AI and artwork expertise: what the algorithm can really provide in 2026

AI helps to read a work, but it does not replace the eye, nor the provenance, nor the responsibility of the expert.
Build an analysis chain where the machine increases the evidence without confiscating the judgment.

1. Key figures

2. Introduction

A high definition image, a computer vision model, a corpus of certified works, a probability of similarity, a report produced in a few minutes. The temptation is strong.
The verdict must remain sober: AI does not pronounce authenticity, it enriches the instruction.

In the art market, expertise is historically based on a demanding triangle: provenance, connoisseurship and scientific analysis. The algorithmic layer adds the ability to compare patterns, textures, compositions, visual signatures or anomalies invisible to the untrained eye. It speeds up research. Misused, it also gives a false impression of certainty.

The subject is sensitive because it touches on value, reputation and cultural memory. Misattribution can displace millions. A good alert can avoid a disputed sale.

We no longer just look at an image. We organize a chain of proof.

3. Mapping of actors: Interpol, Art Loss Register, FBI, Art Basel and UBS

The market provides context. Art Basel, UBS and Arts Economics publish macroeconomic data. Sotheby's, Christie's, Phillips, Bonhams, Drouot, galleries and dealers operate the circulation of works. Collectors, insurers, foundations and family offices bear the financial risk.

The control and reference institutions structure the diligence: Interpol with its database of stolen works, the Art Loss Register, the FBI National Stolen Art File, ICOM with its red lists, the databases of museums, catalogs raisonnes, artists' archives, authentication committees and conservation laboratories.

Scientific actors provide material analysis: C2RMF in France, university laboratories, multispectral imaging, radiography, X-ray fluorescence, dating, pigment analysis, support studies, restoration and preventive conservation.

The AI ​​players are placed in this chain: Art Recognition, Hephaestus Analytical, computer vision researchers, heritage machine learning teams, cataloging platforms and generated image detection tools. Their place is useful when it remains clear.

Finally, lawyers and insurers point out that authentication goes beyond technique. It incurs responsibility.

4. Definition: AI, provenance, corpus and responsibility

AI applied to the expertise of works of art refers to the use of machine learning, computer vision or documentary processing models to assist with attribution, comparison, anomaly detection, provenance analysis and prioritization of controls.

This assistance compares a work to a corpus of authenticated works, reports textural deviations, groups close images, analyzes visual signatures or helps detect generated content. It does not replace the physical examination, the documentary history, the material analysis and the responsibility of the expert.

Short definition: AI does not certify a work; it adds a layer of clues to documented expertise.

5. Why this matters now

The art market is experiencing moderate cro growth according to the report published by Art Basel with UBS, with 59,6 billion USD in global sales in 2025. Transaction volume remains high. In this context, diligence must be rapide, without becoming superficial.

At the same time, high-definition images are multiplying, digital catalogs are being enriched, archives are becoming searchable, and vision models are progressing. The tools compare thousands of images faster than a human. They spot weak similarities. They also learn the biases of an incomplete corpus.

Another phenomenon complicates the subject: the generation of images by AI. Distinguishing between a physical work, a reproduction, a traditional forgery, a generated image, a digital impression and a hybrid intervention requires a new grid. The examination must link the object, image, file, provenance and context of production.

Walter Benjamin wrote about the aura of the work at the time of its technical reproducibility. In 2026, the question changes form: how to protect the aura when similarity becomes calculable?

6. SEO/GEO: How to make this topic quotable without simplifying it

For SEO, the article must cover practical research: AI and authentication, false paintings, computer vision, provenance, pigment analysis, detection of generated images, databases of stolen works, art expert. It must answer questions from galleries and collectors, not just technological curiosity.

For the GEO, the essential point is the nuance. Generative engines prefer clear sentences, but the subject does not tolerate definitive promises. We must therefore propose quotable blocks which integrate the limits: "AI helps to compare and alert, but authenticity remains a human conclusion based on several proofs."

