Who this is for: AI providers and deployers operating in or targeting the Vietnamese market, foreign AI companies evaluating Vietnam market entry, compliance officers at multinational technology firms, legal counsel advising on ASEAN AI regulation, and GRC architects building cross-jurisdictional compliance frameworks.
In force since March 1, 2026. Vietnam's Law on Artificial Intelligence (Law 134/2025) is Southeast Asia's first comprehensive, binding AI law. Transitional grace periods: 12 months for general AI systems (full compliance by March 1, 2027) and 18 months for finance, healthcare, and education (full compliance by September 1, 2027). Foreign providers of high-risk AI must establish commercial presence or appoint an authorized representative in Vietnam.
Contents
1. Overview 2. Risk Classification 3. Key Obligations 4. Obligation-to-Procedure Mapping 5. Detailed Procedure Cards 6. Transition Timeline 7. Quick Reference 8. Quick Start 9. References1. Overview
Vietnam's Law on Artificial Intelligence (Law No. 134/2025/QH15), passed by the National Assembly on December 10, 2025 and effective March 1, 2026, is the first comprehensive, binding AI law in Southeast Asia. The law covers the full lifecycle of AI systems: research, development, provision, deployment, and use within Vietnam's jurisdiction.
Modeled partly on the EU AI Act but with Vietnam-specific provisions, Law 134/2025 introduces a risk-based governance framework that classifies AI systems into Low, Medium, and High risk tiers, each carrying graduated regulatory obligations. The law applies extraterritorially: foreign providers offering high-risk AI systems in Vietnam must establish a local commercial presence or appoint an authorized representative.
The framework is built on five regulatory pillars:
- Risk-based classification with tiered obligations scaled to the potential impact of each AI system on individuals and society
- Conformity assessment for high-risk AI, conducted by registered bodies or through self-assessment, completed before market entry
- Transparency and labeling requirements including disclosure of AI interaction and machine-readable marking of AI-generated content
- Technical documentation and logging requiring risk management records, architecture documentation, and continuous audit trails
- Incident handling and governance structures mandating detection, response, and reporting protocols with designated responsible parties
The law provides transitional grace periods: 12 months for general AI systems and 18 months for systems operating in finance, healthcare, and education. These grace periods allow organizations currently operating AI in Vietnam to bring systems into compliance without immediate enforcement exposure.
2. Risk Classification
Law 134/2025 establishes a three-tier risk classification that determines the level of regulatory obligation applied to each AI system. Classification is based on the potential impact on individuals, society, and national security.
| Risk Tier | Scope | Key Requirements |
|---|---|---|
| High Risk | AI systems with significant potential impact on health, safety, fundamental rights, or critical infrastructure. Includes biometric identification, credit scoring, employment decisions, law enforcement, and essential public services. | Conformity assessment (registered body or self-assessment), risk management system, technical documentation, continuous logging, transparency disclosure, incident handling protocol, human oversight. Foreign providers must establish local presence or appoint authorized representative. |
| Medium Risk | AI systems with moderate potential impact. Includes general-purpose content generation, customer service automation, and recommendation systems affecting consumer behavior. | Reduced documentation requirements, self-assessment permitted, transparency disclosure, basic logging, incident reporting for serious events. |
| Low Risk | AI systems with minimal potential impact. Includes spam filters, search optimization, basic automation, and internal productivity tools. | General transparency requirements only. Must disclose AI involvement when interacting with individuals. |
3. Key Obligations
| Obligation | Applicability | Compliance Deadline |
|---|---|---|
| AI Interaction Disclosure | All risk tiers | March 1, 2027 (general) / September 1, 2027 (finance, healthcare, education) |
| Synthetic Content Labeling | All AI-generated content | March 1, 2027 (general) / September 1, 2027 (finance, healthcare, education) |
| Technical Documentation | Medium and High risk | March 1, 2027 (general) / September 1, 2027 (finance, healthcare, education) |
| Conformity Assessment | High risk only | March 1, 2027 (general) / September 1, 2027 (finance, healthcare, education) |
| Governance Structure | High risk, all foreign providers of high-risk AI | March 1, 2027 (general) / September 1, 2027 (finance, healthcare, education) |
| Non-Discrimination | High and Medium risk | March 1, 2027 (general) / September 1, 2027 (finance, healthcare, education) |
| Human Oversight | High risk | March 1, 2027 (general) / September 1, 2027 (finance, healthcare, education) |
| Incident Handling | High and Medium risk (serious incidents) | March 1, 2027 (general) / September 1, 2027 (finance, healthcare, education) |
| Audit Logging | High risk, continuous | March 1, 2027 (general) / September 1, 2027 (finance, healthcare, education) |
| Data Governance | High and Medium risk | March 1, 2027 (general) / September 1, 2027 (finance, healthcare, education) |
4. Obligation-to-Procedure Mapping
Each obligation under Law 134/2025 maps to SWT3 witness procedures that produce cryptographically anchored evidence of compliance.
