Every study we have found about AI in the region, summarized in plain language. Pick a topic or a country to narrow it down.
175 studies, newest first
ThailandHealth
Shows that having an AI rank how clinically important each fact in a Thai doctor-patient conversation is before writing the visit note makes the notes more complete, and lets a small model trained on synthetic data match a leading commercial one.
Preprint · October 1, 2026 · arXiv 2610.01234
VietnamLanguage and models
Proposes a model that blends image and text features at several levels of detail to answer questions about pictures in Vietnamese, and reports that it beats earlier approaches on a standard Vietnamese test set.
Preprint · October 1, 2026 · arXiv 2610.01637
VietnamLanguage and models
Builds a large test set from real Vietnamese legal consultations and finds that AI systems answer fluently but still struggle to find the specific legal provisions that back up an answer.
Preprint · September 30, 2026 · arXiv 2609.39189
IndonesiaSociety and elections
Finds that an AI and virtual-reality batik experience won users over through immersion and a sense of authenticity, and that checks on faithful cultural symbols mattered more than image quality alone.
Preprint · September 29, 2026 · Research Square
Southeast AsiaGovernment and rules
Argues that the real test of AI rules in the region is whether any body can actually halt a deployment, and finds the answer differs widely: Singapore relies on sector regulators, Vietnam's new powers are untested, and Malaysia and Indonesia are still building theirs.
Report · September 29, 2026 · ISEAS Perspective 2026/70
ThailandLanguage and models
Releases small open Thai-English translation models trained on carefully filtered examples that follow instructions on terminology and formatting without losing accuracy, matching or beating same-size rivals.
Preprint · September 28, 2026 · arXiv 2609.34770
ThailandSchools and universities
Finds that Thai learners of English stayed willing to speak with ChatGPT because it felt socially present yet clearly not human, understood them and never judged their mistakes.
Journal article · September 28, 2026 · Technology in Language Teaching & Learning
MyanmarSchools and universities
Finds that Myanmar students misuse generative AI mainly when peers treat it as acceptable and they can excuse it to themselves, and that visible enforcement by universities does most to curb it.
Journal article · September 27, 2026 · Journal of ICT and Education
SingaporeSociety and elections
Finds that a fact-checking chatbot helped people in Singapore, India and the US spot fake image claims on the spot, but the benefit had vanished a week later, falling most in Singapore and India.
Preprint · September 26, 2026 · arXiv 2609.32739
Southeast AsiaLanguage and models
Shows a small model trained on regional data matches images to text in seven Southeast Asian languages better than a larger general-purpose model, while running faster on ordinary computers.
Preprint · September 25, 2026 · arXiv 2609.30739
VietnamSchools and universities
Describes an AI tutor for Vietnamese schools that runs on local hardware to keep student data in the country, roughly halving response time and improving accuracy on complex tasks compared with standard setups.
Preprint · September 25, 2026 · arXiv 2609.31568
VietnamSchools and universities
Finds that Vietnamese English teachers rate ChatGPT's comments on essay content above its grammar corrections, and that their willingness to adopt it hinges on usefulness and on backing from their institution.
Journal article · September 25, 2026 · Electronic Journal of Foreign Language Teaching
Southeast AsiaSchools and universities
Finds that Southeast Asian English teachers mostly use AI to produce model texts, help students brainstorm and proofread, and that these narrow uses reflect gaps in training rather than the tools' limits.
Journal article · September 25, 2026 · Electronic Journal of Foreign Language Teaching
SingaporeSchools and universities
Finds that an automated marker for Singapore radiology trainees agreed with human graders on 97.5 percent of short answers, though disagreement grew as answers got longer.
Journal article · September 25, 2026 · Singapore Medical Journal
CambodiaLanguage and models
Releases the first maintained open-source tool for turning written Khmer numbers and symbols into spoken words and back, which got almost all test cases right despite Khmer having no spaces between words.
Preprint · September 25, 2026 · arXiv 2609.30984
VietnamLanguage and models
Releases nearly a thousand hours of Vietnamese speech across five dialect groups with matched fake voices, and shows existing deepfake detectors fail badly when the fake closely imitates a real speaker.
Preprint · September 24, 2026 · arXiv 2609.30005
ThailandSchools and universities
Finds that Thai lecturers plan to use AI when they expect it to help, enjoy it, see colleagues using it and have made it a habit, while ease of use and cost make little difference.
Journal article · September 24, 2026 · Human Behavior, Development and Society
IndonesiaSchools and universities
Finds that unedited AI-made vocabulary exercises for Indonesian pupils were vaguer and less culturally local, while materials teachers guided with local prompts fit the lesson and context far better.
Journal article · September 24, 2026 · Technology in Language Teaching & Learning
MalaysiaSchools and universities
Finds Malaysian lower-secondary students fairly ready to use generative AI overall but least sure how to apply it to learning Malay, pointing to a need for subject-specific guidance.
Journal article · September 24, 2026 · International Journal of Research and Innovation in Social Science
Timor-LesteFarms, climate, and disasters
Shows that a model first trained on larger foreign water data can classify groundwater quality in Dili accurately from a small local dataset, sharply cutting missed cases of unsafe water.
Journal article · September 24, 2026 · Automation (MDPI)
IndonesiaFarms, climate, and disasters
Finds that machine learning appears to fix satellite solar estimates in Central Java under usual testing, but every model does worse than the raw satellite data when tried on a station it never saw.
Preprint · September 24, 2026 · Research Square
VietnamGovernment and rules
Builds a tool that drafts the data-processing records Vietnam's new personal data law requires, and shows a language model run on an organisation's own hardware does about as well as a cloud model, keeping data in the country.