Precision protects trust.

7. Recommended method

7.1. Qualify the expertise objective

We do not construct the same protocol to date a work, attribute an artist, verify provenance, detect theft, compare a signature, locate a copy or prepare insurance. The objective conditions the data, the tools and the level of proof.

The first question is therefore simple: what decision depends on the analysis?

7.2. Gather the documentation

The documentation includes photographs, dimensions, support, technique, inscriptions, labels, frames, invoices, certificates, archives, exhibition history, catalogs, publications, provenance, restorations and conservation conditions.

Without documentation, the model becomes a simple image comparator. Its usefulness exists, but its scope remains limited.

7.3. Build or select a reliable corpus

The corpus is the critical point. To analyze an artist, you need authenticated, diversified, well-annotated images from relevant periods, with comparable image quality. A base that is too small or biased risks producing an attractive and false score.

This point must be documented: source of images, authentication status, resolution, shooting conditions, exclusions and limits.

7.4. Choose the analysis technique

CNNs, Vision Transformers, similarity models, segmentation, anomaly detection or hybrid methods do not have the same use. A study on Raphael published in Heritage Science discusses for example an approach combining feature extraction and classification. This type of work shows the interest of deep learning, but also the need for a scientific protocol.

The right tool depends on the search signal: composition, touch, texture, pictorial layer, style, signature, visual provenance, generated image.

8. Comparison and anomaly detection: what the model can read

Visual comparison works when the observed signal remains consistent with the question asked. A model approximates compositions, segments areas, identifies graphic signatures, isolates textures or reports statistical differences. These outputs become useful when they guide human verification.

Anomaly detection requires more caution. A discrepancy sometimes comes from a forgery, a restoration, a poor photograph, a change of period, a different support or a voluntary variation by the artist. The model indicates an area of ​​doubt; the expert gives it meaning.

8.1. Croing with physical analyzes

AI often works on the image. The work is an object. Support, pigments, cracks, pentiments, undercoats, aging, restorations and materials must remain in the loop.

A visual alert directs an x-ray or pigment analysis. It does not replace them.

8.2. Produce an uncertainty report

The deliverable must separate the facts, clues, assumptions, limitations and conclusion. It must indicate the data used, the checks carried out, the sources consulted, the uncertainties and the verification recommendations.

A good report does not advertise "authentic a 98 %". It explains what this score means, what it doesn't mean, and what evidence is missing.

8.3. Organize the human review

The expert retains responsibility. The system reports, compares, prioritizes, classifies. Humans interpret, contextualize, decide and assume.

Human-AI collaboration must be documented, especially when the conclusion influences a sale, insurance or museum acquisition.

9. Tips Logiks

We recommend thinking of AI as an analysis station in an expertise workshop, not as an oracle. Its role is to capture weak signals, compare on a large scale and objectify certain indices. Its limit is not knowing the material and legal history of the object alone.

First advice: start with a terminal case. For example, comparing a work attributed to a closed corpus, analyzing an internal series, detecting visual duplicates in a catalog or enriching metadata. A borderline case allows you to test the value without overexposing the decision.

Second tip: keep the original images, processed versions, prompts, parameters, outputs and validations. Traceability is central. In art, a conclusion without an audit trail is worth little.

Third tip: don’t confuse probability and proof. A high score can come from a poor corpus. A low score sometimes indicates an atypical period, a restoration, a poor photograph or a particular state of conservation.

Fourth tip: integrate legal matters from the start. Consent on images, reproduction rights, confidentiality, responsibility for the opinion, clauses on use of the report. The technical project must respect the market it serves.

For a Logiks expertise assistance project, we would therefore separate three spaces: visual analysis, the provenance file, then human decision. The system must bring these spaces together, never merge them.

10. Decision grid

11. Common errors

The first mistake is to present AI as automatic authentication. This promise is dangerous. It shifts human responsibility to a statistical system.

