ZharfAI Team
AI can measure patterns in a painting that no unaided eye could count: microscopic surface relief, pigment distributions, canvas weave, craquelure, or relationships among thousands of catalogued images. That makes it a useful instrument for triage and research. It does not make the model an authenticator, establish lawful ownership, or produce a defensible market value by itself.
The practical opportunity in 2026 is therefore narrower—and more valuable—than an “AI art detective.” Museums, conservators, provenance researchers, appraisers, insurers, and auction specialists can use computation to organize evidence, expose inconsistencies, and decide which expert examination should happen next. The decision remains an interdisciplinary judgment tied to a specific object, question, jurisdiction, and valuation date.
“Is this genuine?” hides several different questions. A curator may ask whether an object belongs in an artist’s catalogue raisonné. A conservator may need to distinguish original paint from a later repair. A title researcher may investigate whether a transfer was lawful. An insurer may need replacement value, while a tax authority may require fair market value on a defined date.
Each question needs different evidence and a different accountable professional. A useful intake record identifies the object, owner or custodian, purpose, effective date, legal context, permitted tests, conflicts of interest, and the person authorized to issue the conclusion. An algorithmic score belongs inside that record as one observation, never as the assignment itself.
This division of labor is consistent with the American Institute for Conservation’s professional guidance: examination and testing require justification, sampling needs consent and documentation, and declarations about age, origin, or authenticity should rest on sound evidence. It also warns about conflicts among conservation, authentication, appraisal, and dealing.
Before model inference, create a stable evidence packet. It should include calibrated visible-light photographs, scale and color references, dimensions, medium, support, labels and inscriptions, frame and backing, condition, treatment history, acquisition documents, known exhibitions, publications, prior opinions, and every unresolved gap. Record who captured each item, when, with what device, and under what lighting.
That discipline matters because digital twins can silently become cleaner than the object. Cropping may remove a suspicious edge. White-balance correction can change pigment appearance. Restoration may dominate a scan. Compression can invent texture. A reproducible packet keeps the untouched capture, derivative files, processing parameters, hashes, and operator notes. The broader workflow resembles evidence-first AI assurance: a reviewer must be able to reconstruct what the system saw and why a flag was raised.
Provenance is the history of ownership and transfer, not merely a seller’s typed list. The National Gallery of Art’s provenance guidance describes former owners, transfer dates, and transaction types as part of that history. A production system should keep assertions separate from supporting records: invoices, dealer stock books, collection catalogues, customs records, archival photographs, exhibition labels, correspondence, and loss databases.
AI can accelerate entity resolution across spelling variants, languages, dates, dealer names, and handwritten records. It can suggest that “Mme. X” in one ledger may match a named collector elsewhere. It cannot turn an inferred match into a documented transfer. Every proposed link needs a confidence label, source citation, contradiction field, and human disposition. Unknown years must remain unknown; a smooth generated narrative is not provenance.
The revised ICOM Code of Ethics for Museums adopted in June 2026 makes the institutional boundary especially clear: museums should exercise due diligence over origin and complete history, maintain secure documentation, consult relevant expertise, and use technology without misrepresenting information. Computation supports that duty; it does not discharge it.
Visible, raking, ultraviolet, infrared, X-radiographic, hyperspectral, X-ray fluorescence, and surface-topography methods reveal different aspects of an object. Registration software can align these modalities; segmentation can highlight retouching, underdrawing, seams, repairs, or anomalous elemental regions; anomaly models can rank areas for a conservator to inspect.
But an anomaly is not a forgery. A later varnish, historical restoration, workshop collaboration, reused canvas, artist’s revision, or sensor artifact can all create departures from a reference set. Imaging conclusions should state the acquisition conditions, spatial resolution, preprocessing, reference materials, known blind spots, and plausible alternative explanations. If sampling is proposed, the owner’s consent, minimum material removed, chain of custody, and retained sample must be documented.
Computer vision can compare color, edge, composition, or brushwork. Surface profilometry adds height information at scales smaller than an entire stroke. A primary study, “Discerning the painter’s hand”, trained convolutional networks on profilometry from controlled paintings made by art students and reported accuracy ranging from 60% to 96% depending on patch size and comparison. The paper explicitly framed this as a controlled study and noted that aging and conservation would challenge real-world use.
That is promising evidence for a measurement technique, not a universal authorship test. Operational validation must ask: Were reference works securely attributed? Do training and test images come from different capture sessions? Can the model distinguish artist from scanner, varnish, studio material, or conservator? How does it behave on workshops, assistants, copies, and changed technique over a long career? A result without a representative reference population is a similarity score with uncertain meaning.
Pigment and binder analysis can reveal that a material was unavailable in the claimed period, or that a sampled layer is later than another layer. Databases and classifiers can help match spectra, prioritize candidate compounds, and compare stratigraphies. Negative evidence can be powerful when the sampling location and historical availability are well established.
