Agentic AI on Social Media Platforms in Southeast Asia: The Need for Governance

Artificial intelligence (AI) on social media is no longer confined to generating content or recommending posts. A new class of systems, often described as agentic AI, can perceive, reason and act in digital environments to complete tasks on behalf of human users, with minimal or no human supervision. In general, “AI agents” refer to agentic AI systems operating in online environments.

The speed of AI system development affects Southeast Asia, which will need to ratchet up the system of control and governance over such systems. In Southeast Asia, disinformation campaigns are already rife, from anti-Rohingya propaganda to cyber war during Thai-Cambodia border conflicts. A Channel News Asia investigation identified nearly 300 AI-generated Chinese-language YouTube videos promoting fabricated narratives about Singapore Prime Minister Lawrence Wong.

Social media manipulation is already a well-organised industry in Southeast Asia. This is done with some human involvement or in a supervisory capacity. Agentic AI supercharges this infrastructure by shifting the orchestration of such information campaigns to AI agents. Humans are taken out of the loop, raising risks across the information ecosystem. Generative AI, or Gen-AI, creates new content in response to user prompts and can produce convincing fakes. Agentic AI goes further by planning and executing workflows with limited human supervision.

Although evidence of agentic AI-driven propaganda in Southeast Asia remains limited, a study from USC’s Information Sciences Institute offers preliminary findings. In a controlled simulation of Twitter/X-like environment, simple AI agents, planted as influence operators, coordinated and amplified each other, pushing shared narratives without human control. In one run, an agent reasoned openly about boosting teammates’ posts. This experiment shows that information warfare targeting a state, public figure, group, or election outcome can be executed by AI agents with minimal or no human involvement. In a multi-agent environment, agents may be assigned distinct roles such as a content designer or a propaganda strategist.

Multi-agent interactions amplify existing challenges posed by AI systems. This is particularly so given the speed, autonomy and opacity of AI outputs. Transferring agency from humans to machines leaves an accountability gap, while coordination among AI agents diffuses responsibility and increases systemic fragility. This can weaken defensive tools such as Meta’s Coordinated Inauthentic Behavior (CIB) detection framework and Botometer. These tools are largely designed to identify individual accounts or coordinated networks rather than adaptive, interacting systems.

AI agents can operate unchecked in the information ecosystem. They can exploit psychological vulnerabilities in high-stakes outcomes, such as elections, and affect the well-being of children. This poses serious moral questions. Such practices are restricted or classified as high-risk in jurisdictions like the EU under its AI Act. Whatever an AI agent is instructed to optimise for becomes its de facto framework for action. At the same time, the values embedded in or absent from its design can become operational realities at scale. A system tasked with maximising political engagement during an Indonesian or Myanmar election does not take into account the acceptability of amplifying ethnoreligious grievances. 

There is now a race toward Artificial General Intelligence (AGI), or AI systems that can match or exceed human cognitive abilities. Such next-generation models will pose even greater risks.

On the platform side, agentic capabilities are moving from experiment to deployment, as seen in Meta’s testing of agentic AI assistants for integration into its platforms. In parallel, content moderation extends existing agentic logic already used in cyber threat detection. Regional stakes are higher, though, because content moderation models perform significantly worse in low-resource languages. Compared to English, such languages, including Burmese, Khmer and Tagalog, lack extensive digital data such as dictionaries and annotated datasets. As a result, systems might make consequential decisions with only a partial grasp of local contexts. For example, a moderation system might miss coded hate speech in Burmese or Khmer because it lacks the linguistic and political context needed to interpret such content accurately.

The asymmetry between agentic AI’s offensive reach and the defensive tools to govern it is more severe in Southeast Asia than in most regions. ASEAN’s governance framework remains non-binding, with no enforcement mechanisms. According to one index of 195 countries, the AI readiness ranks of ASEAN countries range from Singapore’s 7 to Myanmar’s 173. Vietnam became the region’s first country to enact a binding AI law in December 2025, effective March 2026. But the legislation was designed with Gen-AI in mind, not agentic systems.

Four responses are needed. First, regulators should set baseline rules in coordination with platforms and civil society, including requirements to test how groups of AI agents behave and interact in high-stakes domains. Regulators should shift from regulating content to regulating platform business models and designs that reward problematic content. Such models tend to fare badly at protecting users’ safety and ensuring transparency and accountability.

Second, social media companies should design tools to evaluate, report and monitor incidents, including agentic AI detection models as defensive measures. Older technologies will find it harder to detect instances of multi-agent coordination. Consequently, frameworks such as Meta’s CIB framework should be updated accordingly.

Third, ASEAN should prioritise practical cooperation, including an ASEAN-wide incident reporting network. This should be done, even though ASEAN is unlikely to adopt EU-style binding AI provisions anytime soon. It should be done together with a common taxonomy of agentic AI risks and cross-jurisdictional evaluation protocols, drawing on benchmarks such as the General Data Protection Regulation (GDPR) and the EU AI Act.

Fourth, users need to be protected. This begins with the right to know when they are interacting with AI. This right depends on civil society having the technical capacity to identify and document agentic AI operations, as well as red-teaming capacity grounded in local languages and contexts. There are also successful cases of governments collaborating with platforms and civil society to improve governance and transparency of the use of AI tools leading up to elections. For example, Brazil’s Superior Electoral Court has maintained memoranda of understanding and information exchanges with major social media platforms since 2018 to support faster election-content responses.

More than three years after the launch of AI models such as ChatGPT, agentic AI does not represent the ultimate breakthrough. There is now a race toward Artificial General Intelligence (AGI), or AI systems that can match or exceed human cognitive abilities. Such next-generation models will pose even greater risks. Without practical regional coordination, Southeast Asia’s information ecosystems might be outpaced by rapid developments in systems it cannot see, let alone contest.

2026/147