Who owns AI-generated work?

Who owns AI-generated work? In Malaysia, copyright still needs a human author and sufficient effort. A purely machine-made draft is unlikely to be protected. The UK can treat a computer-generated work as owned by the person who arranged its creation, though the government has now proposed to remove that special protection. The United States will not register a work with no human author. Vendor terms that say "you own the output" do not rewrite those statutes.

A client reads a letter your firm sent last week and asks: if an associate used AI to draft it, do we own it? That is three questions pretending to be one. Can we use it? Can we stop someone else copying it? If a paragraph was lifted from someone else's work, who pays?

This is general information, not advice on a live file. The Malaysian statute is under review.

What question are we actually asking?

What you put in (the prompt, the client's contract, the emails) is already yours or the client's. Pasting it into a tool does not hand copyright to the vendor. It can leak, store, or, if the terms allow, use the text to train. That is contract and the PDPA, not a transfer of title.

The model itself is never yours. You buy a licence to use it.

The output is the messy part. There is a contract answer (what the vendor promised). A copyright answer (whether the statute recognises an author) and an infringement answer (whether the output copied someone else). Those three can point in different directions on the same page.

If the draft was prepared for a client, the engagement letter should say the client gets the work product. That does not give the client better title than the firm has. First ownership, if copyright exists at all, vests in the author.

Does Malaysia copyright AI-generated work?

The Copyright Act 1987 does not mention AI or computer-generated works.

Two gates have to be cleared.

Originality. Section 7(3) says a literary, musical or artistic work is not eligible unless sufficient effort has been expended to make it original in character, and the work has been written down, recorded or otherwise reduced to material form. The Federal Court in YKL Engineering Sdn Bhd v Sungei Kahang Palm Oil Sdn Bhd held that the test is not novelty. What matters is that sufficient effort has been expended in terms of time, labour and skill to produce the work as original.

Authorship. Section 3 defines "author", for a literary work, as "the writer or the maker of the works". For photographs, films and sound recordings it is the person who undertook the arrangements for the making. Every limb of the definition points to a person. Duration for literary, musical and artistic works runs for the life of the author plus fifty years. Moral rights in section 25 are framed around the author's name, honour and reputation, and survive for exercise by the author's personal representative after death. An AI system is not a "qualified person" under section 3 (a Malaysian citizen or permanent resident, or a Malaysian body corporate). Section 10 also lets copyright subsist if the work is first published in Malaysia or made in Malaysia, but those connecting factors still presuppose a work with an author.

So: if a lawyer chose, structured and rewrote the draft, that human contribution is the arguable copyright. If the machine produced the text with almost no human work, protection is uncertain, and the better view is that there may be none.

That is also the government's public line. On 4 August 2026, The Edge reported Deputy Minister Datuk Dr Fuziah Salleh telling the Dewan Negara that the Act protects works that show originality through human effort, skill and intellectual contribution. Fully AI-generated works without significant human creativity do not qualify. She also said Malaysia has not yet recorded a court case on copyright infringement arising from AI.

MyIPO consulted in July–August 2026 on amendments that would clarify when AI-assisted or AI-generated output can be protected. Baker McKenzie's note on the consultation said protection is still expected to hinge on sufficient human authorship under section 7(3) and the statutory concept of an "author" under section 3. That is not yet the amended Act.

Training is a separate fight. The same parliamentary answer, as reported, said using a copyrighted work to train a model needs the owner's permission.[10] The consultation asked whether Malaysia should add a text-and-data-mining exception. Until Parliament amends the Act, treat "permission required" as the government line, not a decided case.

What does UK law do that Malaysia does not?

Section 9(3) of the Copyright, Designs and Patents Act 1988 says that for a literary, dramatic, musical or artistic work which is computer-generated, the author is "the person by whom the arrangements necessary for the creation of the work are undertaken." Section 178 defines "computer-generated" as a work generated by computer in circumstances such that there is no human author of the work. If the work is computer-generated, copyright expires 50 years from the end of the calendar year in which it was made.

That is still the current statute. The UK can say even a no-human output can have an owner. Malaysia and the US usually cannot.

Who made the "arrangements"? In Nova Productions Ltd v Mazooma Games Ltd, Kitchin J held that composite frames generated by a game program were computer-generated artistic works. The programmer who devised the appearance, rules and logic, and wrote the program, was the person who made the arrangements. The player was not an author: playing the game was not artistic skill or labour, and the player had not undertaken the arrangements. That is a 2006 video-game case, not ChatGPT. The UK government's March 2026 report reads the current law as meaning that, for a general-purpose AI which generates output in response to a user prompt, the author "will usually be the person who inputted the prompt." That is a government reading of the statute, not a decided case on generative AI.

The same report then proposes to remove this special computer-generated-works protection, while keeping copyright for works created with AI assistance. Until Parliament amends the CDPA, section 9(3) remains in force.

If a human did author the work and AI was only a tool, ordinary copyright applies to the human parts. Section 9(3) is the fallback where there is no human author at all.

