Does AI-Generated Video Have Copyright? China, the US and the EU Give Three Different Answers

SinCoSphere EditorialUpdated 2026-09-027 min readLegal & compliance

Three jurisdictions, three answers. China looks at the process, not the tool: prompt design, parameter tuning, iterative selection and post-editing can amount to original intellectual input (Beijing Internet Court 2023, Changshu 2025); a few simple prompts cannot (Suzhou Intermediate Court 2025). The United States requires a human author: prompts alone do not qualify, purely AI-generated output cannot be registered, and registrations must disclose AI-generated material. The EU protects only an author's own intellectual creation and adds AI Act transparency duties from 2 August 2026. For a drama team the work is the same everywhere: keep evidence of human authorship, check each tool's output terms, clear real people's likenesses and voices.

The question is not whether AI video has copyright; it is whether your series does

None of the three jurisdictions gives “AI-generated video” a blanket answer. Each asks the same underlying question: what did a person do in this content?

Chinese courts look for “original intellectual input”. The US Copyright Office looks for a “human author”. EU copyright law looks for the “author’s own intellectual creation”. The wording differs; the effect is the same. Where a human contributed expression, that contribution is protected. Where a human only pressed generate, that segment is not.

For an AI short drama this means copyright is almost certainly layered. The script, storyboard, dialogue and editorial arrangement are human-authored layers and protected everywhere. An individual AI-generated shot depends on the generation record in China, is generally unprotected on its own in the US, and has no settled EU-level case law.

This is not legal advice; the primary texts control, and a specific project needs counsel.

Three jurisdictions side by side

Dimension	China	United States	European Union
Threshold	Original intellectual input by a person (case law)	Human authorship (Office policy, upheld by courts)	Author’s own intellectual creation (directives, CJEU case law)
Prompts only	Simple prompts: no; elaborate prompts plus parameters and iterative selection: yes	Prompts alone are insufficient for authorship	No settled ruling; generally insufficient under the existing test
Purely AI output	Judged on the process, not rejected outright	Not protectable, not registrable	Not protectable
Registration / disclosure	No mandatory registration; voluntary registration must be truthful	Registration must disclose AI-generated material (March 2023 guidance)	No registration system
Key rulings / documents	Beijing Internet Court, 27 Nov 2023; Changshu, Mar 2025; Suzhou Intermediate, Apr 2025	Copyright Office Part 2 report, 29 Jan 2025; Thaler v. Perlmutter	AI Act, Regulation (EU) 2024/1689
Labeling / transparency	Labeling Measures from 1 Sept 2025	No federal labeling mandate (state laws vary)	AI Act Article 50 from 2 Aug 2026
Training data	Civil Code likeness and voice provisions; Copyright Law	Copyright Office Part 3 (pre-publication May 2025); litigation ongoing	DSM Directive Art. 4 TDM opt-out; AI Act Art. 53 copyright policy

Checked 2 September 2026. Summary only; the texts and judgments control.

China: three rulings draw the line

Beijing Internet Court, case (2023) Jing 0491 Min Chu No. 11279, decided 27 November 2023, widely called China’s first AI text-to-image copyright case. The plaintiff generated an image with Stable Diffusion; the defendant used it without permission. The court held that the plaintiff’s prompt design, parameter settings, iterative adjustment and final selection amounted to personalised choices about the image’s elements and composition, showing original intellectual input. The image was a work of fine art and the plaintiff its author. Damages: RMB 500. Neither side appealed.

Changshu People’s Court, reported March 2025, the first such ruling in Jiangsu and the second nationally to recognise copyright in AI-generated content. The plaintiff revised prompts repeatedly and retouched details in Photoshop; the court found this showed distinctive selection and arrangement. Damages: RMB 10,000.

Suzhou Intermediate People’s Court, final in April 2025, the first Chinese ruling that an AI text-to-image output was not a work. Content triggered by simple prompts, without original intellectual input, is not a work under the Copyright Law.

Together these set the Chinese rule: the process matters, not the tool. For a drama team, the generation record is itself the evidence of rights: prompt versions, parameters, reference images, rejected takes, editing decisions. Without a record you are on the “simple prompts” side of the line.

United States: human authorship is the gate, prompts do not pass it, registration requires disclosure

The US Copyright Office published Part 2 of its Copyright and Artificial Intelligence report, on copyrightability, on 29 January 2025. Four conclusions: copyright requires human authorship; prompts alone are insufficient to make the user an author, because the user does not control how the expression is rendered; where a human contributes perceptible expressive elements to an AI-assisted work (creative modification of the output, original selection and arrangement), that contribution is protectable; and decisions are made case by case.

