A post on Twitter with images of “reconstructions of historical figures”, including the alleged face of a Guanche mummy, made with artificial intelligence, has generated much controversy. So much so that a week later the author of the tweet republished them, but this time explaining that it was an “unreliable” reconstruction. He had used Midjourney, one of the platforms capable of creating an image from a sentence, and some archaeologists and historians raised the red flag to warn of the risks they pose to cultural knowledge. They say that reconstructions of heritage made with artificial intelligence (AI) create fakes historical, reduce scientific work to a click, reinforce stereotypes and haggle over copyright.
How to make forensically NOT reliable facial reconstructions with #midjourney.
I will explain to you in detail the easy process and all the prompts used in my most controversial thread of 2023 (so far) so you can experiment too.
Here it goes ???????? pic.twitter.com/YfV45BB9HU
— Javi Lopez ⛩️ (@javilopen) January 15, 2023
Pablo Aparicio, owner of a company that performs virtual reconstructions in 3D or 2D of heritage, details that his work combines historical knowledge and anthropology, with different documentation techniques, to carry out a reliable reconstruction of architectural assets, works of art and archaeological sites. A single project can take several months of effort, even if several people work on it.
To rebuild a castle, for example, Aparicio explains that a laser scan or digital photogrammetry is first made of the remains of the construction itself. With a “perfect geometric base”, he considers the architectural hypothesis of the building with computer programs such as Blender or other 3D software. “We began to make the hypotheses in collaboration with architects and archaeologists who are marking all the steps for us. We show them the progress and they tell us, for example, if the walls would be higher or lower, if it looks different, where a roof is needed. They give us details of how we have to move forward”, affirms the founder of this company, 3D Stoa – Patrimonio y Tecnología. “It is a job that has to be very precise and concrete. We always have to attend to the modifications that those who entrust us with them propose to us ”, he continues.
The first problem with reconstructions made with generative AI, as he argues, is the lack of specificity. “As much as you ask in a prompt Well written (the command) that designs a castle for you in a certain way, it is impossible to give it all the details so that it does something precise for you”, adds Aparicio. The images are very striking, attractive, and work very well on social networks. At the end of the day, despite the beauty, it is a false representation, but with a feeling of credibility that leads to confusion. “They deliver an image that looks real or looks like it could be real. And therein lies the biggest problem. It helps a lot to transmit in networks the fakes historical, that have nothing scientific, but that may seem true”, says the archaeologist and historian.
The cause is the very nature of this type of technology, which uses unknown data banks as inputs, which are often anything that circulates on the Internet. The end result is a mix of everything online. “Artificial intelligences pull what is most widespread. They are not going to represent the Coliseum reconstructed or complete, as it was in antiquity, but they are going to represent it in ruins, as it is today, and that is also totally incorrect”, Aparicio emphasizes.
Alberto Venegas, professor of History and PhD from the University of Murcia, maintains that the main function of these tools is recombination, which is why they reaffirm stereotypes and perpetuate erroneous impressions about the past. “It’s not even remotely similar to the past, just a massive combination of all the images, regardless of source, intent or who made it. It is the perpetuation of myths or images that are already common places that we are very used to seeing in the media”, he underlines.
To exemplify this, Venegas cites the “impressive representation” of the city of Paris during the French Revolution. through the lens of a GoPro camera which also circulated successfully on networks. In his opinion, “any historian” would see that Paris in these images is not the Paris of the Revolution, but the Paris of the Second Empire and the Third Republic. However, in the collective imagination, there are elements that “sound of Paris”, such as slate roofs, chimneys, large doors, stone buildings and large avenues. “The Paris of the 18th century was not like that, it was later when a series of reforms were carried out in the city. In the same way, the current French flag comes out, which is not the same as that of that time, ”he adds.
“The AI has taken everything that can be plausible, that makes us think of the French Revolution, that we have seen on television, in video games, in series and movies. I call it the media past and aesthetic memory, which really makes us assume that time, but it really isn’t,” Venegas explains by phone.
Of course, the AI can be trained and its algorithm refined to improve the results obtained. This is the case of the work of Bas Korsten, who tried to teach an AI to replicate Rembrandt’s works until he was able to imitate them almost perfectly. pic.twitter.com/7ke4NlWjyj
— Alberto Venegas (@Albertoxvenegas) November 7, 2022
Would it be possible to use AI to credibly reconstruct heritage? Venegas suggests yes. “If they feed them the right data, they could come close to the reconstructions that the experts do,” he says. To exemplify it, the teacher remembers a project carried out by design executive Bas Korsten with the Delft University of Technology and the Rembrandthuis Museum in the Netherlands, which brought together 170,000 visual sources of the work of the Dutch painter Rembrandt. By processing this bank of images, the machine was able to generate an almost perfect work of art, following the style and brushwork of the artist.
the reconstruction intelligent It also generates controversy in the field of copyright, since it appropriates the work of thousands of photographers, illustrators, designers and artists. While the discussion of whether it is plagiarism is still open, the tools have the advantage of not being regulated.
Although Venegas and Aparicio criticize the reconstruction of heritage with this type of tool, both agree that they are useful in small stages of the traditional process, such as choosing surface textures and other finishes. Also, in educational training. “At school, at the institute or at the university, they could use them to carry out activities of reflection, criticism or to discuss different images, from different pasts, to draw conclusions. They are positive tools for the educational process”, concludes Venegas.