| Vietnam AI Law Obligation | SWT3 Procedure | What It Witnesses | Evidence Produced |
|---|---|---|---|
| AI Interaction Disclosure | AI-TRANS.1 | User notified of AI interaction | Anchor with disclosure type, recipient type, delivery timestamp |
| Synthetic Content Labeling | AI-MARK.1 | Machine-readable content marking applied | Anchor with content type, marking method, verification status |
| Technical Documentation | AI-MDL.1 | Model documentation maintained and current | Anchor with model hash, version, documentation status |
| Component Documentation | AI-SBOM.1 | AI component manifest maintained | Anchor with component list, version, dependency chain |
| Governance Structure | AI-GOV.1 | Organizational governance and representative appointment | Anchor with governance body, representative details, jurisdiction |
| Non-Discrimination | AI-FAIR.1 | Fairness assessment completed | Anchor with assessment type, protected attributes, bias metrics |
| Human Oversight | AI-HITL.1 | Human review in high-risk decision processes | Anchor with oversight type, reviewer role, decision authority |
| Incident Handling | AI-INCIDENT.1 | Incident detection and response protocol active | Anchor with incident type, severity, response time, resolution |
| Audit Logging | AI-AUDIT.1 | Continuous audit trail maintained | Anchor with log scope, retention period, integrity verification |
| Data Governance | AI-DATA.1 | Training data provenance and quality tracked | Anchor with dataset identifier, provenance hash, quality score |
5. Detailed Procedure Cards
Transparency Disclosure
Vietnam requires: Users must be clearly informed when they are interacting with an AI system. AI-generated content must be labeled in a machine-readable format. This obligation applies across all risk tiers, making it the most broadly applicable requirement under Law 134/2025.
How SWT3 addresses it: The witnessTransparency() call records the disclosure type, recipient category, and delivery timestamp, creating an immutable anchor proving the disclosure occurred before or concurrent with the AI interaction. For content labeling, pair with witnessContentMarking() via AI-MARK.1 to cover both the interaction disclosure and content marking requirements in a single evidence chain.
AI-TRANS.1 anchors must predate or be concurrent with any AI-generated output. For content labeling, cross-reference with AI-MARK.1 anchors to demonstrate that both interaction disclosure and content marking are functioning. Factor A identifies the disclosure type (interaction notice, content label). Factor B records the recipient category. A gap between AI-INF.1 and AI-TRANS.1 anchors indicates undisclosed AI interactions.
Model Registry and Technical Documentation
Vietnam requires: Technical documentation for AI systems, including system architecture, capabilities, limitations, and intended use. High-risk and medium-risk systems must maintain documentation that is current and accessible for regulatory inspection.
How SWT3 addresses it: The witnessModelRegistry() call anchors the model hash, version identifier, and documentation status. Each anchor proves the documentation matches the deployed model at a specific point in time. When a model is updated, a new anchor with an updated hash creates a verifiable chain showing that documentation was refreshed alongside each deployment.
AI-MDL.1 anchors should correlate with deployment events. A hash mismatch between consecutive anchors indicates an undocumented model change. Factor A contains the model hash, Factor B the version, and Factor C the documentation status. Cross-reference with AI-SBOM.1 anchors for component-level documentation coverage.
Governance Structure and Representative Appointment
Vietnam requires: High-risk AI providers must maintain governance structures with designated responsible parties. Foreign providers of high-risk AI must establish a commercial presence in Vietnam or appoint an authorized representative. All other foreign providers must maintain a lawful contact point.
How SWT3 addresses it: The witnessGovernance() call records the governance body composition, representative appointment details, jurisdiction designation, and policy version. This creates verifiable proof that governance structures are established and that a Vietnam-based representative has been formally appointed before market entry.
AI-GOV.1 anchors should include jurisdiction metadata indicating "VN" for Vietnam-specific governance. Factor A identifies the governance body. Factor C records the representative details and jurisdiction. Verify that the appointment anchor predates the first AI-INF.1 anchor for Vietnam-targeted deployments. For foreign providers, this is the primary evidence of compliance with the local presence requirement.
Non-Discrimination and Fairness Assessment
Vietnam requires: AI systems must operate fairly and must not discriminate against individuals or groups. High-risk and medium-risk systems are subject to ongoing fairness monitoring and bias testing obligations.
How SWT3 addresses it: The witnessFairness() call records the assessment type, protected attributes evaluated, bias metrics, and mitigation actions taken. Each assessment produces an anchor that documents both the testing methodology and the outcome, creating an auditable record of continuous fairness monitoring across all protected categories.
AI-FAIR.1 anchors should appear at regular intervals, demonstrating continuous bias monitoring. Factor A identifies the assessment type and protected attribute tested. Factor B contains the bias metric results. A gap in AI-FAIR.1 anchors may indicate a period without fairness testing. Cross-reference with AI-HITL.1 anchors to show that human reviewers are evaluating fairness outcomes in high-risk contexts.