Positive identification is more limited. A period-appropriate pigment was available to many artists and to later forgers. Contamination, conservation materials, and mixed layers complicate interpretation. Reports should distinguish “identified,” “consistent with,” “not detected,” and “inconclusive,” and should record detection limits. The safest product is a material chronology with uncertainty, not a generated artist name.
An attribution panel should see separate lanes of evidence: provenance, connoisseurship, technical imaging, materials, documentary scholarship, and computational comparison. Do not collapse them into one opaque percentage. A 92% model score can dominate discussion even when the model has no calibration for the object’s period.
Instead, record each observation, its source, reliability, dependencies, counterevidence, and reviewer. Invite a red-team reading: What ordinary conservation history could explain the anomaly? Which evidence was selected after seeing the desired answer? Are reference works themselves disputed? What would change the conclusion? “Attributed to,” “studio of,” “circle of,” “after,” and “unresolved” are meaningful scholarly categories; a binary genuine/fake interface destroys information.
Value is not an intrinsic visual feature. It depends on attribution, condition, quality, dimensions, subject, rarity, provenance, legal restrictions, market, sale venue, effective date, and purpose. A model trained on auction results may be useful for locating comparables or detecting inconsistent catalogue fields. It does not know undisclosed private sales, guarantees, buyer’s premium treatment, condition differences, or whether the market segment is thin.
For a concrete regulatory example, the IRS Art Appraisal Services guidance requires a qualified appraiser and, for relevant U.S. federal tax assignments, information such as condition, acquisition history, proof of authenticity, valuation date, method, and specific comparable transactions. The IRS also uses trained appraisers and an advisory panel for certain cases. This is jurisdiction- and purpose-specific, but it illustrates the general rule: computational pricing is analysis for an appraiser, not the signed opinion.
A defensible workflow can be organized as:
Stop the workflow when object identity is uncertain, consent is missing, provenance may implicate theft or illicit export, reference data are unsuitable, a material result cannot be replicated, or a reviewer has a financial conflict. Speed is not a benefit if the workflow makes an irreversible acquisition, deaccession, publication, or market statement on weak evidence.
Model accuracy on a convenient image benchmark is not enough. Useful operational metrics include false-negative rates on known interventions, calibration by period and medium, out-of-distribution detection, inter-reviewer agreement, number of source-backed provenance links, percentage of flags resolved, time to retrieve evidence, reproducibility across capture devices, and rate of conclusions changed after expert review.
Track harm as well: unnecessary sampling, misattribution corrections, legal challenges, privacy exposure in ownership records, and cases where a model score leaked into a catalogue or sale description before approval. For triage, measure whether the system directs scarce expert time toward genuinely informative examinations. The target is a better evidence process, not a higher volume of verdicts.
Begin with search, duplicate detection, image registration, condition-map comparison, provenance-name suggestions, and comparable-sale retrieval. Validate on closed historical cases where ground truth and expert notes can be examined. Next, run prospectively in shadow mode: experts work normally, the system records recommendations, and a review board studies agreement and failure modes.
Only after calibration should outputs enter acquisition, lending, insurance, tax, or publication workflows—and then behind named human approval. Maintain a model card by medium and period, a reference-set register, capture standards, incident procedures, and a route for outside scholars or owners to challenge records. For museum-scale implementation, AI in cultural preservation provides the wider collections and stewardship context.
Computational evidence can show that an object resembles or differs from a reference population under defined measurements. It can expose a provenance inconsistency, locate a material anomaly, or rank comparable sales. It cannot, on its own, establish authorship, lawful title, ethical acquisition, or value.
The credible “digital connoisseur” is therefore not a machine that replaces connoisseurship. It is a controlled evidence system that helps conservators, historians, provenance researchers, appraisers, and legal specialists see more, document better, and disagree explicitly. Its most valuable output may be “the evidence is insufficient”—provided the record shows exactly why.
Reviewed 2026-07-30. The professional and governance boundaries use the 2026 ICOM Code, the AIC Code and Guidelines, the National Gallery of Art’s provenance definition, and IRS Art Appraisal Services. The brushwork example is grounded in the peer-reviewed surface-topography study. The IRS material is a U.S. federal tax example, not a universal appraisal rule; the controlled profilometry results are not presented as real-world authentication accuracy.
AI can improve access to archives while provenance, custody, integrity, and evidential status remain explicit—and every generated description stays attributable.
Read More
Font-generation AI can explore glyphs and spacing, but usable releases require correct encoding, script shaping, legibility, accessibility, rights, and testing.
Read More
A safety-first framework for matchmaking systems that separates candidate generation, ranking, moderation, consent, and relationship outcomes.
Read MoreIf this note maps to a real system in your organization, start with the services page or a shipped case study.