Key takeaway: Under the June 2023 interpretive letters of the Taiwan Intellectual Property Office (TIPO), Ministry of Economic Affairs, using others' copyright-protected works to train an AI model involves "reproduction" and constitutes infringement unless it qualifies as fair use under the Copyright Act; content generated automatically by AI is not protected by copyright if no original human expression is involved. Internationally, U.S. courts began issuing rulings on AI training and fair use in 2025, and their views remain split — the Thomson Reuters case rejected fair use, while the Anthropic case held that training on lawfully acquired books is fair use but that building a library from pirated copies is not, with Anthropic reaching a USD 1.5 billion settlement in September 2025. Businesses should manage risk on three fronts: training data (input), AI output (output), and contracts (supply chain).
1. Taiwan's Current Rules: A Basic Framework Resting on Interpretive Letters
1. Training stage: using others' works to train AI is "reproduction"
In June 2023, the Taiwan Intellectual Property Office (TIPO) of the Ministry of Economic Affairs stated its basic position in interpretive letters: using others' copyright-protected works to train an AI model involves "reproduction" within the meaning of the Copyright Act and, unless it qualifies as fair use under the Copyright Act, constitutes infringement. The practical significance of this sentence: training data is not free to use just because it "can be found online" — public availability and licensed use are two different things, and businesses collecting training data must go back and verify the legality of their sources.
2. Output stage: no human creative input, no copyright
The other key point of the same interpretation: content generated automatically by AI is not protected by copyright if no original human expression is involved; conversely, where a human makes a substantial creative contribution, the human-created portion can be protected. Two reminders for businesses follow. First, purely AI-generated output may leave a company unable to assert copyright against competitors — for important brand assets, copy and designs, keep substantial human involvement and preserve records of the creative process. Second, AI output may itself be so similar to others' works as to raise infringement concerns, so it still needs to be checked before use.
3. The legislative gap: the Control Yuan flags "insufficient legal clarity"
A Control Yuan investigation report of April 2025 pointed out that the copyright issues of generative AI currently rest only on TIPO's interpretive letters, and that this lack of legal clarity makes it difficult to keep pace with generative AI's rapid development. In other words, the current rules are built mainly on administrative interpretation; future amendments or policy adjustments remain possible, and corporate compliance programs need to retain the flexibility to adapt.
2. Timeline of Three Key U.S. Cases: The Battle over Fair Use
The United States is the main battleground for generative AI copyright litigation. Substantive rulings began to appear in 2025, but the courts' views are visibly split:
| Date | Case | Court's finding |
|---|---|---|
| February 2025 | Thomson Reuters v. Ross Intelligence (U.S. District Court for the District of Delaware) | Infringement found; the defendant's fair use defense was rejected. |
| June 2025 | Bartz v. Anthropic (U.S. District Court for the Northern District of California) | Training AI on lawfully acquired books held to be "transformative" fair use; but downloading more than 7 million books from pirate sites to build a library was outside the scope of fair use. On September 5, 2025, Anthropic reached a USD 1.5 billion settlement with the author class plaintiffs (with the settlement terms subject to subsequent court review). |
| June 2025 | Kadrey v. Meta (decided in the same period) | On training and fair use, the court's reasoning differed from the Anthropic decision, showing that U.S. courts remain divided. |
Meanwhile, major cases such as The New York Times v. OpenAI are still pending, and U.S. courts have yet to converge on whether AI training constitutes fair use. The clearer signal so far: the legality of how the data was acquired is the key variable — training a model on lawfully purchased books and on pirated downloads fare very differently in court; and the size of the settlement shows that training data disputes can amount to a major financial risk for AI companies.
3. A Risk Map for Taiwanese Businesses: Input, Output and Contracts
Input: training and fine-tuning data
When a company trains or fine-tunes its own models, using protected works without authorization carries reproduction-infringement risk; even where training is outsourced, who supplied the data and whether its sources are lawful will affect where liability falls. Source legality is the first line of defense in the entire risk chain.
