Short answer: when must AI content be labelled?
In Vietnam, Article 11 of the Law on Artificial Intelligence No. 134/2025/QH15 (in force since 1 March 2026) and Articles 17 and 18 of Decree 142/2026/ND-CP (in force since 1 May 2026) create two layers of duty. Providers of AI systems must apply machine-readable marking to AI-generated audio, images and video. Organisations that publish content (deployers) must clearly notify the public when AI content could mislead, and must attach an easily recognisable label when audio, images or video simulate a real person or re-enact a real event.
Put simply: if viewers could take it for real, label it. If AI only fixed spelling, summarised, translated or improved image quality without changing the substance, the Decree exempts it from a visible label. The sections below quote the provisions, compare them with Google, Meta and YouTube policies, and give a review workflow a content team can run tomorrow.
Article 11 of the AI Law: the transparency provisions
The National Assembly passed the Law on Artificial Intelligence on 10 December 2025. It has 8 chapters and 35 articles, and Article 34 sets the effective date at 1 March 2026. The labelling duty sits in Article 11, “Transparency responsibility”. The content-related clauses are below (our English rendering of the official Vietnamese text):
Note the definition in Article 3: a “deployer” is an organisation or individual using an AI system under its control in professional, commercial or service activities, excluding personal, non-commercial use. A business using AI tools for ads, social posts or product videos will usually be a deployer. For an overview of the whole Law, see Vietnam's AI Law 2025.
- Clause 2: providers ensure that audio, images and video generated by AI systems are marked in a machine-readable format as prescribed by the Government.
- Clause 3: deployers must clearly notify when providing to the public text, audio, images or video created or edited by AI if that content may cause confusion about the authenticity of events or characters, unless the law provides otherwise.
- Clause 4: deployers must ensure that audio, images or video created or edited by AI to simulate or imitate the appearance or voice of a real person, or to re-enact an actual event, carry an easily recognisable label distinguishing them from real content. For cinematic, artistic or creative works, labelling is done in a suitable way that does not hinder enjoyment of the work.
- Clause 5: providers and deployers must maintain this transparency information throughout the time the content is provided to users.
- Clause 6: the Government details the form of notification and labelling (now done in Decree 142/2026/ND-CP).
Decree 142/2026/ND-CP: machine-readable marking and visible labels
On 30 April 2026 the Government issued Decree 142/2026/ND-CP detailing the AI Law, effective 1 May 2026. Article 1 states that it details clause 6 of Article 11 on the form of notification and labelling. Articles 16, 17 and 18 are the ones to read closely.
Article 17: technical marking (provider duty)
AI system providers must apply a technical solution so that audio, image or video outputs are marked in a machine-readable format. Providers are not required to apply machine-readable marking to text outputs unless the law provides otherwise. The marker must indicate that the content was created or edited by AI; where technically possible it may also carry the provider, the system and the time of creation to support provenance and tracing.
- Embedding an identifier in the file structure or content data.
- Embedding an identifier in the file's metadata.
- Using a digital signature, electronic signature or equivalent technical authentication.
- Any other technical solution that makes AI-created or AI-edited content identifiable.
Article 17(5): open-source and free models
For open-source or free AI systems, the provider is deemed compliant if marking is built in, or if it publishes the tools, configuration or documentation that let deployers switch marking on. In that case, a deployer using the system to publish content must apply marking to the output itself. Teams that self-host image or voice models often miss this point.
Article 18: notification and visible labels (deployer duty)
Clause 1 requires deployers to clearly notify when publishing AI-created or AI-edited content that may cause confusion about the authenticity of events, characters or the origin of the content. Clause 2 makes an easily recognisable label mandatory for AI-created or AI-edited audio, images or video that (a) simulate the appearance or voice of a real person, or (b) re-enact an actual event. Clause 3 sets five requirements: clear, understandable and recognisable; shown before or at the moment the audience accesses the content; not designed to conceal; suited to the content type and channel; and not significantly obstructing display or use.
Clause 5 lets deployers choose the form: shown directly on the content; in the title, description or caption; in the platform interface; or as an audio announcement. Clause 6 allows films, art programmes and creative content to place the label in the opening, ending, credits, description or accompanying material. Where AI is used during production (for example by an agency working for a brand), the party creating the content must supply the necessary information and the party publishing it applies the label. Clause 7 tasks the Ministry of Science and Technology with publishing reference technical guidance on label formats.