At registration, the Office’s 16 March 2023 guidance requires applicants to disclose AI-generated material and disclaim it. Omitting the disclosure can affect the validity of the registration.

The leading case is Thaler v. Perlmutter. Dr Stephen Thaler sought to register an image generated autonomously by his AI system, naming himself as author by virtue of owning the system; the Office refused. The D.C. Circuit affirmed the refusal in March 2025, confirming the human authorship requirement. As of the check date, several major law firms report that the Supreme Court denied certiorari on 2 March 2026, leaving the appellate ruling in place. We did not obtain the Court’s order list itself; that document controls.

For a drama team distributing in the US: script, storyboard, editing and music arrangement are registrable; the AI-generated footage itself must be excluded on the application.

European Union: no AI exception in copyright, transparency under the AI Act, and an opt-out for training

The EU has not legislated specifically on AI output. Copyright follows the existing test: a work must be the “author’s own intellectual creation”, the threshold the Court of Justice has applied since 2009. Non-human output does not meet it; AI-assisted work is judged on the human contribution. As of the check date there is no EU-level ruling on the copyrightability of AI video.

What bears directly on a drama is Article 50 of the AI Act (Regulation (EU) 2024/1689), applicable from 2 August 2026. Providers of generative systems must mark output in a machine-readable way and enable detection; deployers must disclose deepfakes clearly. For generative systems already on the market before 2 August 2026, the marking duty is deferred to December 2026. Article 50(4) narrows the deepfake disclosure for content that is evidently artistic, creative or satirical to a form that does not hamper the display of the work. In June 2026 the Commission published a voluntary Code of Practice and three optional labeling icons.

On training data, Article 4 of the DSM Directive (2019/790) lets rightholders opt out of text and data mining, and Article 53 of the AI Act requires general-purpose model providers to adopt a copyright policy that honours that opt-out. For your own series this cuts the other way: a machine-readable opt-out is what keeps your public scripts and character sheets out of other people’s training sets.

Tool terms: who owns the output and whether you can sell it

Copyright law decides whether a right exists; the tool’s terms decide whether you may use the output. Summaries of four major video tools’ published terms follow (checked 2 September 2026; paraphrased, not quoted; the current text controls).

Tool	Output ownership	Commercial use	Licence back to the vendor
Kling (global paid-service terms)	Paid members own the IP in generated content	Paid members’ commercial use is “not restricted”, except to build competing products; free-tier watermark and commercial limits per current terms	Licence to Kuaishou including for training
Dreamina (international terms)	User owns inputs and outputs subject to compliance	Terms describe the service as for private, non-commercial use; a paid plan is not automatic commercial clearance	Perpetual, irrevocable, transferable, sub-licensable, survives termination
Runway (terms updated 11 May 2026)	Section 4.4: the company does not claim ownership of inputs or outputs	Not restricted, subject to compliance with the agreement	Non-exclusive, irrevocable, perpetual, sub-licensable, for training and improving models
Hailuo (MiniMax) (hailuoai.video terms)	Paid subscribers retain all IP in generated, downloaded content	Paid plans include commercial use; free output is watermarked	Non-exclusive licence to operate and improve the service; does not affect ownership

Three patterns recur. First, “you own the output” almost always comes with a licence back, so your finished episodes may enter the vendor’s training set. Second, free and paid tiers carry different commercial rights, usually tied to the watermark. Third, terms differ by region; a tool’s mainland-China agreement is not the same document as its international one.

Real faces and real voices: Civil Code Articles 1019 and 1023

Beyond copyright, the risk a drama team most often trips on is personality rights. Article 1019 of China’s Civil Code prohibits infringing a person’s likeness by defacing it or by forging it through information technology, and bars producing, using or publishing a likeness without consent. Article 1023 extends the same protection to a natural person’s voice.

The Beijing Internet Court applied this in the “first AI voice case”, decided 23 April 2024. A voice actor’s recordings had been turned into a text-to-speech product and sold. The court held that voice rights extend to AI-generated voices where the voice is identifiable, and that holding copyright or neighbouring rights in a recording does not authorise using it to train and synthesise a new voice without the person’s consent. Damages: RMB 250,000.

