Today, the digitization of museum archives is in full swing. It is increasingly rare to come across an institution that has not undergone this transformation. This process has its advantages, but when the result is put online, it is often the lack of precaution that allows “scrapping” algorithms to build “datasets” of images intended to feed artificial intelligence. These algorithms can be written by researchers, companies, or artists. Like the artist Rafik Anadol, for example, or Google with its artificial intelligence called “Imagen” a direct competitor of “Dall-E 2” created by Elon Musk's company, Open AI.

Although the images produced by these artificial intelligences seem to us to be original creations straight out of the overflowing imagination of these “intelligences”. These pieces of software work thanks to a resource hoarding done beforehand, meticulously sorted by humans. It is thus from fragments cut out of heaps of pixels recovered by these algorithms that they produce their work. An AI is thus destined to project the past that it has previously ingested into the future. Nothing is lost, everything is transformed. Although, like a compression algorithm, this process is not without loss. But we will come back to this.

In the process of developing or training and AI, it is possible to support or help it. For example, concerning the distinction of the elements, or the orientation of the result of its production, like the project “Botto” by the German media artist Mario Klingemann. We will speak here of supervised or unsupervised training. Ask a curator or an average person to set up an exhibition, the result will not be identical. The knowledge accumulated by the first is invaluable, and will allow him or her to carry out the task and delight the public. For the other, things will be more complicated, although the result will not necessarily be less interesting. Two choices can be made here. Help your friend to set up the exhibition, by giving instructions on the choice of works, providing informing on why and how, or leave your friend in complete autonomy.

Art critics as well as the artists themselves make a clear distinction between AIs whose learning have been supervised and the others, non-guided learning. Norah Khan, a leading writer and curator of artworks created through artificial intelligence, believes that AIs that have not been supervised during their training produce more fascinating images. For my part, I make the same observation, with one distinction: I believe that it is important to conceive the stage of constitution of the “dataset”, the curation of the images (obtained using scrapping algorithms) as the object of a beginning of supervision. Indeed, the artist who proceeds to a predefined image curation results on the identification of a certain aesthetic direction, categorization clear in the type of chosen medium (sculpture, paintings, drawing) or in the stylistic form (cubism, pointillism, mannerism, etc.). Already presents a form of supervision, even if the AI is left in autonomy through its learning. With this distinction, it is thus ninety-nine percent of the AIs that are supervised. Non-supervision is therefore an ideal form that is rarely reached.

The museums, by this digital transposition, find themselves dissolved of their substances. Without supervision, and therefore, the help of human beings to value the different elements recovered by these algorithms. Museums are doomed to produce totally decontextualized “soft images” on which anyone can project whatever they want. The project “Unsupervised—Machine Hallucinations” by Rafik Anadol is a perfect illustration of it.

In “Unsupervised Meanings”, he museum was scanned from top to bottom using photogrammetry, thereafter it was reconstructed in a 3D environment. In parallel, a scrapping algorithm designed by the participants in the project recovered part of the museum's digital archives. An almost perfect mirror of its physicality, everything is preserved. Works, flooring, signage, light, benches, grids, ventilation. All preserved in the wall textures of our replica. These ingredients are then ingested by Artificial Intelligence designed for the project, whose only mission is to paint the walls of our meta-space.