Lev Manovich
“History is the new continent. There is no Terra Incognita anymore. The last one is history.” Nicolas Bourriaud, 2007
In the following short text, I will discuss the emerging use of AI as a “memory machine” from two perspectives. One is a history of modern media technologies and their use in society. The second is my personal desire to visualize (simulate, synthesize, construct) my memories of growing up in Russia in 1960s and 1970s.
The title of the 1997 book by cultural historian Norman Klein perfectly summarizes the key process of social spatial memory: The History of Forgetting: Los Angeles and the Erasure of Memory. The past is continuously erased. And later its selected bits are brought back as curated and museum-quality images and experiences. The traces of originally poor neighborhoods, of immigrant communities, of the original inhabitants of the area (such as Native-Americans living on the territory of what later became Manhattan) are covered over. Business districts, residential communities, and hip urban areas are constructed over the past? And later the past is brought back as museums or re-synthesized neighborhoods with names such as “Little Italy,” “Little Tokyo” or “Chinatown.” And what about our personal individual memory? Does it function in the same way?
Our memories are very selective: we remember some things and not others. Do we actually remember these things as we experienced them at the time? Or does our consciousness re-synthesizes them as museum exhibits or scenes from Hollywood movies – dramatized, theatrical, exaggerated – foregrounding some details and omitting many others?
Our memories also often lack visual precision and details. They are very “low-resolution”, so to speak. For example, can I visualize precisely my childhood apartment in Moscow, where I grew up in 1960s? Or New York City in 1981, after I arrived there as immigrant? The latter is much easier – because more media materials about New York (digitized photographs, documents, newspapers, film documentaries, artworks, etc.) are easily available on the web. I can also consult English Wikipedia, which contains endless details about this period in New York, across its millions of entries.
But Moscow in 1968? Significantly less visual documentation is available online. Not as many people in Russia owned cameras at that time – and amateur photographs in general preferred to capture important moments, such as birthdays and weddings, rather than the everyday and the mundane. When I search “Moscow 1968” on Google, most results are either official Soviet photographs or images from foreign press – and most of them show the Kremlin, the Red Square and a few other iconic sites and subjects. So while search results for “New York 1981” generate enough images of everyday life, the results for “Moscow 1981” are more likely to represent only the limited official iconology. (I went through many pages of Google Images search results and could not find a single personal non-official photograph.)
All society’s collective memories are shaped by media machines, and since their introduction in the 1820s-1830s (e.g., photography), our media records have been growing. Each new machine multiplied them: sound recording (Edison’s phonograph, 1877), film recording (Edison’s Kinetoscope,1891), video tape recording (1951), lidar (1971), body motion capture (1973), face motion capture (1990), and so on. In parallel, many new types of media storage and dissemination machines were also invented and popularized. Internet becomes a new massive depository of human experiences after January 1993, when the first graphic web browser NSCA was released, and a series of social media platform was developed soon thereafter (GeoCities, 1994; SixDegrees, 1997; Cyword, 1999; Facebook, 2004; Flickr, 2004; Instagram, 2010, and so on). And now, because of the proliferation of mobile phones with high-resolution cameras, the amounts of images that document the last few years is astronomical in comparison to what we had in previous decades and centuries.
In the last fifteen years, new computational memory technologies become gradually more important. Researchers in many fields began utilizing data science methods to study patterns in massive text, image and sound archives. Development of AI gives rise to yet another set of new memory techniques. On YouTube, you can find many short videos of city life from the 1900s that have been enchanted by AI–increasing resolution, sharpening details, and adding color. So now we can see more details in the Lumière Brothers’ 50 second-long “Arrival of a Train” than this film first viewers in 1896.
In 2022, yet a new generation of memory media devices became available for everybody. I am referring, of course, to generative AI and tools such as Midjourney, Stable Diffusion, Dall-E and others. These tools employ artificial neural networks trained on datasets containing billions of images and their text descriptions. When a user provides a new description (i.e. a prompt), the network synthesizes images that more or less correspond to this description.
It is important to note that these new images did not exist before, and they are not part of the training data. They are new hybrids that can combine a representation of certain content with a style of a particular artist or a look of a particular medium: from pencil drawing and various old photographic processes to the looks of specific professional cameras and lenses, both historical and contemporary, and three-dimensional photo-realistic CG rendering.
I can write a short text prompt which makes such software visualize “Moscow in 1968,” “New York in 1981,” or any other space and time, and receive many detailed images within minutes. But how reliable are these synthetic historical representations? Their precision partly depends on how much visual documentation exists online about a particular topic or a time period, since the billions of image-text pairs used for training a neural network come from the web. But even in the cases of well-documented subjects, this software has strong biases. For example, when I ask for “Moscow,” the images generated by Midjourney–at least in 2022–inevitably included churches or other historical 19th century buildings; as opposed to, for example, modern buildings constructed after 1960.
So while the new AI memory machines can offer unique opportunities for any individual to “see” into both her personal past and the collective historical past, the images they can generate so far are unreliable. Of course, the same can be said about any historical representation which always represents particular perspectives and shows some things at the expense of others. And this is not necessarily bad. In literature and in cinema, the representation of the events, people and details of reality from a particular subjective perspective is what we find pleasurable. Many modernist writers further expand this by, for example, including narrations of the same events from the perspectives of multiple characters in the same work.
However, until now we clearly differentiated between literary, cinematic, pictorial and visual fiction, and other types of representations that at least aim at objectivity–such as maps or academic articles and books about this or that historical periods. This fundamental distinction is now being questioned by the adoption of generative AI. It is a new type of modern media and a new type of memory machine. It effortlessly and endlessly generates new representations that seamlessly blend facts and hallucinations, reality and its interpolations. For AI generative media, history is indeed “the new continent,” to use the phrase from Nicolas Bourriaud. As an artist, I enjoy exploring the possibilities of AI’s “unreliable memories” – and I hope that you will also find my works in the exhibition to be interesting and provocative.