Generative artificial intelligence has turned the manufacture of convincing falsehoods into a low-cost, high-speed exercise. For election administrators across Southeast Asia, that shift matters less because of any single fake video than because of what it does to the information environment around a vote: it makes verification harder, spreads faster, and often arrives in languages and formats that existing moderation systems handle poorly.
What follows is a factual overview of the documented threat, the figures behind it, and the responses that are actually in place. It also separates the technical problem from the political noise, and flags where the evidence is thinner than the headlines suggest.
What turns a fake into a cybersecurity problem
A deepfake is audio, video, or imagery generated or manipulated to depict a real person saying or doing something they did not. The term covers sophisticated synthetic video, but also simpler edits such as face swaps, voice cloning, and spliced clips that misrepresent context.
Analysts usually distinguish deepfakes from “shallowfakes” or “cheapfakes” – misleading voice-overs, selective cuts, or doctored screenshots. The distinction matters for policy, because the simpler formats remain far more common than fully synthetic media, even as AI tools become cheaper.
Framing this as a cybersecurity issue, rather than only a media-literacy issue, reflects how closely the tools overlap with fraud. INTERPOL’s Project SynthWave study groups the criminal misuse of synthetic media into four categories: synthetic identities for digital fraud, non-consensual imagery, disinformation, and the mass creation of malware. The same generative models used to impersonate a candidate can be used to impersonate a customer, a bank, or an executive.
The study records that global deepfake prevalence rose by 245 per cent in 2024. In the Asia and South Pacific region, INTERPOL’s 2025/2026 cyber threat assessment found that discussions of deepfakes on criminal forums popular among regional actors increased by 600 per cent between February and June 2024, and that distributed denial-of-service attacks against government and financial targets surged by 92 per cent in 2024. Government websites were a leading target in the first half of that year – a period that overlapped with major elections.
Election-specific uses sit inside that broader picture. The recurring patterns that law enforcement agencies describe include impersonation of public figures in fake endorsements, fabricated audio attributed to candidates, spoofed news websites populated with machine-generated articles, and identity fraud tied to voter-facing services.
Why Southeast Asia is unusually exposed
Several structural features make the region a demanding environment for election integrity.
- Young, mobile-first populations with high social media engagement, which shortens the distance between a fabricated post and a large audience.
- Dozens of languages and local dialects, which slows fact-checking and content moderation that is often optimised for English.
- Uneven platform moderation, ranging from active labelling and fact-checking to minimal intervention.
- Rapid digitalisation of public services, which widens the attack surface for impersonation and fraud.
Indonesia’s 2024 election illustrates the scale of the problem. The communications ministry said it identified 3,235 election-related hoax posts between July 2023 and March 2024 and secured the removal of 1,923 of them, while major platforms removed millions more under their own moderation policies, according to reporting by BenarNews.
Exposure on that scale does not translate automatically into belief. A nationally representative panel survey of about 2,000 voters by the ISEAS–Yusof Ishak Institute found low reported exposure to well-known political deepfakes, and still lower belief that the events depicted were real. The researchers described “selective belief”: voters were more likely to accept content that aligned with their existing partisan views and to discount content that challenged them. That pattern is uncomfortable, because it suggests the harm from synthetic media may be less about changing minds than about reinforcing existing divisions.
The playbook: produce, distribute, target
Three things have to come together for a synthetic-media campaign to matter: the content has to be produced, it has to be distributed, and it has to be aimed at an audience that will act on it. AI has mostly collapsed the cost of the first.
Distribution remains the bottleneck and, for defenders, the main point of leverage. That is why the most consequential interventions tend to focus on how quickly false content is amplified, labelled, or removed, rather than on whether it can be generated at all.
A September 2026 threat-intelligence report from the AI developer Anthropic described a commercial influence operation that used a chatbot to build a network of roughly 1,000 social media accounts alongside a fabricated news outlet. The activity was aimed at voters in a Malaysian state election and drew on census and electoral data. The company said it disrupted the operation and shared intelligence with authorities and industry partners. The case is instructive less for its content than for its workflow: what once required teams of operators can now be assembled from a small number of prompts and publicly available data.
The region’s election calendar is crowded. Recent political developments, including state-level contests in Malaysia, illustrate how quickly political attention shifts, and each campaign creates a fresh market for accurate, verifiable information.
How governments and platforms are responding
Responses fall into three rough categories: law, institutions, and technology. A selection of documented measures and indicators is summarised below.
| Country or scope | Documented indicator or measure | Policy response |
|---|---|---|
| Region-wide | Deepfake discussion on criminal forums +600% (Feb–Jun 2024); regional DDoS attacks +92% (2024) INTERPOL, 2025/2026 |
ASEAN Cybersecurity Cooperation Strategy; INTERPOL regional operations |
| Indonesia | 3,235 election-related hoax posts flagged, 1,923 removed (Jul 2023–Mar 2024) Communications ministry, reported March 2024 |
Voluntary AI ethics circular; party self-regulation agreement during campaigning |
| Philippines | 200–300 deepfakes detected per day ahead of the May 2025 midterms Cybercrime Investigation and Coordinating Center, March 2025 |
COMELEC guidelines requiring disclosure of AI use; National Deepfake Task Force; detection tool deployed |
| Singapore | Law targeting manipulated online election advertising Elections (Integrity of Online Advertising) Act 2024 |
Corrective directions; fines up to S$1 million for non-compliant social media providers |
| Malaysia | Influence operation using a fabricated outlet and about 1,000 accounts reported in September 2026 Anthropic threat-intelligence report |
Regulator review of the report; calls from political parties for stronger laws on generative AI misuse |
Singapore’s Elections (Integrity of Online Advertising) Act 2024 is among the region’s more explicit instruments. It prohibits the publication of online election advertising that realistically depicts a candidate saying or doing something they did not, from the issue of the writ of election to the close of polling. The rules cover content produced with both AI and older editing techniques, and they apply to sharing and reposting as well as original publication.
