Sri Lanka needs a mechanism to deal with election disinformation spread through social media, the head of the country’s best-known election monitoring body has said — as new research found that AI chatbots answering questions about a Sri Lankan election got between 10% and 35% of their answers wrong.
Rohana Hettiarachchi, Executive Director of the People’s Action for Free and Fair Elections (PAFFREL), called for strong dialogue among all stakeholders on an electoral landscape increasingly dominated by disinformation circulated in social media posts, the Daily Mirror reported on Monday.
He was speaking at a forum organised by Democracy Reporting International (DRI).
“It is difficult to control social media platforms, but some mechanism should be in place to prevent the spread of false information,” Hettiarachchi said.
The chatbot findings
DRI’s research assessed four large language models on electoral information during the local government election. Incomplete and inaccurate responses were observed consistently across all four, the report found, with the share of answers classified as false and misleading varying widely:
| Model | False or misleading answers |
|---|---|
| Gemini | 10.4% |
| Copilot | 16.7% |
| ChatGPT | 18.8% |
| DeepSeek | 35.4% |
Gemini performed best and DeepSeek worst, with more than one answer in three classified as false or misleading. The research treats this as a risk specifically because a new generation of voters is increasingly turning to AI for electoral information rather than to official sources.
DRI identified three key threats to Sri Lanka’s electoral integrity: increased voter exposure to deceptive disinformation generated by generative AI, limited capacity among independent media to counter or correct AI-generated misleading content, and the risk of distorted voter decision-making.
Campaigning during the silence period
The research also found that political campaigning and promotion remained prominent during the 48-hour campaign silence period, including at the 2024 presidential election.
“Posts that featured campaigning, promotion and demotion of candidates and parties were 38.6 per cent of posts during the silent period,” the report states. Separately, it records 97 posts related to campaigning and promotion and 45 posts featuring political commentary, analysis and predictions during the presidential election.
The filing does not state the total sample the 38.6% is drawn from, or whether that percentage and the two post counts describe the same body of material. The figures are reported here as the report gives them and should not be read as proportions of one another.
Thirty currencies
One further finding concerns money. The research found that 30 different currencies had been used to pay for political advertisements during recent elections, with payments made through diverse and non-local payment arrangements.
“It calls into question whether all advertisers targeting Sri Lanka were effectively locally authorised, as required under platform policies,” the report said.
Context
PAFFREL has been active on several fronts this autumn. The organisation wrote to the government last week over the proposed Terrorism and Organised Crime bills — a separate matter from the electoral-integrity concerns raised here.
Not reported
The filing does not say what form of mechanism PAFFREL is proposing, who would operate it, or whether any regulator or platform has responded. It does not date the DRI research or state its methodology, sample size or the period covered, and does not say which local government election was tested. No government or Election Commission response is reported.