For a drama team: a real face in a reference image or a real voice in a reference clip needs a written licence that expressly covers AI training and synthesis. A licence to use the photograph or recording as a work is not enough. If you build character sheets from an actor’s headshots, the contract must say so. See AI voice for drama and reference audio.

Six things to do before licensing a series

Build the evidence chain: script drafts, storyboard scripts, prompt and parameter logs, reference image provenance, rejected takes, editing decisions, filed per episode. This is the shared evidence for China’s “original intellectual input” and the US “human contribution”.
Inventory rights by layer: script, storyboard, dialogue, edit, music selection. Mark which layers are human-authored and which are AI-generated. License layer by layer.
Archive each tool’s terms: the version in force, the account tier (free or paid), the commercial clause and the licence-back, captured at the time of generation.
Clear likenesses and voices: written consent covering AI synthesis for every real reference image and audio clip; otherwise use preset voices and invented faces.
Map labeling duties: mainland China under the Labeling Measures and the NRTA tiers; the EU under Article 50. See the China compliance timeline.
Put it in the contract: disclose AI share and tools to the licensee accurately, and limit warranties to the layers you can actually warrant.

A fuller list is in the copyright and labeling checklist.

Key figures and sources

Figure	Value	Source	Date
Beijing Internet Court AI image case	(2023) Jing 0491 Min Chu No. 11279, decided 27 Nov 2023; prompt design, parameters and selection showed original intellectual input; image protected as a work of fine art; RMB 500 damages; no appeal	m.thepaper.cn	2026-09-02
Changshu AIGC copyright case	First in Jiangsu, second nationally to recognise copyright in AIGC; revised prompts plus Photoshop retouching showed distinctive selection and arrangement; RMB 10,000 damages	epaper.legaldaily.com.cn	2026-09-02
Suzhou AI image not-a-work case	Final April 2025; output triggered by simple prompts without original intellectual input is not a work	legalinfo.moj.gov.cn	2026-09-02
US Copyright Office Part 2 report	Published 29 Jan 2025; human authorship required; prompts alone insufficient; human expressive contributions protectable	www.copyright.gov	2026-09-02
First AI voice case	Beijing Internet Court, 23 Apr 2024; voice rights extend to AI-generated voice; RMB 250,000 damages	www.chinanews.com.cn	2026-09-02
EU AI Act Article 50	Transparency duties apply from 2 Aug 2026; marking duty for systems already on the market deferred to December 2026	digital-strategy.ec.europa.eu	2026-09-02

FAQ

Is a short drama I made with AI tools my work in China?

It depends on the process, not the tool. Portions where you can show prompt design, parameter tuning, iterative selection and post-editing have been recognised as works; portions produced from a few simple prompts have been refused. Script, storyboard and editing are human layers and unaffected by that dispute.

Is the prompt itself copyrightable?

A prompt is text; a sufficiently long, original prompt can be protected as a literary work. That is separate from the image or video it produces: the US Copyright Office says prompts do not confer authorship of the output, and Chinese courts likewise look at the choices made beyond the prompt.

The tool’s terms say the output belongs to me. Is that the end of it?

No. Terms govern your relationship with the vendor; they cannot create a copyright the law does not grant, and nearly all attach a licence-back and free-tier limits. Check the legal threshold and the contract terms before distribution.

What happens if I do not disclose AI material on a US registration?

The Office’s 2023 guidance requires disclosure and a disclaimer of the AI-generated portion. Failing to disclose can lead to cancellation of the registration or challenges to its validity in litigation. Accurate disclosure is the only safe course.

Did the Supreme Court decide Thaler?

As of 2 September 2026: the D.C. Circuit affirmed the Office’s refusal in March 2025; major law firms report that the Supreme Court denied certiorari on 2 March 2026, declining to hear the case and leaving the appellate ruling in place. The Court issued no merits decision on AI authorship. The Court’s order list is the controlling document.

Do I need a release to build a character sheet from an actor’s photos?

Yes, and it must cover AI synthesis. Civil Code Article 1019 prohibits forging a likeness through information technology; Article 1023 treats voice the same way. The 2024 voice case made clear that owning copyright in a recording is not the same as holding the person’s consent.

What do the EU’s August 2026 transparency rules mean for a drama?

Video produced with generative AI must be machine-detectable as AI-generated, and deepfake content must be disclosed to viewers; for evidently artistic works the disclosure may take a form that does not hamper the display. The Commission’s June 2026 Code of Practice and icons show the expected form.