Preprint · September 23, 2026 · arXiv 2609.27359
VietnamLanguage and models
Finds that AI systems answering English questions about Vietnamese labour law struggle to find the right statutes, and that an added fact-checking step mostly just keeps citations intact rather than making answers better.
Preprint · September 23, 2026 · arXiv 2609.27376
ThailandLanguage and models
Releases a 342-million-word Thai corpus from news, Wikipedia, speech transcripts and social media, machine-parsed for grammar so researchers and language tools can study how Thai sentences are built.
Preprint · September 23, 2026 · arXiv 2609.27558
IndonesiaJobs and the economy
Finds that Indonesian banks that adopted AI earned slightly higher returns and had fewer bad loans, with chatbots giving the biggest lift, though training staff on AI dented profits in the short run.
Journal article · September 23, 2026 · Journal of Central Banking Law and Institutions
IndonesiaJobs and the economy
Finds that AI brought six large Indonesian banks only small changes in headline financial ratios, with the real gains coming in automation, fraud detection and day-to-day risk handling.
Journal article · September 23, 2026 · Journal of Central Banking Law and Institutions
Myanmar, LaosLanguage and models
Shows that Burmese and Lao voice-cloning systems pronounce words more accurately when trained first on clean synthetic speech and only then on real recordings, with unreliable automatic transcripts down-weighted.
Preprint · September 22, 2026 · arXiv 2609.25951
MalaysiaSchools and universities
Finds that doctoral business students in Sarawak use AI quietly to cope with workloads and unclear rules rather than to cheat, but that the habit risks weakening their own judgement over time.
Journal article · September 22, 2026 · Journal of Education and e-Learning Research
VietnamSchools and universities
Tracks Vietnamese undergraduates over a semester and finds three learning profiles (optimal, moderate, maladaptive) defined by AI self-efficacy, cognitive offloading and metacognition.
Preprint · September 21, 2026 · Research Square rs-10468699/v1
Southeast AsiaLanguage and models
Technical report for the Nemotron-based SEA-LION v4.8 30B and 120B mixture-of-experts models, reporting SEA-HELM gains across seven SEA languages.
Preprint · September 16, 2026 · arXiv 2609.18310
Southeast AsiaJobs and the economy
Warns that most of the region rents frontier AI from a handful of foreign providers who can reprice or cut off access, and argues governments should make that dependence diversified, measurable and reversible rather than chase self-sufficiency.
Report · September 8, 2026 · AMRO Policy Perspectives PP-26-03
Southeast AsiaGovernment and rules
Maps the region's dependence across chips, cloud, data centres, energy and models, and makes the case that diversified sourcing and local capacity (AI resilience) is a more realistic goal than sovereignty.
Report · September 1, 2026 · Trends in Southeast Asia 2026/24
MalaysiaJobs and the economy
Explains how Johor became one of the world's largest emerging AI infrastructure clusters and the policy, geopolitical and grid problems that follow.
Report · September 1, 2026 · Trends in Southeast Asia 2026/22
SingaporeJobs and the economy
Finds that Singapore firms adopting AI grow revenue and headcount, with gains rising as AI capabilities deepen, but show no measurable productivity or profit improvement within four years of first adoption.
Report · August 11, 2026 · Economic Survey of Singapore 2Q 2026 feature article
PhilippinesJobs and the economy
After ChatGPT's release, employment growth in Philippine BPO stalled in the provinces most exposed to generative AI while continuing elsewhere, an early sign of displacement in an outsourcing hub.
Working paper · August 1, 2026 · SSRN
Indonesia, Philippines +2Schools and universities
Finds roughly 48 percent technical and 46 percent soft-skills gaps in the fintech workforce of four ASEAN economies, with AI and cybersecurity among the scarcest capabilities, and proposes a regional skills framework.
Report · August 1, 2026 · ADB report
Singapore, Malaysia +2Government and rules
Compares four ASEAN legal regimes to show how sandboxes, data-sharing rules and AI assurance mechanisms, not tax incentives, drive AI commercialisation and adoption.
Journal article · July 31, 2026 · IntechOpen (perspective chapter)
IndonesiaLanguage and models
Pancasila-Dilemmas evaluates how LLM responses to moral dilemmas align with Indonesia's state ideology values rather than Western value frameworks.
Preprint · July 20, 2026 · arXiv 2607.18066
ThailandHealth
Documents the multi-year build of a Thai chest X-ray AI (Inspectra CXR) that generalizes across hospitals with AUROC near 0.97 and is rated highly usable by radiologists.
Preprint · July 10, 2026 · arXiv 2607.09305
Southeast Asia, Indonesia +6Jobs and the economy
Puts nearly 80 million ASEAN workers (22.9 percent of employment) in GenAI-exposed occupations yet finds no sign of broad displacement so far, with women overrepresented in the most exposed jobs and a gap between exposure and institutional readiness.
Report · July 8, 2026 · ILO Brief
VietnamJobs and the economy
Surveys the evidence on how AI is reshaping work, digital skill demand and productivity in Vietnam as the country moves off a low-cost labour model, flagging skills gaps, adoption costs and job quality as the main constraints.
Working paper · July 1, 2026 · IZA Discussion Paper No. 18825
PhilippinesFarms, climate, and disasters
Finds same-day and prior rainfall dominate flood prediction in Western Visayas, with a sharp risk threshold around 27 to 32 mm per day.
Preprint · June 15, 2026 · Zenodo 20697452
MalaysiaJobs and the economy
Runs GenAI and agentic-AI scenarios for Malaysia's services sector and estimates roughly 680,000 jobs could be at risk, enough to lift unemployment from 3.5 to 5.5 percent, arguing current policy tools are inadequate.
Report · June 2, 2026 · ISEAS Perspective 2026/41
VietnamSchools and universities
A Vietnamese-curriculum LLM scores 87 percent on the 2025 national exam and cuts lesson-planning time from hours to under an hour with teacher oversight.
