Intellectual Property in AI-Generated Outputs under Thai Law
Thai businesses are already using AI to create commercial materials, already signing contracts that assume rights in those outputs can be transferred, and content is already being collected to train models. The Copyright Act B.E. 2537 (1994), Patent Act B.E. 2522 (1979), Trade Secrets Act B.E. 2545 (2002) and Trademark Act B.E. 2534 (1991) all predate generative AI and contain no provisions addressing it. The gap is being filled, for now, by administrative practice, contract drafting and emerging judicial guidance.
Three questions are usually combined but must be answered separately: whether an AI-generated output can be protected and who owns it; whether an output infringes the works the model was trained on; and whether the training itself is lawful. The answer to one does not determine the answer to another.
One distinction organizes much of the analysis. Copyright and patents depend on who created the thing, so without a qualifying human act there is no right. Trade secrets and trademarks do not, because they turn on what the subject matter is and how the owner treats it. AI removes the human from the act of creation but does not change the subject matter, so it troubles the first group and barely touches the second.
Copyright and patents: the human requirement
Thai commentary often says the Copyright Act requires a human creator. The wording supports this less than people assume and the structure of the Act supports it more.
Section 4 defines a creator as the person who makes, or who causes to come into existence, any creative work. It does not say “natural person” or “human,” and nothing in section 4 itself rules out a machine. The requirement comes from how the rest of the Act is built: section 8 ties protection to the creator’s nationality or residence, sections 19 to 22 count the term from the creator’s death, and section 18 gives moral rights. Each assumes someone who can hold a nationality, die, or have a reputation. The conclusion that an AI system cannot be a creator is correct, but it is drawn from the structure of the Act rather than stated in it.
Does writing the prompt make you the creator? Not automatically, and the question is open. The argument in favour rests on the words “causes to come into existence.” Against it, section 10 shows that Thai law already separates causing a work to exist, by commissioning it, from creating it. Section 6 paragraph 2 excludes ideas and methods, and a prompt is normally an instruction, so a longer prompt is simply a longer idea and Thai law asks for the creator’s own skill, effort and judgment in the expression, which with a generative model comes from the model.
Where a person genuinely selects, arranges or edits generated material, protection follows through the ordinary rules on adaptations and compilations, but it covers only what the human added, not the AI-generated material underneath.
Patents raise the same requirement with a sharper consequence. Section 10 of the Patent Act gives the inventor the right to apply and to be identified as inventor, and an AI system can do neither. Section 54 provides that a patent is invalid if granted without compliance with section 10, so naming an AI system as inventor is not a procedural error correctable during examination, it may leave a vulnerability a third party can later raise as a ground of invalidity. For most AI-related applications, however, inventorship is not the principal obstacle: section 9(3) excludes computer programs and section 9(2) excludes scientific and mathematical rules, so the real question is whether the subject matter can be framed as a technical process or apparatus providing a concrete solution to a technical problem.
The development that actually matters
In recent months the Department of Intellectual Property (“DIP”) changed its copyright notification form. It now asks whether the work was made using generative AI and, if so, requires the applicant to show how AI was used and to what extent.
Two points follow. The form does not ask a yes-or-no question that disqualifies the applicant; it asks how much, and it asks for proof. The DIP is therefore treating this as a matter of degree rather than of which technology was used, adopting case-by-case assessment through the design of a form rather than through guidance or legislation. And the operative test is whether AI functioned only as a tool: did it help turn an idea the applicant already had into a finished work, or did it supply the creative content itself? A German court reached the same test independently, holding that an AI model must be closer to a mere tool (Hilfsmittel) than to something that creates on its own. The evidence must therefore exist now — prompt histories, sketches made before using AI, and an account of which decisions were the applicant’s.
However, refusal does not mean that copyright does not exist, as copyright arises upon creation and the notification merely records it. Nevertheless, a refused applicant loses the evidential benefit that the system is intended to provide. The amendment also leaves a gap: the form asks about the degree of AI involvement without published criteria for what level of human contribution suffices.
Training data: a problem of statutory wording
Under section 15 of the Copyright Act, building a training dataset out of protected works is copying, and the question is whether an exception applies. Section 32 paragraph 2 contains a closed list of permitted uses, but it is a closed list of purposes that assume a human is reading or looking at the work: research or study “of the work,” criticism, comment, news reporting and teaching. None describes a machine that reads nothing and simply processes form. Thai law does not prohibit mining; its wording does not reach it. That leaves paragraph 1, which asks about the effect on the market and contains no test of purpose. Thailand therefore cannot simply adopt Japan’s approach, which permits mining where the user is not enjoying the work’s expression, because that is a test of purpose and would have to be added to paragraph 2.
In Civil Red Case No. Tor Por 118/2563, the Central IP & IT Court held that feeding theses into a plagiarism-detection system and storing them permanently was copying, but not infringement: the purpose was to compare texts for similarity, which differed from how the owner would normally exploit the work and had no effect on the market. The court applied section 32 paragraph 1 on its own. This is the closest Thailand has to a text-and-data-mining rule, but it does not extend to generative training, because a model trained to produce text, images or music competes directly with what it was trained on.Scraping is not mainly a copyright question
Much of the material collected may not qualify for copyright at all: section 6 excludes ideas, procedures and systems, and section 7 excludes daily news and facts that are merely information, so prices, statistics and basic factual reports may fall outside protection. A compilation may still qualify, but Thailand has no sui generis database right to cover the gap.
The Computer Crime Act B.E. 2550 provides another layer, operating independently of copyright. Sections 5 and 7 address unauthorized access to a computer system or data protected by specific access-prevention measures, so where scraping bypasses such measures the critical question is not whether the material is copyrighted but whether an access control was circumvented. Application to particular techniques remains fact-sensitive, and Thai judicial authority appears limited. The PDPA may also apply where scraping captures personal data; public accessibility does not necessarily take the data outside the Act, and the analysis may extend to overseas operators. Robots.txt tells automated crawlers which parts of a site they may access. In the European Union, text and data mining is permitted subject to conditions, one of which is an opt-out: the permission applies only where the rights holder has not reserved its rights in machine-readable form. Thailand has no equivalent regime, so a reservation has no independent legal force here. It is still worth publishing one. Section 32 paragraph 1 asks whether a use unreasonably prejudices the owner’s legitimate rights, and a clear, dated, machine-readable reservation is evidence of what those rights were understood to be. Two things that make Thailand different Thai is a relatively low-resource language, and much of its high-value online content sits with a limited number of publishers, broadcasters, forums and public-sector websites. This leaves individual rights holders particularly exposed, but it may also make collective licensing more practicable, because the valuable content is held by a relatively identifiable group of owners. Thailand’s copyright industries reportedly account for 8.01% of GDP and more than 900,000 jobs. The second difficulty is transparency. Thai law does not require AI developers to disclose the sources of their training data, so a copyright owner may have no effective means of knowing whether its work has been included in a dataset. A rule may prohibit unauthorised copying in principle, but its value is limited if rights holders cannot discover that it has occurred.