Mandatory or exempt: a quick reference table
Article 18(4) of Decree 142/2026/ND-CP lists four cases where a visible label is not required. The table maps common marketing and communications scenarios to the relevant provisions. It reflects our reading for internal guidance; borderline cases should go to a lawyer.
| Scenario | Duty | Legal basis |
|---|---|---|
| Video with an AI voice imitating a real person (a celebrity, a company executive) | Easily recognisable label required | Law Art. 11(4); Decree Art. 18(2)(a) |
| AI image or video re-enacting a real event (a launch, an accident, a disaster) | Easily recognisable label required | Law Art. 11(4); Decree Art. 18(2)(b) |
| AI-generated article or image that could mislead about events, people or origin | Clear notification | Law Art. 11(3); Decree Art. 18(1) |
| AI used for spelling, grammar, summarising, paraphrasing or translation without distorting the original | No visible label required | Decree Art. 18(4)(b) |
| AI denoising, sharpening or colour correction that does not change substance or context | No visible label required | Decree Art. 18(4)(a) |
| Drafts and internal-only material not released to the public | No visible label required | Decree Art. 18(4)(c) |
| Content created during research, development or testing and not released to the public | No visible label required | Decree Art. 18(4)(d) |
| Films, music videos and creative works using AI | Label in a suitable place: opening, ending, credits, description | Law Art. 11(4); Decree Art. 18(6) |
Deepfakes and prohibited conduct
A label does not make everything acceptable. Article 7 of the AI Law prohibits developing, providing, deploying or using AI systems to use forged or simulated elements of real people or events to deliberately and systematically deceive or manipulate human perception and behaviour in a way that seriously harms people's legitimate rights (clause 2(b)), and to create or disseminate fabricated content that may seriously harm national security, social order and safety (clause 2(d)).
Article 7(5) prohibits concealing information that must be disclosed and erasing or falsifying mandatory information, labels or warnings in AI activities. In practice, cropping out watermarks, stripping AI metadata or hiding a label where nobody sees it is a legal risk, not just a process slip. Under Article 29, violations may lead to administrative penalties or criminal liability depending on their nature and severity, plus compensation for any damage.
For content involving real people, a label is not enough on its own: get written consent from the person being simulated and check image rights and copyright. A safe internal rule: do not create content showing a real person saying or doing something they did not say or do, even with a label, unless you have written consent and a clear purpose.
Platform policies: Google, YouTube, Meta
Vietnamese law is the floor; content published on global platforms must also follow each platform's own rules. The points below come from the platforms' official pages at the time of writing and may change.
Google Search
Google does not penalise content simply for using AI, but using automation, including AI, to generate content primarily to manipulate rankings violates its spam policies. Google Search Central's documentation notes that generative models can produce inaccuracies (hallucinations) and that it is “critical” to manually fact-check and review all AI-generated content before publishing, including titles, meta descriptions, structured data and alt text. Google also suggests telling readers how content was produced. For Google Merchant Center, AI-generated images must carry IPTC DigitalSourceType TrainedAlgorithmicMedia metadata.
YouTube
YouTube requires creators to disclose when they use AI to meaningfully alter or generate realistic content, for example making a real person appear to say or do something they did not, altering footage of a real event or place, or generating a realistic scene that did not happen. The disclosure is the “AI use” setting under Attributes when uploading in YouTube Studio. Production assistance such as AI-drafted outlines, scripts, titles or thumbnails, and minor aesthetic edits, does not need disclosure. YouTube says disclosure does not limit audience or monetisation, that it may apply labels to undisclosed content, and that it carries forward “Made with AI” information from Content Credentials (C2PA) 2.1 or higher.
Meta (Facebook, Instagram, Threads)
Meta shows an “AI info” label on Facebook, Instagram and Threads when it detects industry-standard AI indicators in a file or when people disclose that they are uploading AI-generated content. For ads, Meta labels images created or significantly edited with its own generative AI features or with third-party AI tools. Keeping AI metadata intact when exporting files therefore helps platform labels work correctly and lines up with the Decree's machine-readable marking duty.
C2PA and Content Credentials: a technical option for machine-readable marking
Decree 142/2026/ND-CP does not mandate a specific standard; it allows metadata, digital signatures or equivalent solutions, and the Ministry of Science and Technology will publish a list of reference standards and technical guidance. Until then, C2PA is worth considering because it combines two of the methods the Decree names: metadata and digital signatures.
The Coalition for Content Provenance and Authenticity (C2PA) publishes an open technical standard called Content Credentials for establishing the origin and edit history of digital content. Each Content Credential (a C2PA Manifest) contains assertions about creation, edits and capture devices; it is digitally signed with X.509 certificates and bound to the file by cryptographic hashes, so any change becomes detectable. The specifications are released under a royalty-free licence and can record actions performed by AI systems.