Preprint · June 2, 2026 · arXiv 2607.28647
IndonesiaLanguage and models
IndoBias measures ethnic, religious and regional stereotypes in LLM outputs across Indonesian languages using both explicit and implicit probing tracks.
Preprint · May 31, 2026 · arXiv 2606.01260
CambodiaLanguage and models
Compares embedding and generation models in a retrieval-augmented QA system over Khmer telecom documents, one of the few LLM studies with Khmer as the primary case.
Preprint · May 21, 2026 · arXiv 2605.22099
SingaporeGovernment and rules
Shows that people who engage with AI ethics questions use AI more and support its governance more, while an age, education and gender divide shapes who engages at all.
Report · May 1, 2026 · RSIS Policy Report
PhilippinesSchools and universities
Identifies profiles of AI dependency among Filipino college students and links them to self-reported declines in writing, research and critical-thinking competencies.
Journal article · April 30, 2026 · arXiv 2604.27349 / Proc. 9th Int. Conf. on Education and Multimedia Technology (ICEMT), 2025, pp. 299-302
Malaysia, IndonesiaLanguage and models
Produces annual 10 m oil palm plantation maps for Malaysia and Indonesia (2020 to 2024) by learning from noisy historical maps without new annotation.
Preprint · April 26, 2026 · arXiv 2604.23776
VietnamJobs and the economy
Finds 20.8 percent of Vietnamese employment (11.5 million jobs) is potentially exposed to GenAI but only 1.8 percent sits in the highest-exposure tier, with exposure concentrated in Hanoi, Ho Chi Minh City and Da Nang, in clerical work, and among women.
Report · April 1, 2026 · ILO Brief
PhilippinesHealth
Among Filipino students, habitual AI use and feeling emotionally safe with AI were the strongest predictors of willingness to turn to AI for mental-health support.
Preprint · March 30, 2026 · arXiv 2603.27994
LaosGovernment and rules
Provides the first structured assessment of Lao PDR's readiness to govern AI, identifying priority sectors and gaps in policy, institutional coordination, human capacity and infrastructure ahead of a national AI strategy.
Report · March 17, 2026 · UNESCO AI Readiness Assessment Methodology country report
Southeast AsiaLanguage and models
SEAHateCheck extends functional hate-speech tests to low-resource Southeast Asian languages so moderation systems can be checked for systematic failure modes beyond English.
Journal article · March 17, 2026 · arXiv 2603.16070 / TALLIP Accepted
MalaysiaGovernment and rules
Sorts the AI risks Malaysian stakeholders actually worry about into competitive risk, unintended harm and malicious use, and ties each to readiness measures.
Report · March 7, 2026 · KRI report
ThailandLanguage and models
ThaiSafetyBench introduces culturally grounded Thai jailbreak and harm prompts, including Thai legal and social taboos, to evaluate LLM safety.
Journal article · March 5, 2026 · arXiv 2603.04992
Indonesia, Malaysia +4Government and rules
Tracks how SEA-6 digital governance shifted from AI ethics principles toward binding rules, sovereign compute investment and safety-testing regimes during 2025.
Report · March 4, 2026 · TFGI annual report
Southeast AsiaGovernment and rules
Finds that Southeast Asian governments have leaned on voluntary frameworks and soft law backed by data-protection statutes, and traces the drivers pushing some members toward harder AI rules.
Report · February 23, 2026 · ISEAS Perspective 2026/13
MyanmarLanguage and models
BURMESE-SAN is the first holistic benchmark for evaluating LLMs on Burmese across understanding, reasoning and generation tasks.
Preprint · February 21, 2026 · arXiv 2602.18788
CambodiaJobs and the economy
Collects Cambodian applied research on using AI in agriculture, SMEs and public administration to map readiness and suggest policy paths.
Report · February 19, 2026 · CADT / KAS Digital Insights
PhilippinesJobs and the economy
Shows that active use of digital financial platforms is strongly tied to formal account ownership in the Philippines and that AI in finance is concentrated in large institutions, mostly in fraud detection and credit scoring.
Working paper · February 18, 2026 · PIDS Discussion Paper No. 2026-03
Cambodia, Indonesia +6Government and rules
Regional digital-rights review that includes AI-enabled surveillance (notably Myanmar's PSMS) and platform accountability across eight countries.
Report · February 16, 2026 · DigitalReach annual report
PhilippinesJobs and the economy
Estimates that about 12.7 million Philippine jobs (over a quarter of employment) have some GenAI exposure, with only 3.6 percent in the highest-risk band and women and young service and clerical workers most affected.
Report · February 5, 2026 · ILO Research Brief
SingaporeSchools and universities
Proposes a single national AI literacy framework for Singapore that consolidates existing competency models across schools, workplaces and the public.
Working paper · February 1, 2026 · IPS Working Papers No. 71
PhilippinesFarms, climate, and disasters
An offline phone app for Philippine cacao growers classifies disease with about 97 percent validation accuracy and agrees with expert technicians 84 percent of the time in the field.
Preprint · January 30, 2026 · arXiv 2602.00216
MyanmarGovernment and rules
Tallies 2025 internet shutdowns and online-speech arrests and details the AI-backed Person Scrutinization and Monitoring System, Tiangou gateway and Huawei banking AI used to identify dissidents.
Report · January 30, 2026 · MIP annual report
MyanmarGovernment and rules
Explains how the junta assembled facial recognition, biometric databases and telecom monitoring from Chinese vendors into an integrated surveillance state.
Report · January 30, 2026 · Fulcrum commentary
Indonesia, Philippines +2Jobs and the economy
Builds development-stage-adjusted AI exposure measures and shows workers in lower-income Asian economies, including Indonesia and the Philippines, are less exposed than those in rich countries mainly because tasks within the same occupation differ.
