deceives by offering the illusion of objective certainty". Her institute receives more than 100 Schiele submissions a year, and about 95% of them prove to be fakes, forgeries or misattributions.

Where does forgery sit in the contemporary market?

For most contemporary lots between €800 and €50,000, fraud lives in the documents around the work more than in its brushwork. Screenprints, lithographs, multiples and editioned photographs are produced mechanically or with studio assistance; their visual style is reproducible by design. What separates a genuine Warhol screenprint from a counterfeit is the chain of paper: the edition number and its match with the published edition size, the stamps and their position, the printer's records, the certificate and the ownership history.

Can AI authenticate art made in editions?

Style analysis has the least to say about editions, because every legitimate impression shares the same image. The fraud cases of 2026 show where buyers lose money. Leslie Roberts, former owner of Miami Fine Art Gallery in Coconut Grove, was sentenced on 2 October 2026 to 46 months in federal prison after pleading guilty to conspiracy to commit wire fraud. He admitted more than $2.5 million in counterfeit sales. He persuaded buyers that the works came directly from the Andy Warhol Foundation, using forged authentication documents, stamps and identification numbers, and a co-conspirator impersonated a Phillips employee to "verify" paintings when a family challenged them. A private civil suit had alleged losses of $6.7 million. A Christie's appraisal first identified the forgeries, and the fake Phillips verification collapsed when the family contacted Phillips directly.

The Orlando Museum of Art case followed the same logic. In April 2023 Michael Barzman, a former auctioneer, admitted that he and an accomplice created 20 to 30 fake Basquiats on cardboard in 2012, spending five to 30 minutes on each, sold them on eBay, and supported them with a notarized false provenance claiming they had been found in a screenwriter's storage unit. The FBI seized 25 of them from the museum in June 2022. The paintings fooled buyers and a museum through a story on paper; among the evidence agents later put to Barzman was a mailing label bearing his name, painted over on the back of one cardboard work.

Ownership and title: the risk no image reveals

Title is the legal right to own and sell a work, free of liens, claims and restitution demands. The dealer Inigo Philbrick defrauded collectors and lenders of more than $86 million between 2016 and 2019 by selling more than 100% of the same works and pledging them to several lenders at once. The fraud concerned ownership rather than authorship, and ownership is invisible to any style model. The art market still has no public register of ownership, fractional interests or liens, and an appraisal remains a point in time opinion. Our guide to restitution and title checks sets out the searches that close this gap.

Estate disputes add a further layer. In April 2026 a Manhattan jury awarded $102 million to the Morgan Art Foundation, Robert Indiana's former agent, in a case against a publisher producing unauthorised works. Caroline Sayan of Cadell + Co describes how stricter provenance standards have changed the reading of old paperwork: "What used to be a golden ticket… is now seen as a red flag." A collector relying on a certificate of authenticity for art needs to know who issued it, whether that body still stands behind it, and whether the catalogue raisonné agrees.

Reading an AI report before you bid: a ten point checklist

An AI report is evidence of a specific kind, and a collector should interrogate it with the same discipline applied to a condition report or a provenance file. These ten points separate a useful report from a decorative one.

1. The model's training set. Ask how many accepted works trained the model, from which catalogue raisonné, and whether the period of the lot is represented. A model trained mainly on late works says little about an early one.
2. The negative examples. Ask whether known forgeries, workshop pieces and followers were included. A model that never saw a good copy cannot distinguish one.
3. The probability and its scale. Read the exact percentage and ask how the provider calibrates it. An 86% result leaves a 14% chance the model is wrong, before any other uncertainty is counted.
4. The image analysed. Confirm the resolution, lighting and whether the photograph shows the whole work, including edges and signature. A phone image under gallery lights is a weak input.
5. Who commissioned the report. A report paid for by the consignor serves the consignor's interest. Ask whether the house or an independent party ordered it.
6. The medium. For screenprints, lithographs, multiples and photographs, treat any style score as secondary to edition records, stamps and printer's documentation.
7. The documents. Match the edition number, the edition size and every stamp against the catalogue raisonné. Contact the issuing foundation or studio directly; the Miami case turned on a buyer doing exactly that.
8. The ownership chain. Request the provenance with dates and named owners, and ask whether the house has checked stolen art and loss registers. A gap of decades needs an explanation.
9. The condition report. Read the professional condition report for restoration, trimming, fading or later additions. A restored signature changes the reading of any attribution.
10. The scientific evidence. For a unique work at a significant price, ask whether pigment, paper or fibre analysis exists. Scientific testing of a contemporary work can rule an attribution out; it rarely proves one.

