Fraudsters are using documents produced by artificial intelligence to “prove” the authenticity and ownership of artworks when obtaining appraisals or making insurance claims, according to reporting by the Financial Times and subsequent industry coverage. Chatbots and large language models are being used to generate sales invoices, appraisals, provenance documents, and certificates of authenticity — adding a new dimension to the art market’s oldest problem: fakes.
How the schemes surface
Olivia Eccleston, a fine-art insurance broker at Marsh, has described AI-generated paperwork as a growing feature of fraudulent claims. In one case reported to the FT, a fine-art loss adjuster reviewing a claim on a collection of decorative paintings received dozens of appraisal certificates that initially looked credible — until it emerged that the description fields were identical across different works, suggesting the certificates had been produced by an automated writing system.
Industry specialists distinguish between two failure modes. Some AI use is plainly malicious: deliberate forgery of provenance documents, the chain-of-ownership records that give an artwork its market value. But other cases begin as honest mistakes — collectors ask AI models to find references to their artwork in historical databases, and the model fabricates or garbles results. Angelina Giovani of provenance research firm Flynn & Giovani has noted that language models are prone to inventing answers when pressed: given enough prompting, a model will guess, and the guess reads like a finding. She has also encountered a document in which AI appeared to have been used to apply a signature to a painting.
An old crime with new tools
Experts stress that document fraud is not new to the art world — forgers have always counterfeited not just paintings but the paperwork that accompanies them, without which works can be close to worthless. Fake ledger numbers and forged wartime stamps have appeared on provenance documents for decades. Wolfgang Beltracchi, the German forger who produced hundreds of works passed off as artists including Max Ernst and Fernand Léger, famously staged fake vintage photographs to manufacture provenance for his paintings.
What has changed is the cost and quality of the forgery. Where fraudsters once had to steal or mock up letterheads, AI now produces realistic documents in seconds. Filippo Guerrini-Maraldi, head of fine art at insurer Howden, told the FT that forged documents have long been part of the landscape — but AI has made them markedly more realistic. Harry Smith, executive chairman of appraisers Gurr Johns, put it simply: AI makes long-standing fraud a little easier and a little faster, removing even the need to invent a plausible-sounding expert.
The detection arms race
Insurers and their loss adjusters are responding in kind. Grace Best-Devereaux, a fine-art loss adjuster at claims specialist Sedgwick, examines metadata on digitally filed documents for signs of AI involvement, and adjusters are themselves using AI tools to help authenticate provenance paperwork. But she warns the window may be closing: recent improvements in generative models are making fraudulent text harder to flag even for trained eyes, to the point where “the text looks wrong” may soon stop being a usable signal.
Limitations and what to watch
The scale of the problem is hard to quantify — the reporting is based on practitioner accounts rather than systematic industry statistics, and no reliable figures yet exist for how much AI-assisted document fraud costs insurers or collectors annually. The distinction between deliberate forgery and AI “hallucination” also complicates enforcement, since a collector who submits confabulated provenance may not have intended fraud at all. Buyers and insurers can reduce exposure by insisting on independently verifiable provenance — original documents, checkable archive references, and expert examination of the physical work — rather than digital paperwork alone. Coverage of the underlying story is available from CSO Online and Futurism.
Related reading on this site: how AI behaves differently when it knows it is being watched.
