Ason Zhang and Ge Wang discuss copyright determination for AI-generated works, highlighting the practical value of creation-trail records and token quantification, viewed through recent judicial precedents in China.

Emerging judicial trends in generative AI copyright disputes

Technology has made the determination of copyright in AI-generated content (AIGC) one of the most closely watched frontier issues in IP law in recent years. Since 2024, Chinese courts have heard a series of copyright disputes involving AI text-to-image works and AI designs, and the underlying judicial reasoning has visibly shifted: from an initial focus on “whether the final generated result satisfies the elements of a work” toward a substantive examination of “the degree of human contribution throughout the entire creative process”.

The disputed work - Spring Breeze Brings Gentleness

The case “Spring Breeze Brings Gentleness” [Beijing Internet Court (2023) Jing 0491 Min Chu No 11279 Civil Judgment Li v Liu, Copyright Infringement Dispute, 2023-11-27] heard by the Beijing Internet Court, is the starting point in this area. In that case, the plaintiff completed an AI image creation through prompt design, parameter tuning, and multiple rounds of iteration; the court held that the plaintiff’s operations reflected intellectual investment and personalised choices, constituting a work protected under copyright law, with copyright vesting in the plaintiff. The “human originality/intellectual investment” standard established by this case has provided a basic point of reference for the adjudication of similar cases since.

But not all AI-generated content can obtain copyright protection. The “Phantom Wing Transparent Art Chair” case [Zhangjiagang People’s Court (2024) Su 0582 Min Chu No 9015 Civil Judgment], heard by the Zhangjiagang People’s Court, offers a contrasting reference point: the plaintiff claimed that an AI-generated design drawing for an art chair constituted a work, but because he could not provide original creative materials such as prompts or parameter-iteration records, the court held this insufficient to prove the plaintiff had made a creative contribution to the image, and ultimately dismissed the claim.

The plaintiff appealed to the Suzhou Intermediate People’s Court but failed to pay the appeal fee by the deadline. Therefore, the court issued a ruling treating the appeal as withdrawn, and the first-instance judgment became final [Suzhou Intermediate People’s Court, Jiangsu Province (2025) Su 05 Min Zhong No 4840 Civil Verdict Feng v Zhu et al, Copyright Infringement and Unfair Competition Dispute, 2025-04].

The plaintiff’s AI text-to-image work – Phantom Wing Transparent Art Chair

The chair manufactured by the defendant

The AI-designed plush toy case concluded by the Yangzhou Intermediate Court in 2026 [Yangzhou Intermediate People’s Court, Jiangsu Province. Second-Instance Civil Judgment, Wan v a Yangzhou Company, Copyright Ownership and Infringement Dispute, 2026] further clarified the boundary: the plaintiff had merely entered common descriptive keywords such as “cartoon snake”, “plush texture” and “3D” obtaining the image after multiple refresh-clicks; the court held that such conduct “could not exert strong constraint over the software’s output”, so the generated content could not reflect personalised choice or original contribution, and did not constitute a work.

Together, these cases sketch out the basic framework of current judicial adjudication: AI is merely a creative tool and lacks legal subject status; the key to determining originality lies in the degree of human control over the generative process; and the original record of the creative process is the core item of evidence, whose absence may directly lead to a losing outcome. In other words, “process review” is becoming the dominant method for adjudicating AI copyright cases.

The disputed work – cartoon snake toy

Notably, in August 2026 the Kaifu District People’s Court in Changsha, in the case of Ma v a Changsha kindergarten for infringement of the right of communication through information networks [Kaifu District People’s Court, Changsha, Hunan Province (2026) Xiang 0105 Min Chu No 8031 Civil Judgment. Ma v a Changsha Kindergarten, Infringement of the Right of Communication Through Information Networks, 2026-08 (the first-instance judgment had taken effect after the appeal period)], further refined “process review” into an operationalised, layered analytical approach.

The court divided the AI generation process into three layers – front-end conception, generation control, and back-end processing – examining the originality contribution at each layer separately to avoid conflating the assessments. This more refined method of review confirms the trend noted above – judicial practice is moving from a general inquiry into “whether there was human contribution” toward a more granular breakdown and verification of the entire creative process.

Creation-trail records and token quantification: the value and limits of two new analytical tools

Once “process review” became the consensus approach, a practical question followed: how does one prove the human contribution made during the creative process? Under the traditional adjudicative model, originality determinations rely heavily on the subjective judgment of the judge, and there is no objective yardstick for what counts as a “substantive contribution” or how to distinguish “tool use” from “creative direction”. Two tools that have emerged in practice in recent years offer a new way to address this problem.

The first is the creation-trail evidentiary system. As compliance systems on AI platforms have matured, mainstream generative platforms now retain complete records of user operations, including prompt version history, parameter-adjustment logs, the iterative generation process, records of selection and deletion, seed-value modifications, and other end-to-end metadata. These records can fully reconstruct the entire creative process and prove whether the user genuinely engaged in sustained, intellectually substantive conduct. In the “Phantom Wing Transparent Art Chair” case discussed above, the plaintiff lost precisely because he lacked this kind of original record. The court specifically noted that after-the-fact simulated operations cannot substitute for the original generative process, since the hardware/software environment, input prompts, and operational steps all lack identity and comparability with what actually occurred.

The second is the token-investment quantification mechanism. The token is the basic unit by which large models process information; a human’s prompt input, multiple rounds of iteration, and manual corrections during the creative process all consume a certain number of tokens. By calculating the proportion of human-attributable tokens out of the total tokens consumed across the entire generative chain, one can objectively reflect, from a data perspective, the volume, complexity, and depth of iteration in human informational input. Compared with purely subjective aesthetic judgment, the token-share ratio offers a standardisable, horizontally comparable quantitative indicator.

