On 27 January, the British Museum published on Instagram and Facebook a series of images of a young woman contemplating pieces in its galleries. The accompanying text invited one to pause: looking with calm is always worthwhile. The images were generated with artificial intelligence, and this was not stated in the caption. They were tagged: for those who went to look at the tags, the account of a generated model and that of the agency that had produced them. Six hours later, following criticism from archaeologists and heritage professionals, the museum withdrew them. A spokesperson stated to the specialist press that they do not publish images created with AI and that, acknowledging their potential sensitivity, they had removed them. They added that they were developing internal guidelines regarding the use of these tools.
I shall begin by declaring my own position, for without doing so I would have no right to write this text. The images that head my articles are generated using artificial intelligence. I do so with a prompt of my own design that I have been refining for years, and I have explained this on a page of my website where I detail what I use it for and what I do not. I do not believe that using these tools is illegitimate, nor shall I feign a modesty here that I do not possess. I am interested in something else, and it is not what occupied the discussion.
The discussion concerned whether a museum ought to use generated images. Steph Black, an archaeologist, maintained that the institution was testing the extent to which the public accepts these images in order to subsequently avoid hiring creative professionals, and she attributed this to cost-cutting. That is her interpretation, and as a suspicion it is reasonable within a precarious sector, but it is unproven and I do not subscribe to it. Her other request seems to me much more robust and less cited: that the museum explain what had occurred and who authorised it.
The problem is not the image. It is that it did not state its origin.
It is not merely some incidental error. The authority of a museum such as the British Museum rests entirely upon one operation: attesting to the provenance of objects. Everything contained within that building is accompanied by a documentary chain that explains where it originated, who created it, and where it has been. Without that chain, an object is not a museum piece: it is an object. Provenance is not an administrative requirement of the institution; it is literally that which the institution sells, and in its case, it is also the area where it is most scrutinised, because a significant portion of its collection is in dispute precisely due to how it arrived there.
Thus, we have an institution whose function is to accredit the origin of what it displays, publishing an image without accrediting its origin. It is not hypocrisy, and I would not call it that: it is an asymmetry, and I believe a very revealing one. What is displayed inside is subjected to an extremely high standard of provenance. What is published outside, through the channel by which the institution speaks to the world, is governed by marketing standards. As if communication were an exempt zone.
This is not an isolated case, and there is a figure that measures it. In a survey of more than three hundred gallery professionals conducted in February of this year, sixty-one per cent responded that none of their artists use artificial intelligence in their practice. Fifty-seven per cent responded that they themselves use it to draft their emails and communications. Read the two figures together. Almost the same institutions that declare that there is no AI in their art employ it to speak about that art. There is no deception in that, and likely none of them have considered it a contradiction. It describes with precision where this technology has entered the art world: not through the work, but through the voice.
The precedent that is usually cited is that of Christie’s, which in February 2025 announced the first auction composed entirely of work created with AI. More than three thousand people signed an open letter calling for its cancellation, arguing that the models employed exploited human artists by using their work without permission or payment. The auction was not cancelled and closed at just over seven hundred and twenty-eight thousand dollars, a modest figure for what was billed as a historic event. There is a difference between the two cases, and it is not minor. Christie’s did something debatable and did so openly, by announcing it. The British Museum did something less debatable and did not state it. It is the latter that erodes trust, not the former, because what is discovered afterwards carries more weight than what is declared beforehand.
As of today, 2 August 2026, the transparency obligations of the European Artificial Intelligence Act are in force: certain generated or manipulated content must be labelled in a clear and visible manner and carry machine-readable marks. The question of whether it is advisable to declare it ceases to be a matter of institutional style and takes on a legal answer. And beneath the regulation, there is already infrastructure to comply with it: since 2021, there has existed a technical standard for digital content provenance, promoted by a coalition that includes Adobe, the BBC, Microsoft, and Intel, whose credentials function, according to their own description, like the nutritional label on a food product. They record what was used to create an image and what was done to it subsequently. It is exactly what a museum has been doing with objects for centuries, applied to files.
The fact that the sector has not adopted it is telling. There are recommendations—the European museum network was already calling in 2024 for the maintenance of integrity in the use of AI and highlighting the importance of trust and transparency, and this year UNESCO and ICOM have launched a global consultation to prepare a joint report—but there are no labelling guides as yet, and there is no evidence that any major museum has publicly adopted the technical standard that already exists. The principle is being debated while the tool is available.
A museum can afford many things, but there is one it cannot: for its public to discover that something of its own was not what it seemed. Not because it has lied, but because its entire profession consists of one being able to trust what one is told regarding the origin of things. Each time such an institution publishes something without declaring where it comes from, it expends a little of the only thing it cannot replenish. Declaring it costs nothing. It is a line of text at the foot of the page. And in an institution that lives by provenance, that line is not a mere formality: it is consistency with what is done in the galleries.
On the open conversation
This text continues my interest in authenticity, understood as a documented chain rather than an inherent property of an object, and extends it to the sphere where cultural institutions discuss themselves. I do not maintain that the use of generated imagery is illegitimate; I use my own and declare it on my website. I maintain that an institution whose authority rests upon the accreditation of provenance cannot afford to publish without accrediting it. If anyone wishes to intervene from the perspectives of museum communication, conservation, professional illustration, or regulatory compliance, this notebook remains open.
Sources
Artnet News, 2 February 2026, and ARTnews, 3 February 2026: coverage of the British Museum episode (publication on 27 January 2026, removed the same day) and statements from the museum spokesperson.
The Art Newspaper, 10 February 2025: open letter with over 3,000 signatures against Christie’s 'Augmented Intelligence' auction (announced 7 February 2025; held 20 February to 5 March 2025; results published by Christie’s).
NEMO (Network of European Museum Organisations). Policy recommendations on artificial intelligence and museums, 20 March 2024.
UNESCO and ICOM. Global survey on the use of artificial intelligence in museums, launched 20 May 2026.
European Commission. Transparency obligations of the Artificial Intelligence Act in force since 2 August 2026, and code of practice on the transparency of generated content.
Coalition for Content Provenance and Authenticity (C2PA), established 22 February 2021; Content Authenticity Initiative (Adobe, 2019).
Artsy. The Artsy AI Survey 2026, 18 March 2026 (more than 300 gallery professionals surveyed in February 2026).