Are today’s deepfakes tomorrow’s real experiences?
In the past year many real brows have been furrowed over “deepfakes”. This deep learning technique can be used to generate video that can, for example, convincingly replace one person in a video with another. The concern is its potential for abuse: from deep-fake revenge porn to misrepresenting political adversaries, to impersonating loved ones asking for financial help, the possibilities are limitless. Some believe we are headed toward an era where we won’t be able to trust anything.
It may be premature, though, to assume deepfakes present a new and unique threat to our ability to discriminate between real and fake. If we take a step back, deepfakes are just the latest in a long history of technologies that have challenged our ability to make such distinctions. Laws against forgery, for example, go back at least to the Romans. Are deepfakes fundamentally different? Imagine for a moment showing the famed deepfake of Queen Elizabeth’s Christmas Greeting to, say, Ben Franklin, who was brought back to life just to let me make my point. “It’s fake,” we’d explain.
But what does “it” refer to? Is it that the queen isn’t really behind this “window” (i.e. the monitor)? That the events are not live, or may not have occurred in that order? That the frame of the “window” is not real wood? And finally, oh yes, maybe it’s that it’s not really the queen. To a bewildered Franklin the “deepfake” aspect is arguably among the least fake and confusing aspects of the video.
We modern day sophisticates, on the other hand, have learned not just to discriminate real from fake in a wide range of situations, but also to implicitly define what counts as real and fake in different contexts. We are actually fine with most “fake” things because we expect them. When we go to a movie, we know if it’s fiction or a documentary, and we also know it’s not live. We know whether it’s based on a true story, or a faithful recreation. Even within a fictional movie we will draw distinctions between “real” and CGI, and we may want to know if an actor did her own stunts.
Our expectations are set by long established social conventions. It is not generally seen as dishonest to wear makeup, or crop unpleasant features from photographs. A video of a magic trick is deemed “real” if it only involves deception performed by the magician. If it is done through video editing, it is “fake” — never mind that we all know the woman was not sawn in half in either case. Deception, in sum, is far less an issue of the technology in play than the reputation and intent of the content producer, how it is presented, and the established conventions for that situation.
While we won’t be as confused as Franklin, there’s something a bit deeper we’ll have shared with him. We’ll both have experienced the queen dancing and have a memory of the spectacle. Real or fake, it may leave an impression on us that can’t be undone in the same way as correcting a fact. This affective response is, to a great extent, beyond our control: “You can’t unsee that”, as the saying goes.
The ability of manufactured depictions to make an emotional impact is hardly unique to deepfakes. It is one of the hallmarks of art. What is new, however, is the ease with which deeply personal content can be generated. The point is not the possibility of deception, but rather of the possibility of a largely new class of highly personal experiences — experiences that carry both promise and risk.
Today a widow may look back at photos and videos of their lost spouse. Tomorrow the spouse may be added to videos of an event, engaging with the participants. After all, people are added and removed from photographs all the time today. It is not the technical ability that is provocative. It is the meaning and emotional reaction. On one hand “it’s fake”. On the other, is there anything more genuine than, say, seeing a grandparent congratulate their grandchild on their wedding day? Is it very different from reading a letter pre-written for the occasion, or showing a prerecorded video? Elderly infirm relatives present at the wedding might choose to have their younger selves deliver those comments.
Early examples of using deepfakes in this way exist today. The actor Val Kilmer lost his voice to throat cancer. Sonatic, a company using deepfake techniques, effectively “cloned” his voice, allowing Kilmer to have what he wants said expressed in his voice. Is that his real voice? Or is his post-cancer voice his real voice? Is it dishonest to present an audio clip of Val Kilmer saying something in his cloned voice — even if he authorized the statement? Presented appropriately, people will not be misled into thinking he has regained his voice. But now, for example, his children can hear him say he loves them in his original voice.
The result is potentially a profoundly visceral reaction that is impervious to the fact that a deepfake is involved. These responses are operating in the realm of biology, not epistemology. This would be much like the memories evoked by a scent or a song. What makes these memories so powerful is not that they’re associated with accurate information, but rather their ability to enable us to relive an experience.
This ability to target highly personal experiences in an inexpensive way could potentially be applied in a range of situations. Perhaps most obviously, mental health. While a deepfake avatar embodying a loved one won’t generally be able to truly capture the personality and intelligence of the actual person any time soon, that is arguably the wrong basis of comparison. Pets, for example, help us combat loneliness and studies show that they can extend our lives. Leaving the radio on makes the house feel occupied. Saturday Night Live’s Alexa for the elderly parody featured an “Uh huh” mode, which just listened to an elderly person tell stories and interjected with an occasional “Uh huh”. While done for comic purposes, it highlights that often what’s desired is the feeling of having your stories heard. Adding the dimensions of embodying similar capabilities in a deepfake of a loved one’s appearance and voice might just amplify these benefits.
People often talk about their need for closure for a traumatic event: they never got to say goodbye to a parent; they never confronted an abuser, or apologized to someone they wronged. Already today we will take symbolic actions for their therapeutic effect. We might write the letter we’d want to send, but know the intended recipient will never see it. We don’t write it for them. We write it for the feelings evoked by the act of writing it. Engaging with deepfake chatbots might help in the same way.
Beyond therapy, training might be enhanced. Consider training for military operations, where you’ll go into battle with fellow soldiers who you’ve grown very close to. Current simulations may recreate various aspects of a situation. But will you act the same way when you literally see your buddies in danger? In a business environment, can you simulate having a difficult conversation with the actual employee with whom such a conversation is called for?
Finally, of course, there’s entertainment. People may love the opportunity to, say, recast Game of Thrones with their family and friends, for example. The possibilities are endless.
Looking ahead, growing interest in the metaverse and the experiences we may have there may similarly challenge us to reexamine what counts as “real”.
So let’s not get paralyzed by our fear of fraud. I’m willing to wager a genuine Benjamin that in a few short years, the primary applications of this technology will be focused on the kind of honest experiences we can provide, not the deceptions we’ll fall prey to.
To learn more about Accenture’s work in this space, read our latest report, “Deepfake, real value: flipping the script on deepfake technologies.”
