LLMs can help reduce polarisation in elections: report

Current large language models (LLMs) have the capability to accurately represent their voter samples and potentially reduce polarisation, new-age think tank Policy 4.0 said in a report.

The report titled ‘An AI Experiment for Augmented Democracy in India’ shows the positive aspects of using artificial intelligence (AI) in elections as these tools could potentially help overcome traditional obstacles to democratic participation in the future, making them essential in modern political contexts.

Augmented Democracy is the idea of using digital twins to expand people's ability to directly participate in a large volume of democratic decisions.

“Generally, the attitude towards AI in elections is very negative. Automatically, people think that if AI is saying it, it is trying to mislead us,” Policy 4.0 chief executive Tanvi Ratna told ET.

“In the 2024 general elections, tech-savvy youth are increasingly participating in voting decisions making the exploration of AI in elections crucial for India.”

However, cybersecurity experts have voiced concerns about the potential misuse of AI and deepfake technology, among other methods, to sway voters.

The study involving 50 Indian voters aged 18-29 from urban areas was conducted to evaluate the capability of existing LLMs in making complex and customised voting decisions concerning Indian politics. Using the OpenAI API playground,the GPT-3.5-turbo version was used to model AI variants for each voter.

In the study, voter preferences were collected for India's two major national party alliances—the center-right National Democratic Alliance (NDA) and the center-left United Progressive Alliance (UPA).

The report demonstrated that the model's prediction of vote choice had a 90% accuracy rate. The rationale of the voter and the AI twin matched while deciding on the political party. According to Ratna, the experiment helped voters to come to a conclusion based on their preferences.

“AI model exhibited the best mimicking performance for low-confidence fringe voters, decreasing effectiveness for moderate-confidence voters, and was least effective at mimicking high-confidence voters,” the report said.

"As the first experiment involving live Indian voters, this study significantly advances our understanding of how AI could enhance democratic processes in India. While a short study, it opens the door to a wide spectrum of research in the field,” said Prabhu Pradhan, AI scholar at Policy 4.0. and lead researcher in the study.

This study marks the first experiment for augmented democracy in India, although similar experiments have been carried out in other countries.

In countries like Germany and the Netherlands, voter advice tools such as Wahl-O-Mat and Stemwijzer have been introduced, aiding millions of voters in making informed choices. These tools have been offering impartial recommendations. AI is also being used to monitor and analyse election data, aiming to prevent fraud and enhance transparency.