A report that cannot answer the first five points adds little to the file. Points six to ten apply whether or not an AI report exists, and they carry the decision.

Who bears the risk when an attribution fails after the sale?

Under standard auction terms, the buyer carries most of the authenticity risk, and an AI report does not shift it. Christie's conditions of sale state that the description of a lot, any condition report and any statement about artist, period or provenance "are our opinion and not to be relied on as a statement of fact". The house offers a limited authenticity warranty: the buyer's only remedy is to cancel the sale and recover the purchase price, with no compensation for other losses, and the house may require the written opinions of two recognised experts confirming that the lot is not authentic. For South East Asian modern and contemporary art and for Chinese calligraphy and painting, Christie's states that the warranty does not apply at all, because "current scholarship does not permit the making of definitive statements". Warranties of this kind typically exclude claims that could only be proven by a scientific process not generally accepted at the date of sale.

That last clause matters for AI. If a model trained after the sale later casts doubt on a work, the buyer cannot assume the house will refund. Before bidding, a collector should read the conditions of sale for four specific points:

• the remedy available if a work is later shown to be inauthentic, and the deadline for a claim;
• the evidence the house will accept to prove a claim, and whether an AI report counts;
• whether the warranty covers the attribution line only, or also dates, edition numbers and medium;
• whether the warranty passes to a later owner or stays with the original buyer.

A house that publishes clear answers to those four questions has prepared for disputed authenticity before a dispute arrives, and the buyer can then price the remaining risk before bidding.

A worked reading: an €18,000 screenprint with an AI report attached

Consider a representative scenario. A timed online sale offers a Warhol screenprint, catalogued as signed and numbered from an edition of 250, estimated at €15,000 to €20,000. The consignor supplies an AI report showing a 92% probability that the work matches Warhol's screenprint output, alongside a certificate from a private dealer and a provenance listing two prior owners since 1995.

The AI figure adds little here. Every authentic impression of that edition shares the same image, and so does a good counterfeit printed from a photographic source. The decision rests on other checks. The edition number must fall within the published edition size in the catalogue raisonné of Warhol's prints, and the stamps on the verso must match the forms and positions recorded for that edition. The dealer's certificate is only as strong as the dealer; after the Miami case, a certificate claiming to come from the Warhol Foundation itself demands verification with the Foundation. The two prior owners should be named and dated, and the condition report should state whether the sheet has been trimmed, faded or laid down.

The arithmetic of a mistake is simple. At a hammer price of €18,000 with a 20% buyer's premium, the buyer pays €21,600 before shipping and taxes. If the print later proves counterfeit and the conditions of sale cap the remedy at a refund, the best case returns €21,600 after a dispute; the worst case returns nothing. Spending an hour on points six to nine of the checklist costs far less than either outcome. An AI report at around $2,200 would represent roughly a tenth of the hammer price, and for an editioned print it answers the least important question in the file.

FAQ: can AI authenticate art? Questions collectors ask

Can AI authenticate art on its own?

No AI system has yet replaced documentary, technical and scholarly evidence for a significant work. The first sale on AI authentication alone, a Marianne von Werefkin watercolour at Germann Auctions in November 2024, made CHF 15,000. In higher value disputes, such as the Caravaggio "Lute Player" in 2025 and the Rubens "Samson and Delilah" in 2021, AI findings that contradicted specialists left the accepted attributions unchanged. AI works best as one input that flags a question for experts to investigate.

How accurate is AI art authentication?

Providers publish probabilities per work, for example 86% for the Badminton House "Lute Player" and 97% for a Van Gogh in Oslo. Those figures express similarity to a training set, and their reliability depends on the quality of the images, the size of the artist's accepted corpus and the negative examples used. In 2023 two AI systems reached opposite conclusions on the same disputed Raphael, which shows that method and data drive the outcome.

Does an AI report protect me if the work turns out to be fake?

An AI report does not change the contract between buyer and auction house. Under standard conditions of sale, such as those published by Christie's, the remedy for an inauthentic lot is a refund of the purchase price within the terms of a limited warranty, and the house may require written opinions from two recognised experts. Claims based on a scientific process not accepted at the date of sale are typically excluded, which covers later AI analysis.

Is AI useful for checking prints and editions?

