Political campaigns embrace AI to reach voters across language barriers

In December 2023, Indian Prime Minister Narendra Modi addressed attendees of the Kashi Tamil Sangamam in Varanasi, Uttar Pradesh, in Hindi. But the audience heard the speech in Tamil in real time. That was made possible by an AI translation tool, Bhashini , that Modi himself launched in July 2022. Just before this year’s elections, he once again used Bhashini to translate his Hindi speeches into citizens’ native languages during rallies in the southern part of the country.

It was a savvy move for a politician hoping to reach voters in a country of 1.3 billion — particularly a country that speaks over a hundred languages and countless more dialects. AI tools are already widely used in translation and marketing services, but political campaigns around the world have been slow to adopt the technology, in part out of concerns around disinformation. But as the technology proves itself more, campaigns are starting to explore large language models (LLMs) as a new way to do voter outreach — particularly when there’s a language barrier involved.

The most successful approaches combine translation with message testing, using an LLM to generate a variety of messages and employing conventional focus groups to test the most effective ones. In Hungary , an advocacy group called noÁr used the system to develop the most effective calls for action on social media — although results of the test were mixed. In San Francisco, civic groups are experimenting with the use of AI for real-time translation of city council meetings , showing how useful technology can be for outreach to minority language groups.

For campaigns, the benefits are obvious. Nearly 66 million Americans speak a language other than English at home — predominantly Spanish. But among those, 10 million speak Asian American and Pacific Islander (AAPI) languages, and  translation skills for those are much harder to come by than for Spanish. A July 2024 survey revealed that nine in 10 AAPI voters plan to cast a ballot in the U.S. election this year.

Already, researchers in the U.S. are encouraging campaigns to experiment with the new tools. A joint University of Chicago and Stanford white paper , summarizing political science work on LLMs, emphasized that they were particularly powerful for tailoring messages to particular interest groups. “Under-resourced campaigns may face challenges in creating content that appeals to voters. Generative image and text tools can enable these campaigns to draft more compelling speeches, press releases, social media posts, and other materials,” the paper’s authors wrote. “These methods can also be used to create tailored materials for different audiences.”

One group that does voter outreach for U.S. Democratic and progressive policies has tested out an AI model built to do just that. Last December, the political advocacy group AAPI Victory Alliance received a $20,000 grant to develop an AI chatbot that would translate and refine campaign messages into Hindi, Tagalog, Chinese, Hmong, Korean, and Vietnamese. Awarded by the venture fund Higher Ground Labs, the money was meant to fund a three-week test run to determine how useful LLMs could be for campaigns.

The group’s executive director, Varun Nikore, described the system to Rest of World as a kind of multilingual A/B test. Nikore’s team used OpenAI’s ChatGPT and Anthropic’s Claude to translate the highest-performing messages from Democratic campaigns on key election topic areas into the six languages. The team then leveraged AI to reformulate the message “in such a way where … the AI chatbot believes that it would resonate better [with a specific community],” Nikore said.

The system helped save time and, therefore, costs. Where humans took 45 minutes to translate and refine the message, AI tools were able to do it in 15 minutes, Nikore said. He believes that, with refinement, translation could come down to as little as five minutes.

In the final tests, the AI-translated messages performed roughly on par with human translations, winning 52% of the time . Restricted to Hindi translations, the advantage improved to 58%.

Experts say Hindi’s good performance is likely a result of the same data issues that make AAPI languages difficult in the first place. LLMs are more effective in Hindi because “so much work is being done on AI in India and on the Hindi translation,” Deepak Puri, a Silicon Valley veteran who runs nonprofit Democracy Labs, which tests election tech , told Rest of World . With more Hindi-language text to train on, LLMs have more material to work from. “It’s just a question of where the algorithms have had a chance to learn where the mistakes are.”

Puri worked with AAPI Victory Alliance on feeding the AI translation tools prompts. However, tens of thousands of Indian Americans speak other, less-popular Indian languages, such as Gujarati, Urdu, Bengali, and Tamil. Sikh voters, most of whom speak Punjabi, are upping their civic engagement . Telugu, which is among the fastest-growing foreign languages in the U.S., boasts more than 419,000 speakers.

Besides Hindi, most Indian language models are “low-resource,” Sundeep Narwani, co-founder of Narrative Research Lab, told Rest of World . They have “relatively less data available for training conversational AI systems,” he said. “To bridge the gap, we need more data annotation and collection .”

Even as models improve, humans will have an important role to play in crafting the messages sent to voters. “You need to relate with the technology, in a back-and-forth exchange, to guide its outputs,” communications consultant and political advisor Doug Hattaway told Rest of World . “It can produce useful first drafts, but ultimately the user is responsible for the final product.” For AAPI campaign groups, that means hiring native speakers to oversee AI translations, something many Democratic-affiliated organizations are already doing.

Still, many people fluent in low-resource languages have already lowered their expectations for how well an LLM tool can speak their language. Urvashi Aneja, founding director of Digital Futures Lab, told Rest of World that she sees lots of AI skepticism in her work doing regional language outreach in India. While building a contextualized data set for AI tools will help, “it requires resources, and it requires funding,” she said. “I think it also creates a new entry point for [a] kind of bias … What are the information points that you’re putting in? What are the kind of prompts that you’re putting in?”

So far, most locals are unimpressed with the latest generation of tools. “Most of them are saying they are pretty happy with Google Translate,” Aneja said. “They don’t see a huge difference with ChatGPT.” ▰

Ananya Bhattacharya is a reporter for Rest of World covering South Asia's tech scene. She is based in Mumbai, India.