Taiwan's draft copyright guidelines on generative AI: training data, AI output and the AI Basic Act

If you train AI models on Taiwanese content, offer a generative AI service to users in Taiwan, or own content that may end up in someone’s training set, the position to plan around today is what the Taiwan Intellectual Property Office (TIPO) recommended at a public hearing on July 20, 2026: for now, as a rule, obtain a license. The copyright guidelines on generative AI that TIPO discussed at that hearing are still a draft. As of October 19, 2026, no final version has been issued, and the Copyright Act has not been amended for AI.

1. The short answer

Training. TIPO said at the hearing that the first version of the guidelines starts from the Copyright Act as it stands. Because there is no international consensus on whether reproducing other people’s works to train AI is fair use, it recommends, for now, obtaining a license to avoid unnecessary risk. In the same answer TIPO said the guidelines are administrative guidance, not the standard for deciding infringement; whether a particular use infringes is for the courts to decide on the facts of each case.

Output. TIPO’s published interpretation is that content generated entirely by an AI system’s algorithms, with no actual human creative input, is not protected by copyright. Where a person uses AI as a tool and contributes creatively, the result can be protected.

The AI Basic Act. In force since January 14, 2026, it directs the government to support AI development and to review its laws. In our reading, it does not itself create a new copyright exception.

2. Where the guidelines stand

According to the hearing record, the first draft was discussed with experts and government agencies on May 8, 2026, and the public hearing on the “Copyright Guidelines on Generative Artificial Intelligence (first draft)” followed on July 20, 2026. TIPO’s chronicle for July 2026 records about 60 participants, including foreign chambers of commerce, several government agencies, rights-holder groups and the AI industry. The record was posted on TIPO’s copyright bulletin board on August 3, 2026.

At the opening of the hearing TIPO said it expected to issue the guidelines formally this year, and that it would revise them on a rolling basis as international trends and practice develop. It described their core as the copyright questions at the development and training stages of generative AI: whether an AI developer that uses other people’s works in training infringes; whether AI-generated content is protected; and what to watch for when that content is used afterward.

Status check. As of October 19, 2026, TIPO’s copyright bulletin board, general bulletin board, press releases and page of copyright administrative guidance documents show no final version. Nor did we find the draft text on TIPO’s website. This article therefore relies on the hearing record and TIPO’s published interpretations rather than on press reports of the draft.

3. Training a model on other people’s works

What the statute says. Under Article 22(1) of the Copyright Act, authors have the exclusive right to reproduce their works, except as otherwise provided in the Act; Article 3(1)(5) defines reproduction to include direct, indirect, permanent or temporary copying by any means. Article 22(3) excludes transient, incidental temporary copies that are an essential part of a technical process and have no independent economic significance, made solely for lawful network relay transmission or the lawful use of a work (computer programs excepted). Article 65 provides that fair use does not infringe economic rights and that all circumstances are to be considered, with particular attention to four factors: the purpose and nature of the use, including whether it is commercial or for nonprofit educational purposes; the nature of the work; the amount and substantiality of the portion used; and the effect on the work’s current and potential market value.

The Act says nothing specific about AI training. In our reading, whether the copies made in building a training set fall within Article 22(3) or within fair use is a question the statute leaves to case-by-case decision, and nothing in the hearing record suggests that TIPO considers it settled either way.

What was said at the hearing.

• Industry groups and a foreign chamber of commerce asked TIPO not to present a license as the only lawful route and to keep room for fair use under Article 65; to treat data collection, preprocessing, model training, model adjustment and content generation as different stages; to distinguish developers, service providers, deployers and users; and to state that the guidelines are administrative guidance. One group proposed reduced liability for providers that cooperate in taking down infringing content.
• Rights-holder groups supported TIPO’s direction and asked that fair use not be widened for AI; a broadcasting industry group said its members’ content had been crawled for training even though their websites expressly prohibited crawling; several groups called for licensing matchmaking and labeling of AI output.

TIPO answered that widening fair use involves cross-agency discussion and policy decisions that the guidelines cannot settle; that it would study the allocation of responsibility between developers and deployers internally; that it would update the guidelines promptly if clear landmark judgments emerge internationally; and that it was willing to help with a licensing matchmaking mechanism.

