Nick Bostrom, a Swedish philosopher at the University of Oxford, recently articulated a compelling idea around the existential risk of scientific discovery. This idea, called the Fragile World Hypothesis, relates the process of scientific discovery to reaching a hand into an urn and pulling out a ball at random. Most of the time, the ball we pull out will be a nice color, say blue, green, or pink. This is good because we like colors and this process makes life more colorful. However, somewhere within the urn, a black ball may exist.
A black ball is a discovery that will destroy our way of life. Unfortunately, we cannot see inside the urn and we cannot discern the color of a ball by the touch of our curious hands. We may be grabbing a new colorful ball, say the cure to malaria, or we may be grabbing a black ball and we will not know until we pull it out. Once it is out, we cannot put it back in. Thats it. The way of life we know, or life in itself, is over.
There are no implicit examples because we have not pulled out a black ball, but we have come close. For example, a brilliant German chemist named Fritz Haber dipped his hand into the urn and pulled out a very colorful ball. This ball was the Haber Process and it allowed humanity to convert atmospheric nitrogen into ammonia. Ammonia could then be used as synthetic fertilizer for crops and the resulting boom in food production was unprecedented. This was a very good thing, but there was a downside. Haber’s discovery unintentionally opened doors to the mass production of explosives, whose impact scarred Europe in the 20th century and the wounds have never healed. A few decades later in the Cold War, we were a few buttons away from destroying the entire planet.
This was very close to being a black ball. However, we skirted global destruction because of one factor, that making gigantic bombs is expensive. The process required, and still does, tons of resources and infrastructure. This is really only available to affluent countries that are typically governed with many roadblocks to prevent catastrophic action. What if, on the other hand, nature was constructed in a way where the chemical process of making catastrophic explosives was as easy as baking bread? It would have taken one resentful individual with no barrier to action to wipe out existence. Thank God, luck, or what/whomever you feel necessary to express your gratitude that this was not the case. That would have been a black ball.
I laid out the Fragile World Hypothesis because it is an interesting way to view the rapid advancements we are making everyday in AI. Using this lens, we can try to legitimately assess the existential risk of a new technology while it is in its infancy and if needed, build a robust management system. In a sense, this is akin to proactively monitoring moles on our skin and taking appropriate early action if they appear carcinogenic.
This brings me to deepfakes. I recently attended the Global Artificial Intelligence Conference in Seattle where I saw a presentation on deepfakes. Deepfakes are the result of a deep learning process (technical breakdown post to come) that combine images and audio to create realistic, synthetic video. This means that a video can be created of someone doing or saying anything that the creator wishes and in the most advanced cases, the viewer cannot discern its validity. I would like to clearly state that I do not think deepfakes will be an actual black ball. I think we are going to figure this out, but not without destabilizing our trust in information and kicking out the already crumbling legs supporting the idea of truth.
There are plenty of innocuous and hilarious examples of deepfakes on the internet. The Jennifer Lawrence speech where her face is swapped with Steve Buscemi’s is priceless. The countless fake videos of Nicholas Cage are wonderfully absurd. However, while these hysterical pieces of art have everyone laughing, deepfakes are currently causing horrific damage. In one of the first toxic applications of deepfakes, numerous actresses are having their faces realistically planted into pornographic videos. This could be career damaging and emotionally scarring and I couldn’t imagine the level of disgust one might feel knowing their face is pasted on the naked body of a stranger. As deepfakes get more realistic, and they will, imagine what it would feel like to try to explain the truth to their partner or spouse, their parents, their children, or their fans who cannot unsee what they have seen? Imagine if that was done to you? (with over 300 million photos uploaded to Facebook daily, there is likely plenty of material available for someone to do this with your face).
While this is one of the first ways that deepfakes are manifesting in our society, the potential for individual tarnishing is limited only by human creativity. Deepfakes can be made of you badmouthing your boss and getting you fired, or giving a speech at a terrorist rally and causing you to lose your community and legal standing. A really malevolent person could even make a video of a loved one denouncing you. You may unconditionally trust this loved one, but a millennia of evolution has built within us strong circuitry that associates what we see and hear with what is real. Even though we may ‘know’ they are not real, they will feel very, very real. With increasing education and awareness, I am confident we will find strategies to mitigate the interpersonal damage of deepfakes. However, disassociating what we see and hear with what is real is going to be immensely difficult and will not come without a loss of trust in all digital media.
This leads me to the larger societal implications of deepfakes. This is something we need to be ready for, maybe not for this year, potentially for 2020, and surely for 2024. Coming soon, anyone with with some computer literacy and access to a GPU, which can be cheaply rented on the cloud, can make a convincing deepfake of a world leader. The right will make videos of Bernie accepting donations from Halliburton. The left will make videos of Trump and Putin discussing insidious plans. Instead of pulling the fire alarm at school, some Dorito-munching kid can make a video of Kim Jong Un announcing a missile attack on Seoul just because they think it would be funny. With such a low barrier to entry, almost anyone can do this. At the very least, this will disrupt the democratic process, but at the worst, it could cost lives.
Video has long been the final barrier to the spread of misinformation. We have accepted that we cannot always trust information from text and photos. One can write slandering articles with deliberate misquotes and a skilled editor can maliciously photoshop someone. We have all been duped, but we are not evolutionarily hardwired to experience reality through text or static images. Divorcing truth from this kind of media is not a difficult task. Video will be different.
We are hardwired to assign objective truth to video unless explicitly told not to (as in cinema). If we have to divorce this media source from truth, we will not be able to trust any video anymore. The only way we will be able to gather trustworthy information is by exclusively watching multi-source verified live broadcasts or attending an event ourselves, both of which are rare at best. This is so critical because a properly functioning democracy relies on the dissemination of reliable information to the masses. This would severely reduce the amount of reliable information we can consume and thus impact our ability to make educated decisions.
Unfortunately, the fact that we are irrational beings means that we are not always going to skeptically view media. We are going to get tricked. As social media algorithms are designed in a way that fosters the spreading of outrageous media, this will likely increase the polarization that is already so damaging to our democracy. There is work being done in creating equally powerful algorithms to detect deepfakes, but these take time while an outrageous deepfake can go viral in seconds. Unless we see an increase in videos of bulldogs (with a Nicholas Cage face) riding skateboards to dilute our feeds, the outrage machine of politicized social media will become even more outrageous.
Hopefully I am wrong. Hopefully this will not turn out to be a big deal. Hopefully deepfakes do not destroy reputations, cause interpersonal damage, tamper with democracy, instigate geopolitical conflict, destabilize the last bit of trust we have in digital information, or cause our society to lose hope in the very idea of truth. Hopefully we can develop powerful detection algorithms or use cryptography and blockchain methods to effectively validate authenticity. Regardless, I would prefer to be ridiculed for having cried wolf in a wolf-free future than to risk being silent and not having these conversations now.
I always try to look for silver lining in emerging technologies. It would definitely be cool to have an Indiana Jones deepfake teaching history lessons. Deepfakes can enhance artistic potential in cinema and can be really funny when used by the right comedian. When they are respectfully consensual and are sufficiently labeled as fake, deepfakes can be great. However, in the grand scheme of things, are these positives worth the potential negatives? There are likely positive implications I am missing, but I remain unconvinced on how this technology could be a net positive for society.
We certainly cannot go back in time to forget about this technology. Banning it will not make it go away. Suppression will only force it into the shadowy underground. While I do not truly think deepfakes are a black ball, it is surely one of the darkest we have pulled out of the urn. As scientific discovery only occurs in one direction, we cannot put this ball back. It is up to you, and me, and all of us, to figure out what to do with it.