Last Saturday, on the eve of the Bangladesh national elections, two separate videos emerged on Facebook showing independent candidates Abdullah Nahid Nigar and Beauty Begum announcing their withdrawal from the polls. Both the videos were, however, deepfakes - digitally manipulated using deep generative methods.
While Nigar won in her constituency, Begum lost with a margin of less than 3,000 votes. While it maybe impossible to ascertain the affects of the deepfake video of her withdrawal on her final vote count, it sets a worrying precedent for all future elections.
The year 2023 was highly prolific for the machine learning industry, with phrases like "ChatGPT", "Bard" and "Midjourney" becoming major buzzwords on the internet. AI's share of the investment pie among start-ups in the United States was found to be doubled last year, compared to 2022.
A technology once relegated to science fiction films and dystopian novels, is now readily available in the form of a diverse range of affordable tools. And academics, policymakers and human rights groups are worried that dystopian problems will soon follow.
Year Of Elections
The year 2024 will see over 50 countries from around the world head for elections, with The Economist calling it "the biggest election year in history".
This includes India, along with Indonesia, United States, Mexico, and the United Kingdom among others. And the emergence of deepfakes in political disinformation in Bangladesh, although sparse, is a warning for all.
BOOM's Bangladesh team had fact-checked several such videos last year, that turned out to be digitally altered using deep generative methods.
Along with the two fabricated videos of independent candidates purportedly withdrawing from the elections, we also fact-checked a deepfake video of Dhaka Chief of Detective Bureau Harun-or-Rashid purportedly asking people to boycott the elections, while criticising Prime Minister Sheikh Hasina.
A video of Bangladeshi opposition leader Tarique Rahman, who also serves as the chairman of Bangladesh Nationalist Party, went viral early December, where he is seen urging people to not criticise Israel's bombardment of Gaza. This too turned out to be a deepfake, according to fact-checkers at Fact Watch..
Another deepfake clip of acting Chhatra Dal President Rashed Iqbal Khan went viral, where he is seen claiming to be 47 years old.
The appearance of political deepfakes ahead of Bangladesh elections was not as prominent. According to Sumon Rahman, fact-checker and media researcher at the University of Liberal Arts Bangladesh, this is largely due to the boycott of the elections by the opposition.
"Fortunately, the investment in deepfake during Bangladesh election was not big enough, as the competition lost its momentum after the opposition's boycott," he told Decode. However, he highlighted the strong potential for abuse arising from the novel nature of such technology.
"People are more likely to be influenced by deepfakes, as this is a new technology for most of them. The untraceable nature of deepfakes makes it very difficult to debunk even by the fact-checkers," he added.
Seeing is Believing... NOT!
For India, use of deepfakes in elections is already a tried-and-tested matter.
On February 7,2020, a day before Delhi headed for legislative polls, several videos appeared showing Bharatiya Janata Party leader Manoj Tiwari criticising the Aam Aadmi Party government and its leader, Arvind Kejriwal, and urging people to vote for the BJP.
These videos showed Tiwari speak in English, Hindi and a Hindi dialect from Haryana. However, Vice found that only the Hindi version was originally shot by Tiwari - the English and Hariyanvi versions were actually fabricated using a 'lip-sync' deepfake algorithm trained on videos of Tiwari's speeches.
More recently, AI generated images were found being shared unabated on social media platforms, ahead of the Telangana legislative elections.
Decode reported on how online creators were using innovative techniques to bypass the filters on social media platforms that restricted the use of generative AI for political posts. The Election Commission is yet to publicly address the legality of such AI-generated images being used ahead of elections.
Political parties are already using AI voice cloning apps for better reach of political leaders, to make them speak in languages they may not be fluent in. Furthermore, last year Decode reported extensively on the rise of deepfakes using such AI cloning software for the purpose of scamming people, along with the growing prevalence of deepfake pornography on platforms like X (formerly Twitter).
While disinformation has been a consistent issue around elections, deepfakes are expected to worsen the problem by raising the level of deception. The old adage of "Seeing is believing" is no more true, when electorates are facing fabricated and synthetic visuals that cannot be distinguished from the real ones.
Rahman worries this could pose a big challenge "as people during elections are heavily driven by confirmation biases".
"Deepfake can certainly reinforce their false beliefs", he notes. "The more acute problem is that people during election do not often nurture their critical selves, and debunking a deepfake requires little bit more time and effort."
India, A Fertile Ground For Deepfakes
After Financial Times reported on the use of deepfakes to spread political disinformation ahead of the Bangladesh polls, Meta removed the videos in question.
However, as the article pointed out, the problem worsens due to the lack of regulations to hold the bad actors behind such disinformation accountable. Instead, the laws are selectively enforced to crack down on opponents and dissidents. This runs parallel to the failing health of Bangladeshi democracy.
Immediately after the polls, the US Department of State published a press release voicing concerns over "the arrests of thousands of political opposition members and by reports of irregularities on elections day".
While the lack of opposition took away incentive to invest in deepfakes in Bangladesh, Rahman believes that "India has a more grim prospect with such technology" in the upcoming election.
"The nature of political participation, financial investment, communal tensions and caste politics etc. all together will make India a fertile ground for deepfake production," he warns. "There is a strong possibility that its production will be outsourced, alongside home-made ones. Our fact-checking usually takes a bit of time, but deepfakes might not give such time before causing violence."
"Not Well Equipped"
While India is far from being a quasi-single-party-state like its eastern neighbour, selective enforcement of regulations, particularly the Information Technology Amendment Rules, 2023, has been a cause for concern. In such a scenario, bad-actors will find plenty of opportunities to deploy advanced, AI-enabled disinformation tactics.
"The unfortunate truth is that we are not well equipped," said Prateek Waghre, Executive Director at Internet Freedom Foundation.
Speaking to Decode, Waghre expressed that neither the Election Commission, nor the tech platforms have the capacity to handle the surge in deepfakes and AI-generated content, and stressed on the "need for political actors to be transparent on their use of generative AI".
Bringing up the deepfake videos of two independent candidates in Bangladesh withdrawing on the eve of elections, he adds, "It's unlikely that the platforms are going to be able to respond to such videos in time. Because even detection is far from perfect right now."
By the time such videos are taken down, the damage is done. But what about regulations?
"If you look at the history of IT Rules, and how they were amended - it was a very problematic episode. You’re not going to achieve much by blocking YouTube, Facebook or WhatsApp. Because you are going to push a lot more people - beyond those who have committed any error or any crime," he added.
Given the case of Bangladesh, sweeping powers can often result on non-transparent and selective enforcement of regulations. And even if the there was a clear intent to address the problem, Waghre believes it would still be a challenge for the authorities to address such advanced form of disinformation at a national scale.
Given the complex and novel nature of the problem, the upcoming elections might be ground zero for fact-checkers, researchers and policy makers to figure out how to check further abuse of such technology to deceive the electorate at a mass scale.