This article was originally published by IAPP linked here.
Vietnam’s approach to artificial intelligence regulations crosses many topics and sectors, with a common theme emerging: human‑centered, state‑supervised and legally accountable.
The country’s policy direction is clearly reflected in its first standalone Law on Artificial Intelligence, which took effect March 2026. At the same time, Vietnam continues to promulgate sectoral regulations that govern how AI systems are trained, deployed and commercialized.
One such law is Decree No. 134/2026/ND‑CP, issued 6 April 2026 and effective 9 April 2026. It amends Decree No. 17/2023/ND‑CP guiding the implementation of Vietnam’s Law on Intellectual Property regarding copyright and related rights.
Decree No. 134 provides the primary analytical framework for examining AI‑copyright interaction in Vietnam. Against this framework, emerging legal approaches in Vietnam, China and the United States can be compared across key issues of authorship, training data and downstream liability.
Decree No. 134 is a copyright decree, but it plays a broader role in Vietnam’s digital and AI regulatory architecture by providing detailed guidance on authorship, ownership, statutory exceptions, registration procedures and enforcement mechanisms. It bridges copyright law and AI regulation by embedding the principles of human control and legal accountability into the use of AI systems.
In doing so, it introduces Vietnam’s most detailed rules to date on AI‑assisted creation, text and data mining for training, opt‑out rights for rightsholders and downstream compliance obligations.
For organizations developing or deploying AI systems in Vietnam, copyright law is no longer a peripheral consideration, but a compliance obligation that flows alongside data protection, cybersecurity and AI risk management.
Human authorship remains the legal anchor
The most consequential AI‑specific provision within Decree No. 134 is the newly added Article 5a, which governs the emergence of copyright and related rights where AI systems are used in the creative process. Aligning with trends emerging in the U.S., Europe and Latin America, the decree adopts a clear position that copyright protection arises only where a human makes a significant and decisive contribution.
Article 5a reinforces the position through a non‑exhaustive but unusually detailed description of qualifying human involvement, including supplying original input data or technical parameters, designing system architecture or documentation, crafting prompts that meaningfully steer the AI system, selecting or editing outputs and exercising final control so the result reflects human intent rather than automated processes.
This approach broadly aligns with China’s emerging judicial position, where courts have generally declined to recognize AI as an author, focusing instead on whether a human has exercised sufficient intellectual judgment, selection and control over the output.
The key difference lies in technique. While Chinese courts develop this approach through case‑by‑case assessment rather than legislative rules, Vietnam’s Decree No. 134 translates the same substantive logic into an ex ante, process‑oriented legislative framework that prioritizes regulatory clarity and predictability. In April 2026, the Supreme People’s Court of China indicated it is in the process of formulating judicial guidance on the adjudication of AI-related disputes. While no further details are currently available, this suggests greater clarity is on the horizon.
Similarly, AI guidance from the U.S. Copyright Office requires “dominant, creative force” by a human to convey copyright authorship rights. Neither the law nor court cases recognize a large language model or AI as having such rights. The U.S. Copyright Act of 1976 requires all works to be authored by a human being.
Where these conditions are met, the human contributor is recognized as the author. Conversely, Decree No. 134 also implicitly draws a boundary around insufficient human involvement. Activities such as supplying generic prompts, passively accepting AI‑generated outputs or exercising only formal oversight without creative judgment are unlikely to satisfy the threshold of significant and decisive contribution. In such cases, outputs risk falling outside copyright protection altogether, regardless of their commercial or artistic value.
Documentation and proof obligations for AI‑assisted creativity
Article 5a also introduces practical compliance consequences. When asserting copyright protection for AI‑assisted works, creators must be able to demonstrate their creative process and truthfully disclose AI use upon request by competent authorities.
The decree expressly recognizes evidentiary materials such as input datasets, prompts and interaction histories, as well as intermediate drafts and revision records, together with documentation that describes how human control and creative judgment were in fact exercised.
Under the laws of the U.S. Copyright Office, an author must make mandatory disclosures and disclaimers about their creative process by clearly separating and disclaiming AI-generated material while providing a detailed, narrative description of the specific human creative choices made to modify, arrange or combine those outputs into a final, copyrighted work.
By comparison, in China similar considerations have begun to emerge in judicial practice and certain regulatory contexts, but have not yet been systematically incorporated into the copyright regime itself, remaining an evolving area in practice.
Across borders, process documentation is emerging as a key requirement to demonstrate copyright authorship, effectively requiring AI developers and deployers to design record‑keeping into their creative workflows. Documentation is no longer an after‑the‑fact evidentiary exercise, but a governance function that supports attribution, accountability and regulatory auditability across the AI lifecycle.
Lawful text and data mining for AI training
The decree’s second major AI‑related pillar is the introduction of Article 37a, which operationalizes the Law on Intellectual Property’s exception for using protected works in AI‑related text and data mining.
Threshold conditions for training data
Text and data mining is permitted only where the underlying data has been lawfully published, is accessed legally and from a lawful source, and is not obtained by circumventing technological protection measures deployed by rights holders.
Purpose and market‑impact limitations
Even where access is lawful, the TDM activity must be conducted solely for non‑commercial purposes — such as research and experimentation, must avoid prejudice to the normal exploitation of the copyrighted works and must ensure that AI outputs do not replace copyrighted works in the market or create unfair competition.
This framework links copyright compliance not only to how AI models are trained, but also to how their outputs function in downstream markets, an approach consistent with Vietnam’s cautious stance toward large‑scale generative AI deployment.
Opt‑out rights for rights holders
The new Article 37b allows authors and rights holders to reserve their rights against text and data mining for AI training, except where the use fully satisfies the conditions set out in Article 37a.
Opt‑out may be exercised through machine‑readable metadata, technological protection measures or rights‑management information attached to digital works, or through public declarations made via authorized collective management organizations.
For AI developers, this introduces an ongoing obligation to monitor opt‑out signals and reinforces the importance of dataset governance mechanisms that respect dynamic rights reservations.
By comparison, China has not yet codified a comparable opt‑out framework at the legislative level. A similar concept is reflected in the recommended national standard GB/T 45654—2025, which suggests data owners may opt out of AI training through technical measures such as robots.txt. As the standard is non‑binding and has not yet been clearly tested in publicly available judicial practice, its practical implications remain to be seen.
AI developer responsibilities and royalty exposure
The AI implications of the decree culminate in Article 37c, which imposes direct obligations on organizations and individuals conducting TDM for AI systems.
Key responsibilities include retaining technical dossiers and training datasets in accordance with AI‑related laws, being prepared to explain training practices to authorities in disputes or investigations, and respecting opt‑out declarations by rights holders.
Most significantly, even where upstream training was lawful under a non‑commercial exception, downstream commercial exploitation reactivates copyright obligations, including royalty payment. This decoupling of training legality from deployment legality introduces a delayed‑liability model that AI developers must account for early in system design and business planning.
A converging global baseline for AI governance
Viewed together, Vietnam, China and the U.S. seem to be converging on a set of core guardrail principles for AI and copyright. Human creative control remains a non-negotiable foundation, lawful data sourcing functions as a gatekeeping requirement, and commercial exploitation triggers legal and financial exposure. These principles operate across different legal traditions but point toward a common expectation that AI systems cannot be decoupled from human responsibility.
This convergence signals a shift in regulatory focus. The question is no longer whether AI innovation should be regulated, but how tightly it should be integrated with existing legal frameworks such as copyright.
In this context, Vietnam’s approach stands out for translating broadly shared principles into specific compliance rules, offering a more structured preview of where global AI governance may be heading.