Output: the two layers of risk in AI output
The first is a rights vacuum: purely AI-generated output is not protected by copyright, so content a company invests resources to produce may be impossible to protect from use by others. The second is infringement risk: if output is substantially similar to others' works, using it externally can still constitute infringement — "the AI generated it" is not a defense.
Contracts: the supply chain brings foreign risk to your door
U.S. judgments do not bind Taiwanese courts, but if a Taiwanese company uses a foreign provider's AI services or sells its products and services into the U.S. market, the provider's infringement disputes can flow through to it via contract clauses, service interruptions or customer claims; and when delivering AI output to customers, a company may itself be asked to warrant the rights in that output. The flames of international litigation tend to spread along contracts and supply chains.
4. Five Concrete Countermeasures for Businesses
First, audit and license your training data sources: build an inventory of training and fine-tuning data, distinguishing owned, licensed, openly licensed and unknown-source data — do not use unknown sources, and obtain the licenses you lack. Second, rights-check AI output before use: before publication or commercial use, review the output for similarity risk against existing works, and keep substantial human creative involvement in important assets. Third, copyright warranty and indemnification clauses in vendor contracts: require AI service providers to warrant the legality of their training data and to indemnify for output infringement, and check the scope and caps of those clauses. Fourth, internal generative AI use policies: define which tools may be used, what data may be entered (trade secrets and personal data above all), and the review workflow and division of responsibility for output. Fifth, labeling and records: keep records of AI use, of human creative contributions and of data sources — whether you later face a dispute or need to assert rights, records are the strongest evidence.
5. How LegalAI Can Help
LegalAI Law Firm is itself a law firm that uses AI deeply in its practice — we understand the technical reality as well as the boundaries of rights. We provide contract review services and intellectual property services: reviewing the warranty and indemnification clauses in AI vendor contracts, building internal generative AI use policies, and advising on the copyright risks of training data and AI output — so your business can bring AI into its operations with confidence.
Frequently Asked Questions
Does training an in-house AI on online articles infringe copyright?
There is a risk of infringement. Under the June 2023 interpretive letters of the Taiwan Intellectual Property Office (TIPO), Ministry of Economic Affairs, using others' copyright-protected works to train an AI model involves 'reproduction' and, unless it qualifies as fair use under the Copyright Act, constitutes infringement. Articles publicly available online are in principle still protected by copyright — 'findable' does not mean 'usable.' Businesses should first verify the legality of their data sources and obtain licenses where necessary.
Is AI-generated text or imagery protected by copyright?
Under TIPO's interpretive position, content generated automatically by AI is not protected by copyright if no original human expression is involved; where a human makes a substantial creative contribution in the process, the human-created portion can be protected. Before putting AI output to commercial use, businesses should confirm not only whether they can claim rights in it, but also check whether the output risks infringing others' copyright.
Do U.S. rulings affect Taiwanese companies?
U.S. court decisions are not binding on Taiwanese courts, but they carry reference value. The more practical impact comes through supply chains and contracts — if a Taiwanese company uses a U.S. provider's AI services, or sells products and services into the U.S. market, it may indirectly bear U.S. litigation risk through the warranty and indemnification clauses in its contracts, so it should keep following developments in the U.S. cases.
What clauses should businesses look for when contracting for generative AI?
When contracting with an AI service provider, pay attention to representations and warranties on the legality of training data, copyright warranty and indemnification clauses, the ownership of AI output and its permitted scope of use, and the provider's duty to cooperate if a dispute arises. Pair these with internal generative AI use policies and usage records to form a complete risk-control framework.
References
- Taiwan Intellectual Property Office (TIPO), Ministry of Economic Affairs: International trends in generative AI and copyright (Intellectual Property Right Journal)
- Lee, Tsai and Partners: TIPO issues interpretive letters on the copyright issues of generative AI
- Kotto Dori Law Office (2025): U.S. AI copyright rulings begin to move
This article was compiled by our lawyers with the assistance of LegalAI's AI legal assistants, based on the public sources listed above. It is for general reference only and does not constitute legal advice on any specific matter. For the application of specific obligations, please rely on the competent authorities' announcements and the official text of the laws, or consult our lawyers.