Who owns what AI helps you create?

Singapore's Ministry of Law and the Intellectual Property Office of Singapore opened a public consultation on artificial intelligence and the country's intellectual property regime on 26 August 2026, and it closes at 5pm on 22 October. One of its three copyright questions is the hardest to answer: when generative AI is used as a tool, how much human contribution should count, and how should a creator show it. Our view is that the law is clear at the edges and quiet in the middle, and that the quiet middle is where most working creators now live. A work generated entirely by AI gets no copyright, because a machine cannot be an author; an AI-assisted work can be protected where a real human contribution shows in the result. That was an edge case until recently. More than half of Singapore's creative professionals had already used generative AI in their work by 2025, so the question now lands on ordinary jobs rather than curiosities.

MinLaw and IPOS launched the consultation on 26 August 2026 at the Global Forum on Intellectual Property, and responses close at 5pm on 22 October 2026. On copyright it asks three things: how AI training can be made more certain and accountable, and whether the existing computational data analysis exception is clear enough; who should carry the risk when AI output infringes an existing work — the developer, the deployer or the end user; and the nature of human creativity in AI-assisted works. The third is the one a working creative should read, because it is about your process as much as the finished file. It asks how human contribution should be understood, recognised and evidenced where generative AI is used as a tool in creative and business workflows. Our reading is that the consultation is really asking creators to be able to describe how they work.

You cannot copyright something a machine made on its own. That is the settled position, and it is not only Singapore's. Minister for Law Edwin Tong, opening the consultation, put the common ground plainly: "Courts and IP offices around the world broadly agree that an author must be a human being. But where do we draw the line? What kinds of acts or decisions made by human prompters demonstrate sufficient creativity?" Singapore answers the easy half. Copyright may subsist in an AI-assisted work, but not in work that is entirely AI-generated, and the law looks for an authorial creation causally connected to the engagement of the human intellect. In our reading, someone who types one prompt and ships whatever comes back is unlikely to clear that bar, while someone who sets the brief, chooses between several options and reworks the result has a better case.

The consultation paper groups the human acts that might count into three stages. Before generation: context-setting, grounding and prompting — deciding what the piece is for and what it must contain. Between generations: selecting and arranging outputs, choosing some and discarding others, and combining them into a larger work. And after: editing that changes the material in a way the creator authored rather than the model. None of the three is a test you pass simply by doing more of it; the paper is candid that borderline cases stay borderline, and it asks how to separate a protectable AI-assisted work from unprotectable AI-generated content. For a studio, the practical reading is unglamorous and we would underline it: keep the record. The stages you went through, the options you rejected and the edits you made are what a claim would rest on.

The consultation is not happening in a quiet market, and the pressure is not hypothetical. In the United States a judge ordered the AI firm Anthropic to pay a large sum to authors whose pirated books it had used to train its models, while finding that training on books the firm had lawfully bought and scanned was fair use — a reminder that the answer can turn on how material was obtained, not only that it was used. The creative industries are counting their own losses, and the American film, broadcasting and publishing trades have shed jobs on a scale that is hard to read as anything but a contraction. We would not dress that up. Richard Florida, who coined the term creative class, told Fortune he did not think AI would kill it: "It'll change the creative class, but in many ways, it'll also liberate the creative class." We are less sure than he is that the change will be gentle.

Singapore's own test case is on a soundstage. Crooks, billed as the country's first AI-hybrid drama, began shooting in June 2026 with Christopher Lee in the lead. The actors perform on live-action sets, and AI is used after the fact to build backgrounds, virtual sets, relighting and spatial expansion — the world around the performance rather than the performance itself. More than 50 crew were reskilled to work that way. The producing studio's line is that the technology is there to support the story, not take it over: as its founder put it, AI cannot replace the tremble in an actor's voice or the silence between two people, and what it can do is give film-makers more room to build the world around those moments. That is also the cleanest illustration of the legal question, because the same scene would be unprotectable if the faces were generated and protectable as performed. Singapore performers have no collective bargaining framework for AI protections, so what a performer can insist on rests on their own contract literacy rather than a union agreement.

The consultation is worth reading, and we would rather it asked its question early than late. Our view is that the three-stage test is the right shape and will be hard to apply, which is normal for copyright: the honest advice to a creative team is not to memorise it but to document the working. Keep the brief, the rejected outputs, the edit history. It costs almost nothing, and it is the only thing that turns "I directed this" from a feeling into evidence.

We are less convinced that the line between AI-assisted and AI-generated work can be drawn cleanly. The three stages blur in practice — a prompt is a set of decisions, and a choice between two outputs is a creative act — and any test that tries to score them will be gamed by whoever writes the workflow down afterwards. We would rather see a low, clear bar that people can meet honestly than a high one that only well-advised firms can prove.

On the film side we are optimistic in a specific, narrow way. What Crooks does — humans perform, machines build the set — is the version of AI in production we would defend, because the thing an audience actually responds to is untouched. The anxiety in the industry is real and the job losses are real, and neither is settled by a consultation. But a regime that protects performed work and refuses to protect generated work is pointing the money and the credit at the right place.