Working paper · January 1, 2026 · ADB Economics Working Paper No. 836
ThailandHealth
Evaluates the real-world diagnostic performance of an AI chest X-ray reader for pulmonary TB in a Thai district hospital.
Journal article · January 1, 2026 · Journal of Health Science of Thailand (thaidj.org)
PhilippinesGovernment and rules
The annual monitor's theme chapter asks how Philippine institutions can use AI to improve transparency and service delivery while managing digital divides, ethics and data privacy risks.
Report · December 23, 2025 · PIDS Economic Policy Monitor
VietnamGovernment and rules
VLegal-Bench tests LLMs on hierarchical, frequently revised Vietnamese legislation using questions organized by cognitive level.
Preprint · December 16, 2025 · arXiv 2512.14554
PhilippinesSociety and elections
Documents how Philippine newsrooms are using AI tools, what productivity and quality effects they report, and the ethical and workforce challenges that come with adoption.
Working paper · December 1, 2025 · PIDS Discussion Paper No. 2025-39
PhilippinesGovernment and rules
Measures the readiness of Philippine local governments to deploy AI and finds wide variation in digital infrastructure, staff capability and data governance across LGUs.
Working paper · December 1, 2025 · PIDS Discussion Paper No. 2025-48
Southeast AsiaLanguage and models
RICE-VL offers 28,000 expert-curated visual question answering items and grounding samples to test how well vision-language models understand cultural content from every ASEAN country.
Preprint · December 1, 2025 · arXiv 2512.01419
PhilippinesSchools and universities
Surveys Philippine college students on how, why and on what devices they use generative AI, finding heavy reliance on free smartphone tools and low confidence in output accuracy.
Journal article · November 30, 2025 · arXiv 2512.11821 / 2024 2nd International Conference on Technology Innovation and Its Applications (ICTIIA)
IndonesiaHealth
An AI-based decision-support prototype raised Indonesian GPs' cardiovascular risk-assessment accuracy and statin prescribing by roughly a quarter while slightly shortening decision time.
Journal article · November 25, 2025 · Journal of Medical Internet Research 27:e73131 (doi 10.2196/73131)
SingaporeGovernment and rules
Finds Singaporeans broadly back AI use, R&D and regulation and trust local institutions (government, universities, the military) to oversee it far more than the US, China or international bodies.
Report · November 21, 2025 · RSIS Policy Report
LaosLanguage and models
LaoBench is the first large multidimensional benchmark for Lao, covering cultural knowledge, curriculum-aligned schooling and Lao-Chinese-English translation.
Journal article · November 14, 2025 · arXiv 2511.11334 / ACL 2026 Main
Brunei, Southeast AsiaJobs and the economy
Uses an LLM to score occupational tasks for GenAI exposure across ASEAN+3, with Brunei as the worked country case, and finds moderate overall exposure concentrated in clerical, administrative and financial jobs.
Working paper · November 10, 2025 · AMRO Working Paper
Southeast Asia, Indonesia +5Government and rules
Sorts the SEA-6 into principle-led, hard-law and ecosystem-building approaches to AI governance and proposes four regional coordination priorities.
Report · November 1, 2025 · TFGI Policy Brief
LaosSchools and universities
Lao university students and lecturers report AI boosts efficiency but risks hollowing out foundational skills, amid uneven digital access.
Journal article · October 28, 2025 · Asian Education and Learning Review 4(1)
ThailandHealth
Reports how the Thai Department of Medical Services' home-grown AI Chest 4All tool performs at TB screening in a provincial cancer hospital.
Journal article · October 25, 2025 · The Thai Journal of Radiological Technology (ThaiJO)
ThailandLanguage and models
JAI-1 is a 75B-parameter Thai-centric LLM that grows the architecture during Thai adaptation rather than overwriting existing knowledge.
Preprint · October 8, 2025 · arXiv 2510.08620
Southeast AsiaGovernment and rules
Argues that China's lead in economically useful embodied AI and robotics creates openings for Southeast Asian manufacturers but leaves the region dependent on supply chains controlled by the two rival superpowers.
Working paper · October 3, 2025 · ISEAS Perspective 2025/74
VietnamGovernment and rules
Assesses Viet Nam's AI ecosystem after the 2025 Digital Technology Industry Law, documenting a 5,000 to 7,000 AI engineer base against a large IT talent shortfall and a fast-growing AI market.
Report · October 1, 2025 · UNESCO AI Readiness Assessment Methodology country report (SHS/BIO/2025/AI-RAM/VN)
Southeast AsiaGovernment and rules
Uses World Risk Poll data to compare how Southeast Asian publics perceive AI, data-security and privacy risks by country and gender, and checks whether national governance frameworks address those concerns.
Report · October 1, 2025 · CIIP policy brief
MalaysiaSchools and universities
Maps Malaysia's AI-in-education landscape under the National AI Roadmap and Digital Education Policy and compares it with regional peers.
Preprint · September 26, 2025 · arXiv 2509.21858
SingaporeHealth
In Singapore hospital vignettes, a pharmacist-plus-LLM co-pilot caught substantially more drug-related problems than pharmacists alone, especially serious-harm errors.
Journal article · September 24, 2025 · Cell Reports Medicine (PMC12629785); earlier version arXiv 2402.01741
IndonesiaFarms, climate, and disasters
A 2.7M-parameter detector runs at about 29 FPS on CPU with high mAP, aimed at low-cost fish monitoring in Indonesian waters.
Preprint · September 22, 2025 · arXiv 2509.17406
SingaporeLanguage and models
SGToxicGuard benchmarks how easily LLMs produce toxic content in Singapore's low-resource languages through conversation, QA and content-composition attacks.
