Can I use AI to visualize my own personal experiences from decades ago? For this series, I have asked an AI tool to generate a scene showing people in a bus in Moscow in the 1970s. This were some of my prompts:
"tired people inside a bus in Moscow, photo slightly from above, February 9am ambient soft light coming from bus window, ultra-detailed photo, proportional bodies, f70 lens --stylize 900"
"people of different ages inside a bus in Moscow, photo slightly from above, February 9am ambient soft light coming from bus window, ultra-detailed photo, proportional bodies, f70 lens --stylize 900"
After selecting the best outputs, I have slightly adjusted brightness, sharpness and color balance of images in Lightroom, and made some areas darker or lighter - but did not change any of the content AI generated.
Our own personal memories are very selective (we remember some things and not others), and they often lack visual precision and details. Can I visualize precisely today the room where I grew up in my parents Moscow apartment in the 1960s? Or New York City in 1981 after I arrived there as an immigrant? The latter is much easier - because more media materials about New York (photographs, various documents, digitized newspapers, film documentaries, etc.) are easily available online. And I can also consult English Wikipedia which contains endless details about this period in New York across its millions of articles.
But Moscow in 1977? There is significantly less visual documentation available on the web. Not as many people owned cameras - and amateur photographs preferred to capture important moments such as birthdays and weddings rather than the everyday and the ordinary.
Society’s collective memories are shaped by media machines. Since the beginnings of modern visual media in the 1830s (e.g, photography), our media records have been rapidly growing. After photography, we get films and sound recordings (1890s-), TV programs stored on video tapes (1950s-) and later other capture technologies that rely on computers such as 3D capture of bodies and faces and photogrammetry.
Internet became a new massive depository of human experiences after 1993 when the first graphic web was introduced, and is later joined by social media networks such as Facebook and Instagram.
Because of the proliferation of mobile phones with high-resolution cameras, the amount of images that document the last few years is astronomical in comparison to what we have from earlier history.
In the last 15 years, new computational memory technologies have become gradually more important. Scientists, humanists and journalists start using data science to uncover hidden patterns, relationships, and similarities in massive text and visual archives. Development of AI gives rise to yet another set of new memory techniques.
On YouTube there are many short documentary videos of cities from the first decades of cinema that have been enhanced by AI - increasing resolution, sharpening details, and adding color. Suddenly we can see more details in the Lumières’ “Arrival of a Train” than the first viewers of these first short films saw in 1896.
In 2022, new media memory machines became available to everybody. These are Generative AI tools such as Midjourney, Stable Diffusion, Dall-e 2 and 3, and so on. etc. So now I can compose a short text prompt which will make such software to visualize Moscow in 1968, New York in 1981 or any other space and time, and receive detailed images within minutes.
But how reliable are these synthetic memories? Their precision depends on how much visual documentation exists online about any particular topic or time period. But even in the cases of well-documented subjects, this software has strong biases. For example, when I ask for “Moscow,” the output images generated by Midjourney in 2022 inevitably include churches or other historical buildings. So while this new memory machine can offer unique opportunities for any individual to “see” into both her personal past and the collective historical past, the generated “memories” so far are unreliable - they tend to substitute generic and stereotypic for the particular and unique.