AI style analysis is least useful for screenprints, lithographs and multiples, because every genuine impression carries the same image. Fraud in editions lives in edition numbers, stamps, certificates and ownership records. The Miami dealer sentenced to 46 months in October 2026 sold fake Warhols on forged foundation documents and identification numbers, and buyers exposed the scheme by contacting the auction house named in the paperwork directly.

How much does an AI authentication report cost?

Art Recognition, the Swiss provider, charges around $2,200 per work, with the price depending on the level of detail in the report. The analysis needs only a photograph, so it avoids the shipping and insurance costs of sending a work to a specialist. Against a lot estimated between €15,000 and €20,000, that fee represents roughly a tenth of the hammer price.

How LLB Auction applies due diligence before a lot reaches the sale

LLB Auction does not claim that any lot is authentic. Every work goes through internal, documentary due diligence before it is offered, and the house places that evidence in front of the buyer so that the collector can weigh it directly. Three practices define the approach.

An intake threshold. The house turns down roughly 40% of the works submitted to it. A work that reaches the catalogue has cleared certificate verification, a review of its ownership history and a check of its conservation records.

Documented condition. Each lot carries a three page professional condition report, published with the lot. The condition report is the reference, and the catalogue description defers to it. Authentication is carried out internally, and the buyer carries the authenticity risk, which is why the house sets out its evidence in full.

Transparent terms. Fees are disclosed before bidding: a 20% buyer's premium and a 10% seller's commission, with no hidden costs. The house runs timed online sales of seven to 14 days on its own platform, also listed on Artsy, with four sales in 2026 and two more committed for 2027. Its Contemporary Art Spring 2026 sale on 26 May offered 25 lots and sold 23.

Collectors who want each catalogue and its condition reports as soon as a sale opens can register at llb-auction.com, and registering to bid explains the identity checks involved. Ask your questions. We have answers.

Conclusion

AI attribution has earned a seat in the conversation. It analyses formal detail across thousands of images at once, it has matched established scholarship on works such as the Oslo Van Gogh, and it supported a first auction sale in Zurich in November 2024. The 2026 survey data show collectors adopting AI tools quickly, with use rising from 4% to 22% in two years. The same year's fraud cases show where buyers in the contemporary segment lose money: forged foundation documents, invented provenance, mismatched edition numbers and ownership sold twice.

For a collector buying between €800 and €50,000, the practical rule follows from that evidence. Treat a probability score as a prompt for questions, ask who commissioned it and on what images, and give the final weight to the documents, the ownership chain, the catalogue raisonné and the condition report. Can AI authenticate art? It can strengthen a well documented attribution and flag a weak one; the decision to bid still rests on evidence a camera cannot capture.

Also worth reading:

• The art authentication crisis: who verifies a work when the boards are gone
• Certificate of authenticity for art, catalogue raisonné and which proof counts
• How to read a professional condition report
• Restitution and title checks that protect a buyer
• Scientific testing of a contemporary work

Sources:

• The Art Basel and UBS Survey of Global Collecting 2026 reveals Gen Z as the highest-spending generation : UBS, 8 October 2026
• From AI to Gen Z: everything to know about the 2026 Survey of Global Collecting : Art Basel, 8 October 2026
• Wealthy Gen Z collectors are the art market's biggest spenders, survey finds : The Art Newspaper, 8 October 2026
• Miami dealer sentenced to almost 4 years in prison for selling fake Warhols : The Art Newspaper, 21 October 2026
• Estate fallouts expose challenges of posthumous art ownership : The Art Newspaper, 5 October 2026
• Art does not function like other commodities, so why is it being priced like one? : The Art Newspaper, 19 October 2026
• Masterpiece or cheap copy? Art historians and AI may not agree : CNN, 9 March 2026
• In the age of AI, can art expertise be digitised? : The Art Newspaper, 16 January 2026
• First A.I.-authenticated artwork sells big at auction in "pivotal test case" : Artnet News, 5 December 2024
• Swiss auction house to offer works authenticated by AI : The Art Newspaper, 31 October 2024
• NoHo man admits lying to FBI about his role in creating fake Basquiat paintings : U.S. Attorney's Office, Central District of California, 11 April 2023
• Warhol Authentication Board to close : Artlyst, 20 October 2011
• Conditions of Sale for Christie's online-only sales : Christie's Hong Kong, 8 February 2022
• Art Recognition: AI art authentication : Art Recognition AG, 2026