Journal article · September 18, 2025 · arXiv 2509.15260 / 9 pages, EMNLP 2025
IndonesiaFarms, climate, and disasters
Releases an open, expert-labeled Indonesian oil-palm dataset (by growth stage) intended to train and benchmark geospatial foundation models for deforestation monitoring.
Preprint · September 10, 2025 · arXiv 2509.08303
SingaporeSociety and elections
Survey shows ChatGPT is the most used tool, education the most common use, and finds an emerging AI divide plus public worries about harm and ethics.
Report · September 1, 2025 · RSIS Policy Report
Southeast Asia, MalaysiaGovernment and rules
Argues that weak oversight of SEA data centres lets Chinese actors route around US chip export controls, exposing gaps in ASEAN's voluntary AI governance.
Report · August 26, 2025 · CSIS New Perspectives on Asia blog
Southeast Asia, Brunei +10Government and rules
Surveys AI safety and governance moves in every Southeast Asian country, including Timor-Leste, and argues the region is converging on a pragmatic, development-first model distinct from EU or US approaches.
Report · August 25, 2025 · Brookings Governance Studies report
ThailandLanguage and models
WangchanThaiInstruct is a native-written Thai instruction dataset spanning four domains that reveals gaps between translated and native Thai evaluation.
Journal article · August 21, 2025 · arXiv 2508.15239 / Accepted to EMNLP 2025 (Main)
MyanmarGovernment and rules
Documents the junta's AI-enabled biometric database, facial-recognition camera networks and firewall as a pipeline from digital identification to arrest and torture, and names the foreign supply chain enabling it.
Report · August 20, 2025 · Submission to the UN Special Rapporteur on counter-terrorism and human rights
PhilippinesLanguage and models
FilBench evaluates 27 LLMs on Filipino, Tagalog and Cebuano tasks reflecting local NLP priorities and finds large gaps, especially in generation and classification.
Preprint · August 5, 2025 · arXiv 2508.03523
CambodiaGovernment and rules
Maps Cambodia's AI readiness across governance, skills, data and infrastructure and feeds directly into the draft National AI Strategy 2025-2030, noting limited data, compute and talent alongside growing pilot use cases.
Report · July 15, 2025 · UNESCO AI Readiness Assessment Methodology country report
MalaysiaHealth
A web-based AI chatbot delivering HIV self-test kits to Malaysian MSM scored high on satisfaction and recommendation but failed on about 8 percent of interactions involving unscripted questions.
Journal article · July 15, 2025 · JMIR Human Factors (PMC12283058)
Southeast AsiaLanguage and models
SEALGuard trains a multilingual safety guardrail on a new dataset of unsafe and jailbreak prompts across ten SEA languages, substantially outperforming LlamaGuard on those inputs.
Preprint · July 11, 2025 · arXiv 2507.08898
SingaporeLanguage and models
RabakBench localizes a safety benchmark to Singlish, Chinese, Malay and Tamil via LLM-driven red-teaming and human validation, exposing gaps in commercial guardrails.
Preprint · July 8, 2025 · arXiv 2507.05980
MalaysiaJobs and the economy
Estimates that 4.2 million Malaysian workers (28 percent) are highly exposed to generative AI and that nearly half the workforce has at least 40 percent of tasks substitutable, with women, younger, clerical and urban workers most exposed.
Report · July 1, 2025 · World Bank report (joint with Institute of Strategic and International Studies Malaysia)
Timor-LesteGovernment and rules
Reports that Timorese respondents are heavy AI users (most had used it, over half daily) yet the country lacks any AI strategy, and recommends digital literacy and data governance frameworks.
Report · July 1, 2025 · UNESCO Readiness Assessment Methodology country report (unesdoc pf0000395617)
ThailandGovernment and rules
Finds Thailand has a national AI strategy but no AI-specific law or risk-based regulator, a shortfall of about 80,000 AI professionals, and persistent digital access gaps for rural, minority and elderly groups.
Report · June 26, 2025 · UNESCO AI Readiness Assessment Methodology country report
Brunei, Indonesia +5Health
Stakeholders across seven SEA health systems say AI uptake hinges on basic digital infrastructure and data/technology governance, with a sharp readiness gap between high-income and lower-middle-income members.
Journal article · June 16, 2025 · Journal of Medical Internet Research 27:e71591 (doi 10.2196/71591)
MalaysiaSchools and universities
Malaysia's top-performing students overwhelmingly say ChatGPT helps them grasp difficult topics, while worrying about accuracy and misuse.
Journal article · June 14, 2025 · Journal of Informatics and Web Engineering
IndonesiaSchools and universities
Evaluates state-of-the-art vision-language and language models on scoring real handwritten answers from Indonesian grade-4 classrooms and finds they are not yet reliable for automated assessment.
Journal article · June 5, 2025 · arXiv 2506.04822 / Accepted at AIED 2026
PhilippinesHealth
Sets out a governance roadmap for responsible clinical AI in the Philippine health system, drawing on the Mayo-Metro Pacific partnership.
Journal article · June 5, 2025 · npj Digital Medicine (doi 10.1038/s41746-025-01755-3)
IndonesiaLanguage and models
IndoSafety is a culturally grounded safety benchmark for Indonesian, Javanese and Sundanese that exposes safety weaknesses in both general and Indonesia-specific LLMs.
Preprint · June 3, 2025 · arXiv 2506.02573
Southeast Asia, Indonesia +4Jobs and the economy
Argues that robots have so far created more jobs than they destroyed in East Asia and Pacific, that AI exposure is lower in the region's developing economies but rising, and that digital platforms are reshaping informal work.
Report · June 1, 2025 · World Bank East Asia and Pacific Development Studies
Vietnam, Malaysia +3Schools and universities
Compares what students at open universities in Vietnam, Malaysia, Thailand, the Philippines and Indonesia want from emerging technologies including AI in virtual learning.
