Data on the Internet is becoming an important source of raw materials for training generative artificial intelligence (Gen AI). However, a piece of information that is made public does not mean that any organization has the right to collect, exploit and use it for commercial purposes. When the Law on Personal Data Protection takes effect from January 1, 2026, the problem for AI development businesses in Vietnam is not only how much data can be accessed, but also what legal basis is to use it.
However, the ability to technically access data does not mean the legal right to use the data.
Public data is not masterless data
A photo posted on social media, a professional profile, a comment on a forum, or personal information that appears in an article can all become Gen AI training data.
Thanks to automatic collection technology, businesses can gather millions of content from the Internet, process and put it into large language models or systems that generate images, audio, and video. However, the ability to technically access data does not mean the legal right to use the data.
An individual publishes an image to interact with friends does not mean that they consent to the business using that image to train AI. Similarly, the publicity of a professional profile for job search does not implicitly allow that information to be exploited to develop commercial technology products. Obviously, disclosing information is a behavior; allowing a third party to process information for a new purpose is another legal issue. This is an important but easily overlooked boundary.
For Gen AI, this boundary is even more noticeable because data is not only stored or displayed. They can affect how the model operates, creates content, and responds to users. In some cases, the model also runs the risk of remembering or reproducing personal information from the training dataset.
Therefore, the issue that businesses need to pay attention to is "Is this data found on the Internet? ", but "Do we have the full legal right to use this data to train AI?".
How does Vietnamese law regulate?
In the process of training Gen AI, businesses can perform many data processing activities: collecting, classifying, analyzing, combining, storing, using, sharing, or transferring data to partners. These activities should be considered in accordance with applicable laws.
The bottom line is that publicly available personal data cannot be taken for granted as data that is free to be used for any purpose.
Enterprises must determine the legal basis for the processing, the purpose of use and the corresponding data protection obligations. From January 1, 2026, the processing of personal data in Vietnam is governed by the Law on Personal Data Protection No. 91/2025/QH15 and Decree 356/2025/ND-CP. For AI Gen training activities, enterprises must determine the appropriate basis for data processing, rather than relying on the information being made public. The consent of the data subject and cases of processing without consent must be considered in accordance with the statutory conditions. In particular, the right to access a content, the right to exploit the content under intellectual property law and the basis for processing personal data are different legal issues that cannot be interchanged.
If it relies on consent, that consent must meet legal requirements; if it is based on the case handled without consent, the enterprise must clearly define the applicable conditions. In particular, consent for the publication or disclosure of information is not automatically consent for AI training.
In addition to the Personal Data Protection Law, this activity may also be governed by the 2024 Data Law, the law on cybersecurity, the Civil Code, and the Intellectual Property Law.
Article 38 of the 2015 Civil Code protects the right to private life, personal secrets and family secrets; Article 32 regulates the rights of individuals to images. These provisions are especially noteworthy when the training data contains portrait photos, private life information or content capable of identifying a specific person.
A dataset can also contain personal data and copyright-protected works at the same time. Therefore, meeting the requirements for personal data protection does not mean that the copyright issue has been resolved.
What Compliance Is Required When Collecting Data from the Internet
Web scraping, which uses tools to automatically collect content from websites, is a common method for building AI training datasets. But a website that is publicly accessible to users does not mean that businesses are allowed to copy and exploit all of the content on it.
First of all, there is the boundary of personal data. Businesses need to determine whether the information collected can directly or indirectly identify a person. If so, the processing must meet the corresponding legal requirements, regardless of whether the data is obtained from a public source.
Next is the boundary of intellectual property rights. Articles, images, books, artwork, and many types of online content can be protected. It is not possible to rely solely on the fact that content appears on the Internet to conclude that they are allowed to be used for AI training.
Finally, there are limitations on access. Businesses must consider the conditions of use of the service, access control mechanisms, and system security regulations. Bypassing technical safeguards for data collection may give rise to its own legal risks.
In other words, collecting data to train AI is not just a technical problem. It is an activity that must be managed simultaneously legally, technologically, and ethically.
The big risk lies in sensitive data and re-identification
In fact, not all personal data carries the same level of risk. A person can share their experience of treating a disease on a forum to seek empathy. But if that content is collected, combined with other information, and fed into an AI model, the consequences for privacy may go far beyond the original purpose of sharing. Similarly, financial data, biometrics, location, or other private information requires a higher level of caution.
The risk also comes from the ability to combine data. An anonymous comment may not reveal identity, but when paired with a location, occupation, photograph, and online activity history, the individual involved can be identified.
Therefore, deleting a name or phone number does not mean that the data has been safely anonymized. For Gen AI, businesses need to assess both the risk of re-identification and the ability of the model to remember and disclose personal information through the output content. Data sources that contain sensitive information, children's data, or information that has a significant risk of harm need to be screened and applied.
Instead of collecting as much data as possible, businesses should build a data strategy based on legitimacy, necessity, and control.
Removing data from AI models: The problem is not simple
One of the biggest legal and technical challenges arises when the data subject requests the deletion of information that has been used to train the AI.
With conventional databases, businesses can identify and delete a specific record. But in the AI model, the data is already involved in the process of adjusting the parameters, making it complicated to completely remove their influence.
Deleting data from the repository or training training does not necessarily mean that the model has stopped remembering that information. Therefore, businesses need to build a data management mechanism from the beginning, such as: recording the origin, controlling the training version, receiving data subjects' requests, and evaluating technical measures to limit or eliminate the impact of the data when necessary. It is important not to make a commitment to completely delete data from the model if there is no technical basis to verify.
In addition, the use of data labeling partners, cloud providers, or training infrastructure located abroad may give rise to additional obligations for confidentiality, delimitation of responsibilities, and cross-border transfer of personal data.
Enterprises also need to determine the case in which they must prepare a dossier to assess the impact of processing personal data or assess the impact of cross-border data transfer according to current regulations.
If you want to train Gen AI in accordance with the law, what do businesses need to do?
Instead of collecting as much data as possible, businesses should build a data strategy based on legitimacy, necessity, and control.
Before training, it is necessary to clearly define the data source, the type of information collected, the purpose of use and the corresponding legal basis. Data sources that do not prove the right to use should be excluded or further evaluated.
Businesses should prioritize data that is duly licensed, aggregated, or that has been effectively anonymized. Identifying information that is not necessary for the training objective should be eliminated.
In the process of processing, it is necessary to apply security measures, decentralize access rights, track operations and check the risk of the model disclosure of personal data. When deploying the product, businesses must continue to monitor outputs, receive feedback, handle legitimate requests of data subjects, and update protection measures when new risks arise. More importantly, the responsibility for compliance does not only belong to the legal department. The engineering, data administration, information security, and product development teams need to be involved from the design stage. Because an AI model can work well technically, there are still great risks if the training data source is not transparent.
Public personal data is not a resource that can be freely exploited. With Gen AI, the decisive question is not how much data businesses can collect, but how much they have the right to use it and what risks they can control. Responsible AI development must therefore start from the legitimacy of the input data, not just rely on overcoming the consequences at the output, when everything seems to be the end of the story.