Know the limits: C2PA's own explainer acknowledges that provenance is not always complete, for example when a file is cropped in a tool that does not support Content Credentials, and some platforms may strip metadata on upload. Machine-readable marking therefore does not replace a visible label for the audience; you need both.
A 7-step review workflow for AI content
The workflow below turns legal duties into concrete control points. It complements the AI marketing workflow and the human-in-the-loop principle: an AI Agent can draft, check formats and suggest labels, while the decision to approve and publish stays with an accountable person.
| Step | What to do | Owner | Evidence kept |
|---|---|---|---|
| 1. Brief | Record purpose and channel, whether a real person's likeness or voice is used and whether a real event is re-enacted; assign a risk level | Content lead | Brief with “AI used” and “real person / real event” fields |
| 2. Draft | Generate with approved AI tools; switch on technical marking where supported; log tool, version and key prompts | Editor or AI Agent within its permissions | Generation log; original file with metadata intact |
| 3. Fact-check | Verify every number, quote, name and legal reference against its source; include titles, descriptions and alt text | A checker other than the author | List of sources opened; notes on corrections |
| 4. Label | Use the quick reference table to decide label, notification or exemption; choose a form under Decree Art. 18(5); confirm AI metadata is intact | Editor | Screenshot of the label in its final display position |
| 5. Approve | Approve content and label; content with real people needs written consent; send borderline cases to legal | Team lead; legal when needed | Named, timestamped approval |
| 6. Publish | Complete platform AI disclosures (YouTube “AI use”, Meta AI info); check the label shows before or as the audience reaches the content | Publisher | Post URL and disclosure settings |
| 7. Archive | Keep the original file, published version and approval log; monitor feedback and correct or remove when errors are found | Content lead | Content record under internal retention policy |
Sample label wording
Illustrative example (not an official template; update it once the Ministry of Science and Technology publishes reference technical guidance):
- Video with an AI voice imitating a real person: a fixed on-screen caption “The voice in this video was generated with AI” from the first second, repeated in the description.
- Image re-enacting an event: a caption directly under the image “Illustration generated with AI, not a real photograph”.
- Article with AI-drafted passages that could mislead: a note at the top stating which parts AI assisted with and who verified them.
- Podcast or audio ad: a spoken notice at the start of the programme.
Who does what with agencies and AI Agents
Article 18(6) splits roles clearly: whoever creates or edits the content supplies the information, and whoever publishes it applies the label. When hiring an agency, put the duty to disclose AI tools used and to hand over files with metadata intact into the contract. When an AI Agent is part of the workflow, configure it to log tools and versions and to propose a label type, while people carry out approval and publishing.
How EYC approaches it
EYC builds AI for content and media on a human-led principle: AI Agents help draft, record provenance and suggest labels, while editors and accountable owners fact-check, approve and publish. If your team is standardising an AI-assisted content process, contact EYC to discuss a suitable scope.
Frequently asked questions
Does every AI-written article need a label?
No. Article 11(3) of the AI Law requires clear notification when AI-created or AI-edited text may cause confusion about the authenticity of events or characters. Article 18(4)(b) of Decree 142/2026/ND-CP exempts text processed by tools for spelling, grammar, summarising, paraphrasing or translation from a visible label when the original meaning is not distorted.
When did Decree 142/2026/ND-CP take effect?
The Government issued Decree 142/2026/ND-CP on 30 April 2026 and it took effect on 1 May 2026. The Law on Artificial Intelligence 134/2025/QH15 took effect on 1 March 2026.
Where should an AI content label be placed?
Article 18(5) of Decree 142/2026/ND-CP allows the label directly on the content; in the title, description or caption; in the platform interface; or as an audio announcement. Whatever the form, it must be clear, appear before or when the audience accesses the content, and must not be designed to conceal.
Is removing an AI watermark or metadata a violation?
Article 7(5) of the AI Law prohibits erasing or falsifying mandatory information, labels or warnings in AI activities. Deliberately removing a mandatory marker or label is therefore a significant legal risk.
Who labels AI video produced by an agency?
Under Article 18(6) of Decree 142/2026/ND-CP, the party creating or editing the content supplies the necessary information and the party that publishes it notifies and labels. Usually the publishing brand applies the label; set out the split of roles in the contract.
Is C2PA mandatory in Vietnam?
No. Article 17 of Decree 142/2026/ND-CP permits several methods such as metadata, digital signatures or equivalent solutions, and tasks the Ministry of Science and Technology with publishing reference standards. C2PA is one technical option that fits these methods.