Journal article · June 1, 2025 · arXiv 2506.00922
ThailandSociety and elections
Tests how ChatGPT, Copilot and Gemini handle Thai taboo and legally sensitive topics (e.g., lese majeste, religion) and whether their refusals align with Thai social norms.
Journal article · May 27, 2025 · Journal of Integrative and Innovative Humanities 5(1)
PhilippinesSociety and elections
Frames AI as both the engine and a potential antidote for election disinformation in the Philippines' 2025 midterms and proposes regulatory, technical and literacy responses.
Journal article · May 13, 2025 · Data & Policy (Cambridge)
PhilippinesSociety and elections
Documents an April 2025 workshop with Filipino journalists and activists on AI-generated disinformation and red-tagging, and the verification techniques they were trained in.
Report · May 8, 2025 · WITNESS blog (workshop report)
Southeast AsiaGovernment and rules
Notes that only six ASEAN members have explicit AI strategies and proposes using the DEFA to harmonise AI governance, support SME adoption, build regional datasets and green AI infrastructure.
Report · May 1, 2025 · ERIA Policy Brief No. 2025-1
MalaysiaJobs and the economy
Weighs the energy and water demands of Malaysia's AI-driven data centre boom in Johor against the state's investment and sustainability commitments.
Working paper · May 1, 2025 · ISEAS Perspective 2025/43
MyanmarSociety and elections
Documents how AI-generated images and videos spread after the March 2025 Sagaing earthquake, compounding a response already hampered by conflict and telecom shutdowns.
Report · May 1, 2025 · IDSS Paper No. 055/2025
Southeast Asia, Singapore +4Health
Maps the thin but growing evidence base on AI-assisted cancer screening in ASEAN, finding most studies concentrated in a few countries and only half reaching prospective clinical evaluation.
Journal article · April 15, 2025 · BMC Cancer (PMC12001681)
MyanmarLanguage and models
Extends the XNLI inference benchmark to Myanmar and tests inexpensive approaches for a language largely absent from LLM training data.
Preprint · April 13, 2025 · arXiv 2504.09645
Southeast AsiaLanguage and models
Introduces the SEA-LION v3 family of open LLMs continually pretrained on Southeast Asian language data and evaluated on the SEA-HELM leaderboard.
Journal article · April 8, 2025 · arXiv 2504.05747 / Accepted at IJCNLP-AACL 2025 (Main Track)
VietnamGovernment and rules
Scores Vietnam's public-sector AI capacity, finding weak inter-agency data sharing and a shortage of in-house AI expertise, and proposes a roadmap for AI in government.
Report · April 1, 2025 · UNDP AILA country report
Southeast AsiaJobs and the economy
Pools existing studies to argue AI adoption raises Southeast Asian trade volumes, with effects amplified by digital infrastructure and regulation and partly mediated by firm size.
Preprint · March 15, 2025 · arXiv 2503.13529
Southeast AsiaLanguage and models
Builds SEA-VL, a culturally grounded vision-language dataset for Southeast Asia, and compares crowdsourcing, web crawling and generative synthesis as ways to obtain locally relevant images.
Preprint · March 10, 2025 · arXiv 2503.07920
SingaporeSociety and elections
Argues Singapore's government can use AI to listen to citizens at scale and shift public communication from broadcasting to a relational model.
Working paper · March 1, 2025 · IPS Working Papers No. 61
PhilippinesJobs and the economy
Maps which Philippine occupations are most exposed to AI and finds that roughly a third of workers are highly exposed, but most of those also sit in jobs where AI is likely to complement rather than replace them, with the BPO sector as the key watch point.
Working paper · February 21, 2025 · IMF Working Paper No. 2025/043
PhilippinesLanguage and models
Batayan is a holistic Filipino benchmark covering understanding, reasoning and generation, showing that most models still lag on Filipino despite strong English scores.
Journal article · February 19, 2025 · arXiv 2502.14911 / Accepted to ACL 2025 (Main Conference)
SingaporeLanguage and models
Compares alignment methods for reducing toxicity in Singlish and proposes a KTO variant that generalizes safety to a low-resource English variety.
Preprint · February 18, 2025 · arXiv 2502.12485
ThailandGovernment and rules
NitiBench benchmarks retrieval-augmented and long-context LLM setups on Thai legal question answering, showing retrieval and legal structure both limit accuracy.
Preprint · February 15, 2025 · arXiv 2502.10868
ThailandHealth
Thai health stakeholders confirm that infrastructure, policy and organizational barriers identified in the LMIC literature block the data sharing needed for clinical AI, and call for unified platforms and standards.
Journal article · February 4, 2025 · Journal of Medical Internet Research (PMC11836587)
MalaysiaGovernment and rules
Maps Malaysia's fragmented AI policy landscape, groups risks into falling behind, unintended harm and malicious use, and proposes six pathways including a central AI agency and an agile regulatory framework.
Report · January 28, 2025 · KRI report
ThailandJobs and the economy
Builds an algorithm that standardises Thai online job-posting data into occupation codes so that vacancy and skills trends can be tracked in near real time, filling a gap in Thai labour statistics.
Working paper · January 13, 2025 · PIER Discussion Paper No. 228
SingaporeJobs and the economy
Builds a Singapore-specific AI and LLM exposure index from task descriptions and job-posting data to measure how AI is changing labour demand across occupations.
Working paper · January 1, 2025 · SMU ResWORK research project
IndonesiaJobs and the economy
Two 2025 Bank Indonesia papers apply AI and machine learning to Indonesian central-bank problems, one to analyse policy communication and one to detect illegal online-gambling transaction patterns in payment data.
Working paper · January 1, 2025 · Bank Indonesia Working Paper Series
PhilippinesGovernment and rules
Recommends how Philippine universities, regulators and civil society should govern generative AI in research outputs, with attention to intellectual property.
