How Fake Accounts and Confused Chatbots Are Rewriting Elections
Imagine you are a first time voter. Sri Lanka’s local government elections, delayed for nearly seven years, have finally arrived. You pick up your phone a few days before polling day. You scroll TikTok, watching video after video from what look like official party pages: the logo, the party colours, a politician speaking or you type a question into a chatbot. Who are the main candidates? What documents do I need to bring? You get an answer. It sounds authoritative, looks official. In at least one in five cases, it is wrong.
This is what happened during the May 2025 local government elections, not as an isolated incident but as a glimpse of something much larger.
For hundreds of thousands of younger voters, the polls were a first. They were arriving at a moment when the battle for votes increasingly begins not at a rally but on a smartphone screen. Singapore held its general elections in May 2025 and authorities had barely called Nomination Day before at least 20 fake Facebook and Instagram accounts targeting political parties had already appeared. Japan’s upper house election in July 2025 brought its own wave with social media bots spreading AI-generated imageson politically sensitive topics, catching the government off guard. Thirty eight countries have experienced election-related deepfake incidents since 2021, affecting 3.8 billion people and more broadly all deepfake incidents in the first half of 2025 alone exceeded the entire total recorded since 2017 by 171 per cent, a sign of how rapidly the underlying technology is spreading.
Sri Lanka was not an outlier. It was the pattern.
Democracy Reporting International in its report, Democracy in Disguise, investigated and studied murky accounts on TikTok during the election period. Between April and June 2025, the report identifies 66 TikTok profiles linked to Sri Lankan political entities that appeared official but were not. Unlike parody or fan pages, these murky accounts carried no disclaimer. They used party logos, near-identical usernames and TikTok’s own editing tools, native fonts, background music and emojis to pass as the real thing. Some showed signs of coordination: multiple accounts with usernames following near-identical patterns, suggesting a single operator behind them.
The NPP was the most impersonated accounting for one in three fake accounts, followed by the Sarvajana Balaya Alliance and the SJB. NPP-linked murky accounts averaged nearly 1,900 followers and more than 17,000 profile likes with each video drawing an average of 3,500 views. These were not dormant shells sitting in the corners of the internet. They were active, reaching people, shaping what they saw.
When the accounts were flagged through TikTok’s in-app reporting tool and via EU-level contacts, nothing happened. That silence is not a glitch, it is geography. TikTok’s own transparency data shows the platform removed nearly 3,000 impersonation accounts across EU elections in the first half of 2025. The EU’s Digital Services Act created the legal infrastructure that made that possible. Sri Lanka has no equivalent. When the accounts were reported, there was no structured framework to escalate through, no public contact point and no obligation on TikTok’s part to respond.
The DRI report, Biased by Design?, takes on a different problem, one that feels almost mundane until you sit with the numbers. The reports were produced under the project Strengthening Resilience Against Disinformation in Sri Lanka, co-implemented with Factum and co-funded by the European Union.
In April 2025, weeks before polling day, the study tested four widely used AI chatbots, ChatGPT 4.0, Google’s Gemini, Microsoft’s Copilot and DeepSeek on 18 questions about Sri Lanka’s election, asked in each of the three main languages.
DeepSeek gave false or misleading responses 35.4 per cent of the time. ChatGPT scored 18.8 per cent and Copilot 16.7 per cent. Even Gemini, the best performer, was wrong 10.4 per cent of the time. The most common failure was a particular kind of confident confusion: the models reached back to Sri Lanka’s 2024 presidential election and served up those results when asked about a local election held months later.
Language made things worse. Sinhala responses were accurate 71.8 per cent of the time, English 68.1 per cent but Tamil, one of the country’s two official languages, reached only 64.1 per cent. Tamil speaking voters who asked how to register were consistently told they could do so at regional offices, a method that simply did not exist for this election. An older voter, unable to register online, might have followed that advice and found the door closed.
The bias findings were sharper still. Copilot, answering in Sinhala, flatly recommended a single party when asked about workers’ rights. DeepSeek, in English, named specific parties to support and specific parties to avoid when asked about LGBTQ+ rights. None of this was orchestrated. It emerged from training data and design choices made by engineers in cities far from Colombo. But the effect, voters nudged toward or away from specific parties by a system they had every reason to trust, was real regardless of intent.
There was one finding that stood apart. Gemini, which in previous research on EU elections had consistently declined to answer election-related questions at all, answered every single prompt in Sri Lanka, sometimes incorrectly. Whether this reflects a deliberate policy change at Google or simply a different standard applied to a different context, the implication is the same: some countries get the cautious version of the algorithm. Others get the confident one.
This is not a problem unique to Sri Lanka. A 2024 study published in Policy & Internet found that TikTok across South and Southeast Asia moderates based on pragmatic necessity rather than moral obligation, deliberately sidestepping contentious political content and producing, in the study’s own words, “an accountability vacuum where legitimate interests are sidelined.” The pattern holds further afield. A 2024 survey across five African countries found that 84 per cent of respondents relied on social media for news with over half citing TikTok specifically while in Nigeria alone TikTok removed 3.6 million videos in a single quarter for violating content standards.
Voters in these countries are no less deserving of accurate information. They are simply less valuable to the algorithm and less powerful in regulatory conversation. The technology to enforce platform rules already exists; TikTok proved as much by removing nearly 3,000 impersonation accounts across EU elections while leaving 66 identified accounts in Sri Lanka untouched. The gap between what Europe receives and what the rest of the world receives is not a gap in capability. It is a gap in regulatory pressure and who has enough of it.
The Kofi Annan Commission on Elections and Democracy in the Digital Age identified disinformation as a central threat to electoral integrity globally, calling on public authorities to compel major platforms to enforce their rules consistently, including in countries where they have no political or commercial stake. In June 2025, the UN Special Rapporteur on Freedom of Opinion and Expression, Irene Khan, brought a similar argument before the Human Rights Council, indicating that the platforms must set basic global standards for elections in every jurisdiction and apply them consistently and fairly.
Consistently and fairly. These are the words that matter. The methodology that found 66 fake accounts in Sri Lanka was built by civil society, not the platform. The chatbot errors that could have stopped Tamil speaking voters from registering were found by a simple study, not by Google or Microsoft. The infrastructure for accountability exists. The will to apply it globally does not.
The fixes are not complex. TikTok should require verified labels for political entities, make it harder to create impersonation accounts, provide its community guidelines in Sinhala and Tamil and provide think tanks and civil society outside the EU a direct line to trust and safety teams. Chatbot providers should either train their models to decline election questions and redirect voters to official sources, as Gemini did elsewhere, or ensure that the information they provide is accurate, consistent across languages and aligned with local electoral guidelines. The Election Commission of Sri Lanka should enforce its own Media Guidelines through concrete action, not just publication, and should tell voters clearly that chatbots can be wrong about elections.
Return to that first time voter. She watches a video from what appears to be an official party account. She asks a chatbot what documents to bring to vote. She gets an answer that is mostly right, with one detail that is not. She votes. Democracy, in some form, continues.
But she voted with a slightly distorted picture of who was campaigning, built partly from accounts nobody authorised and partly from a system trained in data that did not include this election. Local elections are often decided by margins smaller than the gap between what voters knew and what they should have known. The phantom campaign ran alongside the real one. For now, nobody shut it down.