The business decision. For a developer or service provider, I would treat licensing as the default for the content that drives the value of your model, and budget for it like any other input. Where you rely on fair use instead, record why, factor by factor; those are the factors a court weighs. I would not treat the final guidelines as a shield either: TIPO itself has said they are not the standard for deciding infringement, so a rights holder can still sue, and the outcome will turn on the facts.

4. Is AI output protected?

TIPO set out its view in e-mail reply 1141017b of October 17, 2025. Under Article 3(1)(2) of the Copyright Act an author is a person who creates a work, and under Article 10 the author enjoys copyright upon completion of the work. TIPO reads these provisions to require a creation by a natural person. Whether an AI-generated work is protected therefore depends on whether there was actual human creative input:

• If AI is used as an auxiliary tool, much like drawing software, and a person contributes actual creative input, the resulting creative work can be protected, and the copyright in principle belongs to the person who actually created it.
• If the content is produced entirely by the AI’s algorithms with no human creative input, it is not protected.

TIPO added that copyright is a private right: if it is disputed whether a work involved actual human creative input, the courts decide on the facts of the case. At the hearing TIPO said the draft explains in principle how to tell human creation from AI generation, by the degree of substantive human participation, and will add detail once there are concrete judgments.

An everyday example. Picture a beverage start-up that produces the artwork for its new can with an image generator and ships it. A competitor copies the design. If the start-up wants to rely on copyright, much will depend on whether it can show its own creative contribution: the sketches, the selection among many outputs, the edits and the layout decisions. TIPO’s reply 1141017b does not say whether a prompt, or a long series of prompts, counts as creative input; in our reading that will turn on the facts. Without that record, the start-up may have paid for artwork it cannot protect.

5. What the AI Basic Act adds

The Artificial Intelligence Basic Act passed its third reading on December 23, 2025, was promulgated on January 14, 2026 and took effect on promulgation (Art. 20). Its central competent authority is the National Science and Technology Council (Art. 2). Three provisions bear on copyright:

• Article 11(1) requires the government, when developing, training, testing and verifying the impact of emerging AI technology, to provide “reasonable use” (the official English translation), support and subsidy measures, and to improve the laws on AI research and application. Where the interpretation and application of the relevant regulations conflict with other laws, promoting new technology and services takes priority, subject to the Act’s basic principles in Article 4.
• Article 13(2) requires the government to strive to improve the quality and quantity of data used for AI, so that training and output reflect diverse cultural values and protect intellectual property rights.
• Article 18 requires the government to review its laws and, where they do not comply with the Act or no law applies, to complete the necessary enactments, amendments or repeals within two years after the Act took effect; until then, where existing regulations are silent, the agency in charge interprets and applies them in line with the Act, in consultation with the central competent authority.

In Chinese, “reasonable use” in Article 11 is the same term the Copyright Act uses for fair use, so it is tempting to read the provision as settling the training question. In our reading, Article 11 directs government policy and does not itself widen fair use: under Article 1, matters the Act does not cover are governed by other laws; the Copyright Act’s latest amendment is still that of June 15, 2022 (checked October 19, 2026); and TIPO said at the hearing that widening fair use is a cross-agency policy matter. TIPO also said that labeling AI output and recording how it was generated lean toward obligations on AI businesses, fall within the scope of the AI Basic Act, and are not purely a copyright matter.

6. What to do now

• AI developers and service providers. License first the content that matters most to your model; keep source and license records for each dataset; document any fair-use position against the Article 65 factors.
• Companies using AI to make content. Record the human contribution to anything you will want to enforce, and settle in contracts with agencies and vendors who owns the output and who bears the risk if it resembles an existing work. For brand assets, consider protection another way as well; our trademark practice page covers filing in Taiwan.
• Rights holders. State your license terms where your content is published, and keep evidence of ownership and publication dates.

How a particular model, dataset or work fares depends on its facts, which is better handled in a consultation than in a general article. When you write to us, please tell us what the model or service does and whether it serves users in Taiwan; where the training data comes from and under what terms; and, for output, the work itself, how it was made (prompts, drafts, edits) and where it has been used. We will update this page when TIPO issues the final guidelines.