Report · January 1, 2025 · DLSU Animo Repository, Asia-Pacific Intellectual Property Management and Innovation Book Series
IndonesiaFarms, climate, and disasters
A random forest model maps peat-fire vulnerability in Riau's Bengkalis Regency during El Nino with high accuracy, pointing to road density and rainfall as the key drivers.
Journal article · December 20, 2024 · MethodsX (PMC11741053)
ThailandLanguage and models
Typhoon 2 extends the Thai LLM family to text, vision and audio models with safety tuning, reporting strong Thai performance at small parameter counts.
Preprint · December 18, 2024 · arXiv 2412.13702
PhilippinesLanguage and models
Creates Filipino versions of sexism and anti-queer bias benchmarks and finds that regional SEA models exhibit these biases even more than their base models.
Journal article · December 10, 2024 · arXiv 2412.07303 / https://aclanthology.org/2025.loreslm-1.9/
MalaysiaGovernment and rules
Recommends a phased path for Malaysia that leans on soft law, a central coordinating body, an AI safety institute and a data governance centre rather than rushing to hard regulation.
Report · December 9, 2024 · ISIS Malaysia white paper
PhilippinesJobs and the economy
Assesses how ready Philippine firms are to adopt AI and finds infrastructure gaps, low awareness, skills shortages and scarce financing as the binding constraints, with recommendations for government support.
Working paper · December 1, 2024 · PIDS Discussion Paper No. 2024-35
PhilippinesSchools and universities
Proposes how Philippine universities should redesign curricula and assessment for a generative AI world, written for the EDCOM II education commission.
Report · December 1, 2024 · UP CIDS Monograph Series
Indonesia, PhilippinesGovernment and rules
Reviews the data-protection laws of Indonesia, the Philippines, Bangladesh, Nepal and Sri Lanka to see how far they can restrain risks from AI systems, finding significant shortfalls.
Report · December 1, 2024 · EngageMedia research report (Greater Internet Freedom project)
Timor-LesteLanguage and models
Analyzes real usage of tetun.org, a Tetun machine-translation service, showing that Timorese users, mostly students on phones, translate short texts in ways existing corpora do not reflect.
Journal article · November 19, 2024 · arXiv 2411.12262
Southeast AsiaGovernment and rules
Sketches how generative AI could be applied to health security threats in Southeast Asia and reviews ASEAN-level policy and governance activities relevant to it.
Journal article · November 3, 2024 · arXiv 2411.14435
ThailandHealth
Only about a third of Thai medical students had a basic grasp of AI, yet almost all wanted it in the curriculum and those who understood it better were more upbeat about AI in radiology.
Journal article · October 22, 2024 · BMC Medical Education (PMC11515691)
VietnamHealth
For a coastal Vietnamese province, a lagged-climate negative binomial model beat XGBoost and LSTM at weekly dengue forecasting, though none captured outbreak peaks well.
Journal article · October 21, 2024 · Tropical Medicine and Infectious Disease (PMC11511084)
PhilippinesSchools and universities
Filipino students in Eastern Visayas rate ChatGPT as moderately useful for research help but complain of redundant or inaccurate answers and worry about plagiarism.
Journal article · October 10, 2024 · Salud, Ciencia y Tecnologia - Serie de Conferencias
IndonesiaHealth
Prototypes an LLM pipeline that transcribes and summarizes doctor-patient conversations in Bahasa Indonesia directly into the ePuskesmas community health record system to cut documentation time.
Preprint · September 25, 2024 · arXiv 2409.17054
PhilippinesLanguage and models
Kalahi is a 150-prompt, native-authored evaluation suite that checks whether LLMs give culturally appropriate answers to everyday Filipino situations.
Journal article · September 20, 2024 · arXiv 2409.15380 / https://aclanthology.org/2024.paclic-1.49/
Southeast AsiaGovernment and rules
Explains why ASEAN has not produced a binding regional AI framework and proposes an ASEAN AI working group, reskilling initiatives and elevating AI on the regional agenda.
Journal article · August 30, 2024 · Frontiers in Artificial Intelligence
Southeast Asia, Singapore +5Government and rules
Maps national AI policies across six ASEAN economies and characterises the region's light-touch regulatory stance against its capacity gaps.
Report · August 21, 2024 · SEAPPI report
SingaporeJobs and the economy
Finds Singapore is among the most AI-exposed economies but also among the best prepared, with clerical, administrative and sales roles facing the largest displacement risk and upskilling as the main policy lever.
Working paper · August 13, 2024 · IMF Selected Issues Paper SIP/2024/040
SingaporeGovernment and rules
Reviews deepfake and AI disinformation threats to elections and what Singapore can adopt from other jurisdictions' safeguards.
Working paper · August 1, 2024 · IPS Working Papers No. 56
CambodiaSchools and universities
Synthesizes Cambodian reporting and research to list five concerns (privacy, critical thinking, misconduct, assessment, misinformation) and four benefits of AI in the country's classrooms.
Journal article · August 1, 2024 · Asian Journal of Distance Education 19(2), 2024; also SSRN 4932278
Southeast AsiaLanguage and models
Third-generation SeaLLMs built on Qwen2, adding more SEA languages and a training method that targets language-specific neurons to reduce hallucination and cost.
Preprint · July 29, 2024 · arXiv 2407.19672
SingaporeHealth
Benchmarks classical and deep models on Singapore's weekly dengue counts and finds CNN-based forecasts had the lowest error for 2019.
Preprint · June 29, 2024 · arXiv 2407.00332
IndonesiaGovernment and rules
Reviews how campaigns and buzzers used generative AI and deepfakes in Indonesia's 2024 election and draws regulatory lessons for the region.
