Thanks to lockdown, most of us – me included – got a chance to catch up on a lot of TV shows and online content. One clip I came across on Youtube was of actor Bill Hader doing a spot-on impression of Al Pacino and Arnold Schwarzenegger on The Tonight Show with Conan O’Brien. That Hader is an actor par excellence is known. But there’s a part in the video where Hader’s face effortlessly morphs into that of Al Pacino’s just when Hader is about to start his impersonation of the veteran actor. For the less observant, the next 30 seconds could very well appear to be a young Al Pacino speaking. Welcome to the era of deepfakes.

Deepfake is a portmanteau of the words deep learning and fake. It basically is the latest spin on Photoshopping and video morphing, with the added complexity of voice-driven manipulation. Before the advent of AI, this sorcery was being executed and profiteered off by animation studios worldover. Now, with reams of content being generated online through video, voice and images and with the power of Artificial Intelligence, we’re in a time where it has become a bit more challenging to distinguish the real from the fake.

General Adversarial Networks (GANs) are commonly used to generate deepfakes. In 2014, Ian Goodfellow and researchers at the University of Montreal proposed a framework where two AI models – a generative one and a discriminatory one - are simultaneously trained, with the intention of enhancing the former model identifying errors possibly made by the latter model. GANs can be used for a range of applications, and the most common being for a deepfake – where faces of two people talking can be swapped. If the machine is provided enough data and can generate multiple cycles, it will eventually be able to generate pictures of nonexistent people. When it comes to voice modulation, companies like Adobe have developed a tool like Voco, which basically allows you to modify voice WAV forms using machine learning algorithms.

There are countless examples, similar to the Hader clip on Youtube. In 2018, a video of US President Donald Trump went viral where he’s advising the people of Belgium on climate change and in his characteristic style makes a claim about his bravado in withdrawing from the Paris accord. This video was released by Socialistische Partij Anders, a Belgian political outfit. Understandably, this caused outrage across the Internet against Trump. More recently, there was a video of Nancy Pelosi, speaker of the US House of Representatives appearing intoxicated, which was pulled down by YouTube.

Interestingly, a lot of videos of prominent politicians, not just something for jest like in the case of Hader. This brings us to an interesting juncture as far as democracy and free speech goes. Deepfakes are considered among the biggest threats to modern day democracy. There are tonnes of videos available online of politicians at rallies, or giving speeches earnestly, in a bid to reach out to their voters, but it is now very easy to manipulate their voices and content to derail their political ambitions or tarnish their image online.

There is an underlying risk of deepfakes taking over news as well. In 2018, China’s Xinhua introduced AI news anchors – developed by simulating voice, facial movements and gestures of news anchors. More recently, Indian startup Rephrase.ai have also released a similar software, available in more than 40 languages. Founder Ashray Malhotra in an interview stated that AI won’t take away news jobs just yet, but this may not be the case a few years from now. Not only do deepfakes resurrect the age-old concern of AI jeopardising jobs but also threatening the integrity of online content, and the millions of jobs it supports.

The imminent threat to news and democratic process by technology like deepfakes – the two bedrocks of our society – is a sign of troubling times ahead. We have already seen how the pornographic industry has been thriving thanks to deepfakes. Facebook, after the Pelosi episode, banned deepfakes in January 2020. Youtube and Twitter too have released policies aimed at countering misinformation, especially photos and videos. Researchers from University of California, Berkeley and USC have developed AI models to detect deepfakes. DARPA too released a similar AI model.

As this Barack Obama-led PSA, made by comedian Jordan Peele, says, “We're entering an era in which our enemies can make anyone say anything at any point in time."