The Philippines combined institutional and technical measures. The election commission issued guidelines requiring campaigns to disclose AI use in online material, while a national task force coordinated monitoring. The country’s cybercrime agency reported detecting 200 to 300 deepfakes a day in March 2025 and acquired a detection tool with a reported 30-second turnaround and 95 per cent accuracy, distributing licences to universities and election watchdogs so that verification did not depend on a single government office.
Indonesia took a lighter-touch route during its 2024 cycle, relying on a circular on AI ethics and a voluntary social media framework agreed among parties and candidates. That approach preserved space for legitimate political speech, but it also left much of the enforcement to platforms and to voters themselves.
Where the defenses still fall short
Detection is a moving target. Every published improvement in identification is followed by new generation techniques designed to defeat it, and detectors tend to perform worse on low-resource languages and on compressed, re-shared clips – exactly the formats that circulate most widely.
The more important gap may be conceptual. During the Philippines’ 2025 midterm campaign, a fact-checking coalition analysed 35 unique altered claims and found that 11 likely involved deepfake techniques, while 24 relied on simpler manipulation such as doctored images, selective cuts, and fabricated posts. Investment focused only on AI detection can therefore miss the majority of what actually spreads.
Reporting also lags the perceived threat. INTERPOL’s Project SynthWave study notes that recorded cases in several Southeast Asian countries remain low relative to the scale that law enforcement describes – a gap the study attributes partly to underreporting, limited forensic capacity, and victims’ reluctance to come forward.
There is also a legitimate debate about how far takedown powers should reach. Digital- and human-rights organisations have cautioned that election-time content rules can be drafted too broadly and, in some jurisdictions, used to restrict criticism. Their recommendations typically include narrow definitions, published criteria, clear appeal routes, and explicit protections for journalism and satire. Reporting by The Guardian in October 2026 captured the same tension: experts argue AI lowers the cost of disinformation, while noting that distribution platforms, not generation tools, remain the decisive factor.
What practical mitigation looks like
No single control solves this. The measures that appear most durable are layered and combine technical, institutional, and educational elements.
- Prebunking and media literacy. Teaching voters how synthetic media is made, before they encounter it, tends to be more effective than correcting falsehoods afterward.
- Rapid, transparent correction. Multi-stakeholder task forces that include election bodies, platforms, fact-checkers, and civil society can compress the time between detection and response.
- Provenance and labelling. Content credentials and platform labelling help audiences judge where material came from, though adoption remains incomplete.
- Local-language fact-checking. Verification capacity in regional languages is a persistent gap and a high-value investment.
- Institutional discipline. Media organisations that verify before amplifying can avoid becoming an unwitting distribution channel.
What to watch next
The region’s next few election cycles will test whether current measures scale. Malaysia could hold a general election by the late 2020s, while the Philippines and Cambodia have scheduled polls in 2028 and Indonesia the following year, according to reporting by The Guardian. Each will be the first such contest to be run with widely available generative AI tools at the current level of capability.
Two emerging issues deserve attention. The first is the prospect of campaigns aimed not at voters but at the large language models that many people now query for information, in an effort to shape the answers those systems give. The second is regional coordination: ASEAN’s cybersecurity strategy for 2026–2030 and plans for a regional computer emergency response team could, if implemented, improve threat-intelligence sharing between member states with very different levels of capacity.
Frequently asked questions
What is a deepfake, exactly?
A deepfake is audio, video, or imagery generated or manipulated so that it depicts a real person saying or doing something they did not. It can be created with AI tools or with older editing techniques such as splicing and dubbing.
Are deepfakes actually changing how people vote in Southeast Asia?
The evidence is mixed. Surveys suggest exposure to specific political deepfakes is often limited, and belief depends heavily on a voter’s existing partisan views. The documented risk is less about wholesale mind-changing and more about reinforcing divisions, undermining trust, and creating confusion that can be exploited around polling day.
Why is Southeast Asia more vulnerable than some other regions?
High social media use, young and mobile-first populations, many local languages, and uneven platform moderation all play a role. Rapid digitalisation of public services adds further opportunities for impersonation and fraud.
Are deepfake detection tools reliable?
They are improving but not definitive. Accuracy varies with language, compression, and how many times a clip has been re-shared, and detection alone cannot address simpler forms of manipulation, which remain more common.
What can an ordinary voter do?
Check the original source before sharing, look for provenance labels, and treat emotionally charged audio or video of public figures with extra caution. Verification habits matter most when content is designed to provoke a fast reaction.
The test ahead
The lesson from the past few election cycles is that the threat is distributed, not concentrated. No single platform, statute, or detector carries the whole burden. What appears to work is a division of labour: regulators set clear, narrow rules; platforms enforce them at speed; election bodies coordinate and monitor; fact-checkers work in local languages; and voters develop the reflexes to pause before they share.
That is a less dramatic answer than the technology itself invites, but it is also the one the documented evidence supports. The countries that treat synthetic media as an ongoing operational problem – rather than a one-off crisis to be legislated away – are likely to be the ones whose elections remain legible to the people participating in them.
