← Blog

How much do I have to do for an AI work to be protected?

2026-10-13 · OpenStela

There is no minimum number of hours, prompts or edits. The question both the U.S. Copyright Office and Taiwan's Intellectual Property Office ask is a different one: did a person decide what the expression looks or sounds like? If you only described what you wanted and the model decided the rest, the answer is no. If your own drawing, lyrics, edits or arrangement are visible in the finished work, that part is yours to claim.

So "how much" is the wrong unit. The useful question is "which kind". Below is a ladder of the kinds of contribution, from the bottom up, with what the official sources say about each and where to read it. At the end: what to keep while you work, so that you can show which rung you were on. This is general information, not legal advice.

Effort doesn't count. Control does.

The U.S. Copyright Office published Copyright and Artificial Intelligence, Part 2: Copyrightability on January 29, 2025. Its summary (p. iii) says:

• "Copyright does not extend to purely AI-generated material, or material where there is insufficient human control over the expressive elements."
• "Whether human contributions to AI-generated outputs are sufficient to constitute authorship must be analyzed on a case-by-case basis."

The report is blunt about effort. On revising prompts it says "the time, expense, or effort involved in creating a work by revising prompts is irrelevant, as copyright protects original authorship, not hard work or 'sweat of the brow'" (p. 20).

Taiwan reaches the same place by a shorter route. In 電子郵件1141017b (October 17, 2025), TIPO said that whether an AI-generated work is protected depends on whether there was actual human creative input in the process. AI used as an assisting tool, like drawing software, with real human input: the result can be protected. Content completed entirely by the AI's own computation, with no human creative input: not protected. Neither office publishes a threshold.

The ladder, from the bottom

Rung 0: a prompt, however good

"Based on the functioning of current generally available technology, prompts do not alone provide sufficient control" (USCO, p. iii). The reasoning is on pp. 18–20: a prompt leaves gaps the system fills, and the same prompt can produce very different outputs. Re-running it doesn't change that. "By revising and submitting prompts multiple times, the user is 're-rolling' the dice, causing the system to generate more outputs from which to select, but not altering the degree of control over the process" (p. 20).

Taiwan has said the same about music specifically. Asked about generating songs with Suno, TIPO answered in 電子郵件1140310 (March 10, 2025) that a song completed entirely by the AI's computation, without human input, is not protected.

Rung 1: AI as a brainstorming partner

If you ask a model for ideas and then write or draw the work yourself, the AI never touches the expression. The report: "the user appears to be prompting a generative AI system and referencing, but not incorporating, the output in the development of her own work of authorship. Using AI in this way should not affect the copyrightability of the resulting human-authored work" (p. 12).

Rung 2: your own work as the input

This is the rung most character creators can reach without changing how they work. The report calls these "expressive inputs" (pp. 22–24). If you feed in your own copyrightable work and "that work is perceptible in the output, they will be the author of at least that portion of the output" (p. 24).

The example the Office gives is a registered image called Rose Enigma: a hand-drawn illustration run through an AI system with a prompt. The drawing's outline, the placement of features and the arrangement of leaves were "clearly perceptible in the output". The Office registered it, limited to the human pictorial authorship, with the AI-rendered realism and lighting excluded (p. 23). In a footnote, the report contrasts a refused application, Suryast, where a photograph was restyled after The Starry Night and the input didn't clearly show in the result (p. 23, n. 124).

The lesson for a character designer: your own sketch, turnaround sheet or 3D blockout, if it visibly survives into the output, counts. A photo that gets restyled until it disappears may not.

The same goes for music. TIPO's Suno letter adds that lyrics you write yourself are protected when you finish them, and that no registration is needed (電子郵件1140310). The AI-made audio around them is a separate question.

Rung 3: selecting and arranging

Picking one output out of forty is not, by itself, authorship (see rung 0). Arranging many pieces into a larger work can be. The Office's early decision on the comic Zarya of the Dawn found that "the selection of the images that make up the Work and the placement and arrangement of the images and text on each of the Work's pages" was protectable as a compilation (quoted on p. 25). The Midjourney images themselves were not.

