Preserving our digital memory with a personalized AI twin could be the future.
Recently, I rediscovered my passion for photography after a decade-long hiatus of taking casual iPhone photos. My previous enthusiast camera, a Nikon D200 made in 2005, still sat quietly in my storage container. As I was deciding whether to invest in another camera for an upcoming trip to Rwanda, where I hoped to capture stunning wildlife images, I pondered the role of photography and personal storage, and how they might evolve in the future.
In today’s world, owning a traditional camera may seem outdated. Most people rely on their smartphones, which now boast high-resolution cameras and advanced AI algorithms for exceptional results. Over the past several decades, for most people cameras have shifted from being tools of artistic expression to lifelong recording devices. We capture meals at fancy restaurants and picturesque views on our trips. We hand our phones to strangers to take photos for us, without worrying about their photography skills. Personally, I have amassed over 50,000 photos on my phone, rarely revisiting them. They serve as a visual diary of my daily life.
Unlike computer memory, our memories are not accessed by searching for specific dates or addresses. Instead, our memories are stored through association. We collect travel souvenirs not for their material value, but for their ability to evoke memories of the moments before and after we acquired them. Travel photos function in a similar way as visual souvenirs. They don’t have to be perfectly captured; they serve as triggers that activate the part of our brain storing memories, allowing us to relive our journeys.
Throughout history, our recording media have continually evolved. In the past, people carried diaries to chronicle their lives. In the 20th century, traveling with a camera and films was the standard. Today, DSLRs have become a thing of the past and cell phones are everyone’s pocket camera. With Apple’s new Vision Pro, we’re about to step into the world of 3D spatial photography and videography. Imagine being able to fully immerse yourself in a past event, like your best friend’s wedding, your child’s birthday party, or a casual conversation with a loved one. If the animated portraits in Harry Potter amazed you as a kid, immersive photography can be an unimaginable experience of the future.
Beyond these technological advancements, however, there is an often overlooked question: How do we store and manage our digital memories? It has been nearly two decades since the widespread adoption of digital photography, but the lifespan of a typical hard disk is only around 10 years. This means that many of us likely own broken hard disks or have lost digital files. Every digital storage medium has an expiration date. Hard disks last 5 to 10 years, SSDs 3 to 5 years. Opticals disks with special coatings made of precious metal can last from several decades to a couple centuries, but their storage capacity and read/write speed make them much less attractive. Preserving data is not an easy job. Data professionals follow the 3-2-1 rule: three copies of data on two different storage media, with at least one copy stored in a different location, and refreshing all data copies every 3-5 years.
As managing your own data is so painstaking, many people have turned to cloud storage solutions. Back in the 2000s and 2010s, the prevailing sentiment was that everything on the internet should be free. It was only in 2019 that I purchased my first storage plan with iCloud. As file sizes increase with raw camera files and 4K videos, I find myself constantly upgrading my Google Drive and iCloud subscriptions for more storage space. However, cloud storage does not offer a permanent solution. Ideally, we would want our data to outlive us, but who will maintain it centuries later? As more individuals spend their life in the digital realm, it becomes crucial to consider how to preserve personal data in a sustainable way.
In some jurisdictions, data is treated as an asset that can be included in one’s estate. However, people may be uncomfortable handing over their entire photo or social media archive to their immediate family. Moreover, future generations might forget to renew expensive iCloud subscriptions that cost hundreds of dollars annually. Establishing a dynasty trust to cover subscription fees is an option, but one may question the purpose of preserving data if it cannot be accessed, except perhaps for exhibition in a museum thousands of years later.
An intriguing solution is the concept of creating a personalized AI twin, an AI that shares our life experiences. Existing AI models like ChatGPT and DALL-E learn from vast amounts of internet data. However, instead of relying on this shared pool of information, imagine if AIs were trained using your personal data—your unique perspective of the world and the words you have written or spoken before.
Similar to how Google indexes the entire internet for searchability, AI could serve as a tool for organizing personal memories. It could answer questions related to your memories, such as identifying a restaurant you visited or recalling where you last met your best friend. As AI becomes more personalized, a critical technical challenge lies in enabling AI to learn with less data and more personalized information. This is an area that my students, collaborators, and I are actively researching. Personalized AI twins would process your data locally, ensuring privacy, rather than uploading it to a public server.
The more life moments you record, the more closely your AI twin will resemble your way of perceiving and thinking. It may express thoughts and make decisions similar to yours in comparable situations. While we are alive, it can serve as a memory aid or even act as a substitute when we are preoccupied. And when we eventually pass away, it could retain a digital copy of our essence, continuing to interact with the world as if we were still present.
The AI twin’s interaction with the world strikes an interesting balance between privacy and accessibility. Standard data encryption either allows complete access to the entire archive or scrambles the data beyond comprehension, leaving no easy way to selectively share partial or obscure information. In contrast, an AI twin can provide answers with varying levels of detail and privacy, depending on the recipient. People can query specific information without accessing the entire archive. Additionally, AI enables more efficient data compression, making it a viable option for long-term storage compared to storing complete video footage.
If an AI twin is trained solely on a person’s lifelong recordings, it may even replicate aspects of consciousness, at least partially. The question of what consciousness truly is has boggled philosophers and scientists for centuries. In the past, we lacked the computational tools to simulate the process of learning from an egocentric perspective, which we are starting to have today. Researchers have collected head mounted camera footage from infants, and my lab is currently investigating the possibility of training AI using lifelong visual streams as training data. If successful, this research may shed light on the emergence of our perception of the world. In the wildest of dreams, AI twins could pave the way for a sci-fi future where consciousness can be uploaded and digital humans exist.
As we step into uncharted territory, ethical questions may arise. Should we train AI using personal data? Should everyone have the ability to create their own AI twin if the technology becomes available? Should AI twins be allowed to make decisions beyond answering memory-related questions? In the past, debates focused on the morality of human cloning, but the AI twin, rather than cloning the physical body and DNA, acts as a virtual clone of our memory, and potentially our mind. In the near future, I envision it as a valuable memory tool that resides in our phones, AR glasses, and computers. However, in the distant future, as AI twins assume more tasks, the boundaries of our identities may become blurred. We must exercise caution, especially when allowing AI twins to take actions in the physical world. Compared to a centralized AI that relies on all human knowledge for training, a reality of today’s large language models, a personalized AI twin that draws from limited training data may pose less risk of overpowering individuals.
Having an AI twin may seem like a wild speculation, but it could be right around the corner. My digital album holds its first dated photo from twenty years ago—a snapshot of my visit to Victoria in British Columbia when I was a young child, holding a Canon point-and-shoot camera loaded with Fujifilm. Little did I know then that photography would undergo such a remarkable evolution. Nevertheless, our innate instinct to record rare and memorable moments has always been with us. Perhaps these moments define who we are, and are worth preserving indefinitely.