Training is different. On 18 March 2026 the UK government published its report under the Data (Use and Access) Act 2025. A broad data-mining exception with opt-out is no longer its preferred way forward. It said it will not introduce reforms to copyright law until it is confident they meet its objectives.[18] No new training exception has been enacted. Existing copyright continues to apply unless an existing exception applies. The main existing text-and-data exception, section 29A, is limited to copies made for computational analysis for non-commercial research, by a person with lawful access. It is not a commercial-training licence.

Why is the US the strictest on output?

United States copyright needs a human author. A machine cannot be one.

Stephen Thaler tried to register an image, "A Recent Entrance to Paradise", created autonomously by his system, with the machine named as sole author. The D.C. Circuit, on 18 March 2025, held that the Copyright Act of 1976 requires all eligible work to be authored in the first instance by a human being, and affirmed the refusal of registration (Thaler v Perlmutter, No. 23-5233). The Supreme Court denied certiorari on 2 March 2026 (No. 25-449).

The US Copyright Office said the same in its January 2025 report, Copyright and Artificial Intelligence, Part 2: Copyrightability. Purely AI-generated material is not copyrightable. The Office concludes that, given current generally available technology, prompts alone do not provide sufficient human control to make users the authors of the output. What can be protected is the human's own writing, selection, arrangement, or substantial edits. AI-generated portions have to be disclaimed on registration.

The working example is still Zarya of the Dawn. In a letter dated 21 February 2023, the Office held that Kristina Kashtanova is the author of the comic's text and of the selection, coordination and arrangement of written and visual elements. The images generated by Midjourney are not the product of human authorship.

Training in the US is being fought as fair use. The Copyright Office's Part 2 report said a later part would address training. There is no final, general judicial answer. Do not quote a winner.

Malaysia vs UK vs US

Question	Malaysia	United Kingdom	United States
Pure AI output, no real human edit	Probably no copyright. Government line: no significant human creativity, no protection.	Current statute: can belong to the person who arranged creation (CDPA s.9(3)), 50-year term. Government has proposed to remove this special protection.	No copyright. Cannot register.
Human uses AI as a tool, then rewrites	Human contribution can be protected if s.7(3) originality is met.	Human parts: ordinary copyright. Machine-only parts: s.9(3) while it remains in force.	Human contribution can be protected. AI portions must be disclaimed.
Is a prompt enough?	Unlikely, on the current human-effort reading of s.7(3) and YKL.	Government report: for general-purpose AI the prompter is usually the "author" under s.9(3). Untested in a generative-AI case.	Usually no. Copyright Office: prompts alone are not enough control.
Training on copyrighted works	Government: permission required. Consultation on a TDM exception. No decided case.	No new exception enacted (March 2026). Existing copyright applies unless an existing exception (e.g. s.29A, non-commercial) applies.	Fair-use litigation ongoing.
Vendor terms ("you own the output")	A contract with the vendor. Does not create copyright if the Act says there is no author.	Same.	Same.

If the vendor says you own it, is that enough?

No. The terms of use are a promise between you and the vendor. They bind those two parties. They do not amend the Copyright Act 1987, the CDPA, or Title 17 of the US Code. They also do not wash infringement. If the output reproduces a substantial part of someone else's work, "I paid ChatGPT" is not a defence.

Think of a junior who drafts from your notes, using a library the firm did not licence. Your notes stay yours. The library is not yours. As between you and the junior, you can use the letter. Whether the law calls you the author, and whether a paragraph was taken from a textbook, is a different fight.

What should a Malaysian law firm do on a live file?

The safe fact pattern is the same in all three countries: a human selects, structures and rewrites, and can show it.

1. Do not tell a client they "definitely own" an unedited AI draft as copyright. The honest line is: you can usually use it under the vendor's terms; copyright in a fully machine-made work is unsettled here; keep a human in the edit so there is an author to point to.
2. Keep an authorship trail. Who instructed the tool, what was selected, what was rewritten, who signed off. If you ever have to prove sufficient effort under section 7(3) and YKL, that file is the evidence.
3. Treat vendor terms as a licence to use, then read them. Assignment language does not create a statutory copyright that is not there.
4. Check the output for third-party copying. Ownership of the Word document is not clean title.
5. Say in the engagement letter that the client gets the work product as between firm and client. Do not promise a copyright the firm may not have. If copyright exists, section 26 vests it first in the author, then (unless agreed otherwise) in the employer where the work is made in the course of employment.
6. Ownership and liability are different questions. The Bar has not banned generative AI. Circular No. 242/2025, dated 3 July 2025, still puts accuracy, confidentiality and accountability on the lawyer, and reminds members that all GenAI output must be subject to human oversight. If the draft is wrong, the problem is not "the AI".

A first draft from a tool is fine. A final draft that nobody rewrote is how you end up with nothing to own, and still something to answer for.

FAQ

Does ChatGPT own the letter it drafted?

No. The vendor owns the model. You get whatever right the contract gives you to use the output. That is not the same as copyright existing in the text.

