For a glimpse of where artificial intelligence is heading in elections, look no further than India, the world's largest democracy, where voting begins on Friday.
An AI-generated version of Prime Minister Narendra Modi shared on WhatsApp shows the potential for hyper-personalized outreach in a country of around 1 billion voters. In the video (a democlip of unknown origin), Mr. Modi's avatar directly addresses a series of voters by name.
However, it's not perfect. Mr Modi appears to be wearing two different glasses and parts of the video are pixelated.
Additionally, Mr. Modi's party workers have been sending out videos on WhatsApp in which their AI avatars deliver personalized messages to specific voters about government benefits and urging them to vote.
These video messages are automatically generated in one of the dozens of Indian languages spoken by voters. The same goes for phone messages sent by AI-powered chatbots that use the voices of political leaders to address voters and solicit their support.
Such outreach requires only a fraction of the time and resources spent on traditional campaigning, and can be an essential tool in elections. But as technology moves into the political arena, there are few guardrails to prevent misuse.
Chatbots and personalized videos may seem more or less harmless. But experts worry that as technology advances and becomes more widespread, it will become increasingly difficult for voters to distinguish between real and synthetic messages.
“This year is going to be the Wild West and the unregulated AI space,” said Prateek Wagle, executive director of the Internet Freedom Foundation, a New Delhi-based digital rights group. He added that the technology is entering a media environment already contaminated with misinformation.
Around the world, elections have become a testing ground for the AI boom. This tool has been used to turn Argentine presidential candidates into Indiana Jones and Ghostbusters. During the New Hampshire primary, voters received robocall messages in a voice likely artificially created to resemble President Biden's voice, urging them not to vote.
In India, Modi's Bharatiya Janata Party (BJP) and the opposition Indian National Congress have accused each other of spreading election-related deepfake content online.
One of India's new frontier outposts is in the western desert state of Rajasthan. On the ground floor of a residential building in a dusty back alley, Divyendra Singh Jadun, a 31-year-old college dropout, runs an AI startup, Indian Deepfaker.
His team of nine people creates commercials using AI-generated avatars of Bollywood actors and actresses. But earlier this year, political parties and politicians began demanding that he do for them what he did for celebrities. According to Jadun, he has accepted 14 of the 200 requests.
Among those undergoing AI treatment is 33-year-old BJP member Shakti Singh Rathore. His job this election season is to tell as many people as possible about Mr. Modi's plans and policies. So he decided to create a replica of himself.
“AI is great and it's the way forward,” Rathore said, sitting in front of a video camera in his office at Indian Deepfakers, ready to go digital. “How can we reach so many beneficiaries of Mr. Modi’s programs in such a short period of time?”
As Rasool adjusted the saffron scarf emblazoned with his party's logo around his neck, Jadoon instructed him to “look into the camera and speak as if the person were sitting in front of you.” did.
Jadun got to work with about five minutes of material, including audio recordings and profile photos. He said he uses his AI system, which is open source, and builds his own code on top of it.
First, Mr. Rasool's face was isolated from each frame of the recording.
Data was then collected from facial features such as face size, lip size, and gaze.
Jadoon said the dataset was then fed into an AI model that learned to predict facial patterns.
“You have to keep running the program and tweaking the face until you get the best face possible,” he said.
A “cloning algorithm” also analyzed audio recordings and learned the rhythm and intonation of the voice. Jadun said adjustments often take him six to eight hours to perfect the face and synchronize the lips with the words. The rest is mostly automated.
In one demo, it took about 4 minutes to create about 20 personalized greeting videos.
Jadun said his team can produce up to 10,000 videos a day. For large jobs with deadlines, rent graphics processing equipment.
Generative AI can also remove language barriers, which is especially useful in countries with diverse languages. Rathore's avatar can be programmed to speak regional languages to reach remote parts of India.
In addition to texting video messages to voters, political parties are also using cloned voices to call people directly, all powered by chatbots such as ChatGPT.
Rasool said that in the past, when party representatives called voters, they would be hung up on. “But now, when a local leader mentions a constituent’s name, it immediately attracts attention.”
During the conversation, the chatbot asks about local government programs that provide free electricity or funding for startups. Jadun said the calls were recorded and transcribed for quality control and AI training.
Rasool said he has spent about $24,000 of his own money contacting about 1.2 million people through video messages and phone calls and obtaining information about those who did not respond.He called it an investment in his future with the Bharatiya Janata Party.
Personalized messages can be particularly powerful among Indians, said Nikhil Paffa, editor of Medianama, which covers digital media in India.
“India is a country where people love to take pictures with celebrity impersonators,” he said. “So if they get a call from, say, the Prime Minister and he talks to them as if he knows them, where they live, what problems they have, they actually get excited about it. I will.”
The Internet Freedom Foundation's Waghre doubts whether AI content will be persuasive enough to influence this year's elections. But the long-term effects could be problematic, he said. “Once people's information diet becomes normalized, what happens if a deceptive video exists in six months?” he said.
Mr. Modi himself has discussed adding disclaimers to AI-generated content to prevent people from being “misled”. Jadun and representatives from two other Indian AI startups have created what they call an “AI Coalition Manifesto” pledging to protect data privacy and maintain election integrity. For example, videos from Indian deepfakers are labeled as “AI-generated,” and the company's chatbot announces that the audio is AI-generated, Jadun said.
Narendra Singh Bhati, 28, a resort owner in Rajasthan, received an AI-generated phone call from Rathore this week. Batty said she was impressed by the personalization.
He said he did not realize the call was generated by AI, although it was clear in the script. Finally, “I also said goodbye to Mr. Rasool,'' Mr. Bhatti said.