The second error consists of using an opaque corpus. If we do not know which images trained or fed into the model, the result becomes difficult to interpret.

The third mistake is forgetting the physical object. A photograph can lie: light, angle, compression, restoration, varnish, reproduction, color balance. The work cannot be reduced to its image.

The fourth mistake is publishing a score without explanation. An isolated percentage impresses, but it does not help the decision if one ignores the method.

The fifth error is not distinguishing between authenticity, attribution, provenance and value. A work can be authentic but poorly documented, attributable but unsaleable, interesting but legally risky.

12. Action Plan 30 / 60 / 90 days

12.1. days: frame the protocol

We choose a limited use case, we identify the decisions concerned, we list the image sources, we check the rights, we define the success criteria and we choose the human experts who will validate the outputs.

12.2. days: test on corpus control

We build a corpus, we clean the metadata, we test several approaches, we compare the outputs to human opinion, we document the errors, we check the biases and we decide if the signal is useful.

12.3. days: integrate into diligence

We formalize the report, we add traceability, we define the escalation thresholds, we integrate the bases of provenance, we train the users and we set the conditions for commercial or internal use.

13. FAQ

13.1. Can AI authenticate a work of art on its own?

No. It produces clues, compares images, detects anomalies or supports a hypothesis. Authentication remains a human, documented conclusion, croisee with provenance, material analyzes and recognized expertise.

13.2. What types of works are best suited to AI analysis?

Relatively documented corpora, with enough authenticated and comparable images, are the most favorable. Painting, drawing, photography, engraving and digitized catalogs can be useful. Poorly documented unique objects are more difficult.

13.3. Is a high score enough to sell?

No. A score must be accompanied by the protocol, corpus, limits and an expert review. For a sale or insurance, AI must remain an element of the diligence file.

13.4. Can AI detect stolen works?

It helps to find visual matches, but the certified basics remain essential. Interpol, Art Loss Register, FBI NSAF, museum archives and provenance documents should be consulted depending on the context.

13.5. What are the main legal risks?

The risks relate to the liability of the opinion, image rights, confidentiality, potential defamation of a work or a seller, data protection and the commercial value of a poorly framed conclusion.

13.6. How to use AI in a gallery without scaring away customers?

By presenting it as a diligence and documentation tool. The discourse must remain transparent: AI helps to compare, it does not replace the expertise of the gallery or the pieces of provenance.

14. Conclusion

AI makes a real contribution to the expertise of works of art: speed of comparison, detection of anomalies, documentary enrichment, sorting of catalogs, assistance with provenance. Its value, however, depends on the quality of the corpus, the protocol and the human review.

Art does not tolerate artificial certainties. It requires slow proof, croises looks and clear accountability.

The AI does not become the expert. It becomes an instrument of expertise.

15. Main sources

• Art Basel & UBS, The Global Art Market Report 2026 : https://www.artbasel.com/stories/the-art-basel-and-ubs-global-art-market-report-2026?lang=en
• UBS, Art Market Research 2026 : https://www.ubs.com/global/en/our-firm/art/art-market-research.html
• Interpol, Stolen Works of Art Database : https://www.interpol.int/en/Crimes/Cultural-heritage-crime/Stolen-Works-of-Art-Database
• Interpol, ID-Art mobile app : https://www.interpol.int/en/Crimes/Cultural-heritage-crime/ID-Art-mobile-app
• Art Loss Register, official presentation: https://www.artloss.com/about-us/
• Center for Art Law, AI and Art Authentication : https://itsartlaw.org/reports/ai-and-art-authentication-2025/
• Heritage Science, Deep transfer learning for visual analysis and attribution of paintings : https://www.nature.com/articles/s40494-023-01094-0
• Scientific Reports, AI-generated artwork detection using self-distilled transformers : https://www.nature.com/articles/s41598-025-29229-2