But neither tool can be directly equated with a finding of originality. Creation-trail records can only prove “what operations the user performed,” not that those operations were creative – a single “one-click generation” also leaves an operational record, yet plainly does not constitute creation. Token count, meanwhile, represents only the volume of investment, not the quality of the idea – a long, verbose, simple prompt is not necessarily superior to one precise, creative instruction. Both, therefore, are corroborating evidence rather than substitute standards for establishing rights.

A more sound approach is to combine the two tools: first use creation-trail records to determine whether the human genuinely directed the creative process and whether there was sustained intellectual investment, ruling out low-investment scenarios such as “simple refresh-and-pick”; then use the token-investment ratio to help assess the depth and volume of the human contribution, providing objective data support for the contribution-level assessment; and finally, the judge makes the substantive determination based on the originality of the expression in the final work.

Moving from “was there any operation at all” to “how much was invested” and then to “does it count as creation” forms a progressively layered adjudicative logic. This sequence mirrors the three-layer analysis applied by the Changsha Kaifu Court in the kindergarten case discussed in the first section, which examined front-end conception, generation control, and back-end processing separately: both refine, within the “process review” framework, the method for assessing the degree of human contribution.

Improving the rules

Alongside the rapid development of judicial practice, progress is also being made at the level of formal rules. In November 2025, the national standard GB/T 45654-2025, “Cybersecurity Technology – Basic Security Requirements for Generative Artificial Intelligence Services” [State Administration for Market Regulation, Standardization Administration of China. GB/T 45654-2025, “Cybersecurity Technology – Basic Security Requirements for Generative Artificial Intelligence Services,” effective 2025-11] formally took effect, laying out systematic provisions on corpus compliance, content labeling and complaint-and-reporting mechanisms, thereby strengthening copyright-compliance obligations at the input stage. The State Council’s 2026 legislative work plan proposes accelerating comprehensive legislation for the healthy development of artificial intelligence, with related research and drafting work now underway, which is expected to provide a more systematic legal response to AI copyright issues in the future.

On 7 September 2026, the China Supreme People’s Court, after in-depth research and extensive solicitation of opinions, formulated the “Supreme People’s Court’s Opinions on the Legal Trial of Artificial Intelligence-Related Disputes” [China Supreme People’s Court. Supreme People’s Court’s Opinions on the Legal Trial of Artificial Intelligence-Related Disputes, 2026-09-07] to address prominent issues in the adjudication of AI-related cases.

The “Opinions” establish three fundamental principles: upholding a people-centred approach, supporting innovation-driven development and reinforcing safety safeguards.

Most relevant here is Article 12, in the IP section, which allocates liability among developers, service providers and users when AI-generated content infringes copyright: it lists the factors courts should weigh, requires developers who deny infringement to produce the sources of their training data and records of the training process, and holds users liable where, knowing or having reason to know of an earlier work, they generate substantially similar content without a valid defence. The “Opinions” do not, however, address whether AI-generated content can be protected by copyright, which leaves that question to the courts. These measures hold significant practical importance for standardizing legal application and adjudication in AI-related disputes, safeguarding public rights and interests, and fostering the healthy and orderly development of the AI industry.

Globally, the regulatory approach to AI copyright shows a diverse pattern. The United States has advanced incrementally through judicial precedent and Copyright Office administrative guidance: the series of policy documents issued by the Copyright Office in 2025 [United States Copyright Office. Copyright and Artificial Intelligence, Part 2: Copyrightability, 2025-01] clearly distinguished between “AI-assisted content” and “AI-generated content”, and in March 2026 the US Supreme Court denied certiorari in Thaler v Perlmutter, leaving in place the DC Circuit’s ruling that an AI-generated work with no human author involved cannot obtain copyright protection.

The EU, by contrast, has taken a different approach: the Artificial Intelligence Act passed in 2024 [Regulation (EU) 2024/1689 (Artificial Intelligence Act). Official Journal of the European Union, 2024] and the General-Purpose AI Code of Practice issued in 2025 have built a strict regulatory framework covering training-data transparency and copyright-holder opt-out mechanisms, among other measures. However, neither the AI Act nor the Code of Practice addresses whether AI outputs can be protected, and the EU has not legislated on that question yet. Despite their differing paths, all three jurisdictions are highly aligned on the fundamental position that “human creation is a prerequisite for copyright protection”.

Practical recommendations

For creators using AI tools, there are several practical recommendations worth noting for effectively protecting one’s own copyright under the current legal framework: first, keep a complete record of the creative process, including prompt versions, parameter settings, generation logs, intermediate drafts and final revision records, and where possible fix this evidence using methods such as blockchain notarisation or timestamping; second, increase substantive human creative investment, avoiding “one-click generation” style use, and instead making full, personalised modifications, selections and arrangements of the AI’s output; third, properly retain account records, operation logs, and payment receipts from the AI platform, as these may all serve as important evidentiary material.

Overall, the copyright dispute over AI-generated works is, at its core, a friction between traditional categorical legal rules and a new, digitised mode of creation. While creation-trail records and token quantification cannot substitute for the substantive determination of originality, they can effectively remedy the shortcomings of the traditional rights-determination approach – its high subjectivity, difficulty of proof and blurred boundaries – providing a more traceable, quantifiable and unified practical reference path for judicial adjudication. As the AI industry continues to develop and judicial experience continues to accumulate, a more systematic, scientific, and balanced AIGC copyright rule system is something worth looking forward to.

Ason Zhang and Ge Wang are members of Chispo Attorneys at Law in China. Ason Zhang is also a member of the MARQUES Copyright Team