Report · June 25, 2024 · Fulcrum commentary
Southeast AsiaLanguage and models
SEACrowd consolidates almost 500 text, image and audio datasets in about 1,000 SEA languages into one catalogue and uses it to benchmark models across 36 indigenous languages.
Journal article · June 14, 2024 · arXiv 2406.10118 / https://seacrowd
Southeast AsiaGovernment and rules
Examines how the 2024 ASEAN Guide on AI Governance and Ethics can be operationalised across member states with very different regulatory maturity and identifies the gaps between the guide and national practice.
Report · June 1, 2024 · Trends in Southeast Asia TRS18/24
Southeast Asia, Malaysia +5Government and rules
Shows how Grab and GoTo's algorithmic management isolates ride-hailing and delivery workers, blurs employer responsibility and concentrates platform power across the region.
Report · April 30, 2024 · Carnegie Endowment research
IndonesiaLanguage and models
Cendol releases a family of instruction-tuned LLMs for Indonesian and its regional languages, with gains on unseen tasks but limited ability to absorb human values via adaptation.
Journal article · April 9, 2024 · arXiv 2404.06138
Southeast AsiaLanguage and models
Releases the Sailor family (0.5B to 7B) of open models for Indonesian, Thai, Vietnamese, Malay and Lao, with ablations on data mixing and deduplication for SEA languages.
Preprint · April 4, 2024 · arXiv 2404.03608
IndonesiaLanguage and models
IndoCulture tests whether LLMs know culture-specific commonsense that varies across eleven Indonesian provinces, revealing large geographic knowledge gaps.
Journal article · April 2, 2024 · arXiv 2404.01854 / Accepted at TACL
VietnamLanguage and models
Fine-tunes open LLMs for Vietnamese and builds a broad Vietnamese evaluation framework, showing that data quality and instruction tuning matter more than scale for Vietnamese performance.
Journal article · March 5, 2024 · arXiv 2403.02715
Timor-LesteFarms, climate, and disasters
Tests whether AI-based drought forecasts can serve as reliable triggers for pre-emptive humanitarian and agricultural action in Timor-Leste, one of the few AI studies with the country as primary case.
Working paper · January 1, 2024 · FAO Technical Working Paper
PhilippinesSociety and elections
Makes the case for public investment in computer-vision and AI tools to monitor and track marine plastic litter along Philippine coasts.
Report · January 1, 2024 · UP CIDS Policy Brief Series 2024
Southeast AsiaLanguage and models
Presents SeaLLMs, LLaMA-based models extended to Thai, Vietnamese, Indonesian, Malay, Khmer, Lao, Tagalog and Burmese, with claims of stronger regional and safety performance than ChatGPT-3.5 on those languages.
Journal article · December 1, 2023 · arXiv 2312.00738 / Technical report, ACL 2024 DEMO TRACK
VietnamLanguage and models
PhoGPT-4B is an open Vietnamese generative model and chat variant pretrained from scratch, reported to outperform much larger multilingual models on Vietnamese instruction following.
Preprint · November 6, 2023 · arXiv 2311.02945
Southeast AsiaJobs and the economy
Argues ASEAN should put AI on the regional agenda now, prioritising narrowing readiness gaps between members, managing job displacement, building region-specific AI and coordinating with global regulation.
Working paper · November 1, 2023 · ERIA Discussion Paper No. 488 (ERIA-DP-2023-16)
IndonesiaLanguage and models
IndoMMLU collects 14,000 questions from Indonesian primary to university exams and shows that even GPT-3.5 only clears primary-school level, with especially weak local-language and culture knowledge.
Journal article · October 7, 2023 · arXiv 2310.04928 / Accepted at EMNLP 2023
Singapore, Southeast AsiaLanguage and models
SeaEval evaluates foundation models on both standard NLP tasks and Singapore- and Southeast-Asia-centred cultural reasoning, finding brittleness to paraphrase and inconsistency across languages.
Journal article · September 9, 2023 · arXiv 2309.04766 / Published at NAACL 2024
SingaporeGovernment and rules
Surveys regulatory options for AI harms and positions Singapore's voluntary, framework-based approach against emerging hard-law regimes.
Working paper · August 1, 2023 · IPS Working Papers No. 52
ThailandFarms, climate, and disasters
Builds a LINE-based chatbot that delivers crop cultivation recommendations to Thai farmers as a complement to IoT sensor-based smart farming.
Journal article · July 31, 2023 · arXiv 2308.02524 / 2022 Joint International Conference on Digital Arts, Media and Technology with ECTI Northern Section Conference on Electrical, Electronics, Computer and Telecommunications Engineering (ECTI DAMT and NCON), Chiang Rai university, Jan 2022, Chiang Rai, Thailand. pp.364-367
Laos, VietnamJobs and the economy
Develops a text-similarity method to port US-based AI exposure measures to developing-country skills surveys and finds urban Vietnamese workers face more displacement-type AI exposure than workers in Lao PDR.
Journal article · July 1, 2023 · Journal of Evolutionary Economics 33(3)
Southeast AsiaGovernment and rules
Early civil-society mapping of national AI strategies and rights risks across Southeast Asia, based on interviews with regional experts.
Report · June 6, 2023 · EngageMedia research report
Southeast AsiaLanguage and models
Tests whether multilingual LLMs can produce natural code-mixed text for seven SEA language pairs including Tagalog, Indonesian, Malay, Vietnamese and Singlish, finding most models fail except ChatGPT with caveats.
Preprint · March 23, 2023 · arXiv 2303.13592
MalaysiaGovernment and rules
Argues Malaysia's PDPA 2010 and constitutional privacy protections cannot govern police and commercial facial recognition, drawing on US and UK law for reform proposals.
Journal article · February 1, 2023 · Atlantis Press, Proceedings of ICLD 2022