If two firms prompt the same thing, can either stop the other copying it?

If there is no copyright, neither has an exclusive right. If each lawyer rewrote the text, each may own their own human contribution.

Does a private "legal AI" tool change the copyright answer?

The confidentiality and training questions change. The copyright question does not. A private deployment does not turn a machine into an author under the 1987 Act.

---

Auxilator helps boutique law firms in Malaysia adopt AI safely, inside the firm's own Microsoft 365 or Google Workspace, grounded in the firm's own precedents, with lawyers reviewing every output and no training on client data. If you are weighing how to bring AI onto a live file, connect or send a message. Happy to compare notes.

Notes

1. Copyright Act 1987 (Act 332), s.26(1)–(2) (copyright conferred by s.10 vests initially in the author; deemed transfer to the employer if made in the course of employment, subject to agreement). Consolidated text: Attorney General's Chambers, as at 7 February 2023. https://lom.agc.gov.my/ilims/upload/portal/akta/outputaktap/1736259_BI/Act%20332_Muktamad%207.2.2023.pdf
2. Copyright Act 1987 (Act 332). No provision equivalent to CDPA s.9(3). The words "artificial intelligence" and "computer-generated" do not appear in the Act.
3. Copyright Act 1987, s.7(3)(a)–(b).
4. YKL Engineering Sdn Bhd v Sungei Kahang Palm Oil Sdn Bhd & Anor [2022] 6 MLJ 1 at [97], [103]–[104], [106] (Federal Court; Rohana Yusuf PCA). The test is not novelty; sufficient effort means time, labour and skill to make the work original.
5. Copyright Act 1987, s.3, definition of "author", paras (a)–(f).
6. Copyright Act 1987, s.17(1).
7. Copyright Act 1987, s.25(2)–(3).
8. Copyright Act 1987, s.3, definition of "qualified person".
9. Copyright Act 1987, s.10(1)–(3).
10. "AI training using copyrighted works requires owners' permission", The Edge Malaysia, 4 August 2026, reporting a Dewan Negara answer by Deputy Domestic Trade and Cost of Living Minister Datuk Dr Fuziah Salleh. https://theedgemalaysia.com/node/813250 (parliamentary answer as reported; not the Hansard PDF).
11. Baker McKenzie, "Malaysia: Public Consultation on Proposed Copyright Act 1987 Reforms", 14 July 2026. https://www.bakermckenzie.com/en/insight/publications/2026/07/malaysia-public-consultation-on-proposed-copyright-act-1987-reforms
12. Copyright, Designs and Patents Act 1988 (UK), s.9(3). https://www.legislation.gov.uk/ukpga/1988/48/section/9
13. CDPA 1988, s.178, definition of "computer-generated". https://www.legislation.gov.uk/ukpga/1988/48/section/178
14. CDPA 1988, s.12(7). https://www.legislation.gov.uk/ukpga/1988/48/section/12
15. Nova Productions Ltd v Mazooma Games Ltd [2006] EWHC 24 (Ch) at [104]–[106] (Kitchin J). High Court, not the later Court of Appeal on infringement.
16. Department for Science, Innovation and Technology and Department for Culture, Media and Sport, Report on Copyright and Artificial Intelligence (presented to Parliament pursuant to s.136 of the Data (Use and Access) Act 2025, published 18 March 2026), Section I, para 1. https://www.gov.uk/government/publications/report-and-impact-assessment-on-copyright-and-artificial-intelligence
17. Same report, Section A, paras 46–47, and Section I conclusions: the government proposes that protection for wholly computer-generated works should be removed, while copyright should continue to protect works created with AI assistance. Not yet enacted.
18. Same report, published 18 March 2026, Section A, paras 18, 176: a broad copyright exception with opt-out is no longer the government's preferred way forward; it will not introduce reforms until it is confident they meet its objectives.
19. CDPA 1988, s.29A (copies for text and data analysis for non-commercial research). https://www.legislation.gov.uk/ukpga/1988/48/section/29A
20. Thaler v Perlmutter, US Court of Appeals for the D.C. Circuit, No. 23-5233, 18 March 2025. https://copyright.gov/ai/docs/court-of-appeals-decision-affirming-refusal-of-registration.pdf
21. Thaler v Perlmutter, Supreme Court of the United States, No. 25-449; certiorari denied 2 March 2026. https://www.scotusblog.com/cases/thaler-v-perlmutter/
22. US Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (January 2025), analysis of prompts: "prompts alone do not provide sufficient human control to make users of an AI system the authors of the output." https://copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf
23. US Copyright Office, letter re Zarya of the Dawn (Registration VAu001480196), 21 February 2023. https://www.copyright.gov/docs/zarya-of-the-dawn.pdf
24. Malaysian Bar, Circular No. 242/2025, 3 July 2025, "Updates on the Use of Generative Artificial Intelligence in Legal Practice" (supplemental to Circular No. 342/2023). https://www.malaysianbar.org.my/cms/upload_files/document/Circular%20No%20242-2025.pdf