There is a floor here too. "The selection, coordination, and arrangement of only two or three elements is not generally sufficient" (p. 27, n. 138). An album, a comic, a sequenced video or a character bible can clear it. A single image with a caption probably won't.

Rung 4: modifying the output

Editing what the model gave you is the other main route. Copyright "would extend to the material the human author contributed but would not extend to the underlying AI-generated content itself" (p. 25). The report specifically discusses region tools, using Midjourney's Vary Region as the example: "Unlike prompts alone, these tools can enable the user to control the selection and placement of individual creative elements." Whether a given set of edits is enough "will depend on a case-by-case determination" (p. 27).

Think repainting a render, re-drawing hands, cutting a video to your own timing, re-recording a vocal.

Rung 5: AI inside a larger human work

"A film that includes AI-generated special effects or background artwork is copyrightable, even if the AI effects and artwork separately are not" (p. 27). If the work is mostly yours and AI fills a corner, the work as a whole stays protectable. The AI-made part on its own still isn't.

What "protected" covers when you get there

Climbing the ladder doesn't make the AI-made material yours. At every rung the protection covers your part: the perceptible input, the arrangement, the edits. In the US this shows up at registration. The Office's Registration Guidance (effective March 16, 2023) says applicants "have a duty to disclose the inclusion of AI-generated content" and "to provide a brief explanation of the human author's contributions." AI-generated content "that is more than de minimis should be explicitly excluded from the application." Applicants who aren't sure how to fill in the form "may simply provide a general statement that a work contains AI-generated material", and the Office will follow up.

In Taiwan there's no registration step. Copyright arises when the work is finished (Copyright Act Art. 10, cited in the TIPO letters), and a dispute about whether there was enough human input goes to a court that looks at the evidence. In 電子郵件1140522c (May 22, 2025), TIPO added a separate warning. If an AI image is substantially similar to a work in the training data, using it commercially may infringe. It suggested checking the AI provider's usage rules, so read the current terms for your tool and plan.

Keep the evidence while you work

Both systems end up at the same practical question: what did a person do? Evidence of that is easiest to keep on the day you make the work.

1. Keep your inputs. The sketch, the lyric sheet, the reference photo you shot yourself. Rung 2 depends on showing the input and the output side by side.
2. Keep versions, not just the final. The raw generation and the edited result, as separate files. Rung 4 is the difference between them.
3. Write down what you did, in verbs. "Drew the character sheet; generated 30 candidates; repainted face and hands; arranged 12 panels; wrote all dialogue." Leave out adjectives. If the honest answer is "prompted and picked one", write that. An overstated record is the first thing the other side will test.
4. Fix the date. A note in a folder can be edited later. A record with a fingerprint and a public timestamp can't.

Where OpenStela fits

Step 4 is what we built authorship records for. When you register a work, you give the attribution (a pen name is fine), the creation date if you know it, and the rights you hold. Tick This work was made with AI tools and Your contribution becomes required. It takes up to 600 characters, which is enough for a list of verbs. That text is frozen into the record along with the file's SHA-256 fingerprint and the tools you named. The record is anchored to a public chain in that night's batch. After that, nobody can change those fields, including you.

For rung 4, register the raw generation and the edited version as two works. The two record dates show the order. If the works are bound to a character, the character page also lets you mark one as a "Version of" the other, and that link becomes its own record.

What this gets you is a dated, verifiable record: this file existed on this date, under this name, with this statement of what a person contributed. It is evidence if a dispute comes up. It doesn't decide whether you hold copyright, and we don't check or rate the contribution text. It is your statement, and a court or the Copyright Office decides what it's worth. In the US, use it alongside registration, not instead of it.

Registering works is free. Uploads share 300 MB per account, link-type works take almost none, and there's a daily cap on new registrations (see pricing).

More on the contribution field: Can AI-generated art or music be copyrighted?. On video and footage that isn't yours: Is a Higgsfield video copyrighted?. On what a record proves: What an ID proves.

This is general information, not legal advice. If a real dispute comes up, talk to a lawyer where you live.