Incident Detection and Response
Vietnam requires: AI system operators must implement incident detection and response procedures. Serious incidents involving high-risk and medium-risk systems must be reported to regulatory authorities. Operators must maintain documented protocols for incident classification, escalation, and resolution.
How SWT3 addresses it: The witnessIncident() call records the incident type, severity classification, response time, and resolution status. Each incident creates an immutable timeline anchor that proves detection, response, and resolution occurred within prescribed timeframes. The chain of AI-INCIDENT.1 anchors from detection through resolution provides the complete incident lifecycle for regulatory reporting.
AI-INCIDENT.1 anchors create a complete incident chain. The time delta between the first anchor (detection) and subsequent anchors (response, resolution) demonstrates compliance with response time requirements. Factor A identifies the incident type. Factor B records the severity classification. Cross-reference with AI-SAFE.1 for safety-related incidents and with AI-AUDIT.1 to confirm the audit trail remained intact throughout the incident lifecycle.
6. Transition Timeline
| Date | Milestone | What It Means |
|---|---|---|
| December 10, 2025 | Law 134/2025 passed by National Assembly | Legislative text finalized and published |
| March 1, 2026 | Law enters into force | Regulatory framework active. Grace periods begin. New AI systems should begin compliance planning. |
| March 1, 2027 | Full compliance, general AI systems | 12-month grace period ends. All general AI systems must meet applicable obligations. Enforcement begins. |
| September 1, 2027 | Full compliance, finance, healthcare, and education | 18-month grace period ends. AI systems in regulated sectors must meet all applicable obligations. Enforcement begins. |
Current status (July 2026): The law is in force. Organizations have approximately 8 months remaining for general AI systems and 14 months for finance, healthcare, and education sector systems. Foreign providers evaluating Vietnam market entry should begin governance structure setup, representative appointment, and evidence collection now to avoid compressed timelines.
7. Quick Reference
| Examiner Question | Where to Look |
|---|---|
| Does the AI system disclose its nature to users? | AI-TRANS.1 anchors. Factor A identifies the disclosure type. Anchor timestamp must predate or match the AI interaction timestamp. Applies to all risk tiers. |
| Is AI-generated content labeled in machine-readable format? | AI-MARK.1 anchors. Factor A identifies content type. Factor B records the marking method (C2PA, watermark, metadata). Factor C confirms verification status. |
| Has the foreign provider appointed a Vietnam representative? | AI-GOV.1 anchors with jurisdiction = "VN". Factor C records representative details. Appointment anchor must predate first Vietnam-targeted deployment. |
| Is the AI system classified by risk tier? | Crosswalk metadata on the tenant record. AI-MDL.1 anchors document the system architecture and risk-relevant capabilities. AI-SBOM.1 confirms component manifest. |
| How is fairness monitored across protected categories? | AI-FAIR.1 anchors at regular intervals. Factor A identifies the protected attribute. Factor B contains metric results. Gaps indicate periods without monitoring. |
| Is there meaningful human oversight for high-risk decisions? | AI-HITL.1 anchors. Factor B identifies the reviewer. Factor C confirms override authority. Ratio of AI-INF.1 to AI-HITL.1 anchors shows oversight coverage. |
| What happens when an incident is detected? | AI-INCIDENT.1 anchor chain from detection through resolution. Factor B records severity. Time delta between anchors demonstrates response time compliance. |
| Is the audit trail intact and tamper-evident? | AI-AUDIT.1 anchors. Factor B contains the integrity hash. Factor C documents retention period. Cross-reference with Merkle rollup for daily root verification. |
8. Quick Start
pip install swt3-ai
# Initialize with the international AI profile
swt3 init --profile international-ai --tenant YOUR_TENANT
# Run the demo to see witness anchors generated
python -m swt3_ai.demo
# Or use TypeScript
npm install @tenova/swt3-ai
npx swt3-init --profile international-ai
Full SDK documentation: sovereign.tenova.io/docs
Create a free account: sovereign.tenova.io/signup
9. References
- Law No. 134/2025/QH15 on Artificial Intelligence, National Assembly of Vietnam, December 10, 2025
- Baker McKenzie: Vietnam AI Law Analysis, February 2026
- OneTrust: Vietnam AI Law Overview and Compliance Planning
- IAPP: Southeast Asia AI Regulation Tracker
- Duane Morris: Vietnam's AI Law Takes Effect, March 2026
- Related: South Korea AI Framework Crosswalk
- Related: EU AI Act Crosswalk
- SWT3 UCT Registry, 103 AI procedures across 52 namespaces
- SWT3 Bidirectional Framework Crosswalks, machine-readable JSON
- SDK Documentation, Python, TypeScript, and 5 additional language SDKs