Across 2025 and 2026, AI copyright case law and regulations landed in dense succession — dense enough that the phrase “gray area” has stopped working: the training side, the generation side, and the platform side all tightened at once (Phase 2, 04). For people who write for a living, or half a living, what matters is not memorizing every ruling but knowing which ones will land on you. This post lays out the whole panorama — the training-data fight, how generated works are judged for copyright, platform disclosure obligations, three hard lines in the Chinese context — then compresses it into four things individual creators should do, closing with how this blog itself handles it.

Note: markers like (Phase 2, 04) refer to numbering in internal research reports (Phase 2 is the first-hand corpus and case-law group report); the reports keep clickable source links for every case. Anything the reports marked “unverified” is not cited here.

Training Data: Two Jurisdictions, Two Fates

The American courts’ position can be summarized in one sentence: training is most likely lawful, but the source of the data will be prosecuted. In Bartz v. Anthropic (June 23, 2025), the judge ruled that Anthropic’s use of copyrighted books during training was fair use — the first case where an AI company won fair use on the training side; but the same ruling held that downloading and permanently storing books from the piracy library LibGen constituted infringement (Phase 2, 04). Forty-eight hours later, Kadrey v. Meta likewise dismissed the authors’ class action, because the authors could not produce evidence of “substantial similarity” between model outputs and their works (Phase 2, 04). The implicit rule the two rulings piece together: how you use it is a technical question, where you got it is a legal question. In September of the same year, Anthropic announced a $1.5 billion settlement of the authors’ class action, formally reached in July 2026 — paying precisely for the pirated-storage segment (Phase 2, 04). The case also incidentally exposed a rights-chain problem in class actions: once payouts entered the distribution stage, some publishers asserted copyright claims against the settlement, and the amounts authors would actually receive became contested.

The EU’s position is the opposite: unlicensed training itself can be found infringing. GEMA v. OpenAI (November 11, 2025, Munich) became Europe’s first winning copyright-infringement judgment on the training stage; the court prohibited OpenAI from reproducing the 9 song lyrics at issue in its model, prohibited offering them publicly through ChatGPT, and ordered disclosure of training-data usage information (Phase 2, 04). The accompanying EU AI Act entered full enforcement on August 2, 2026: general-purpose models must submit training-data summaries and copyright compliance documentation, with violations fined up to 7% of global annual revenue (Phase 2, 04). The EU’s play is using transparency to force licensing — only once training data is laid open can rights holders meaningfully refuse or charge.

The two jurisdictions signal different things to creators. In the US, if your work is taken for training, you most likely cannot win — unless you can prove the model output is substantially similar to your work; in the EU, transparency obligations are handing rights holders negotiating leverage. The American CLEAR Act draft bill would mandate disclosure when training data contains copyrighted works, showing the legislature also wants to converge toward Europe — but it remains a draft (Phase 2, 04).

Generated Works: Human Contribution Is the Only Hard Currency

The state of adjudication in one sentence: purely AI-generated output has no copyright; human selection, modification, and iteration create rights. The US Copyright Office’s Part 2 report from January 2025 puts it plainly: outputs without human originality investment enjoy no copyright; where a human’s choices, modifications, and iterations of prompts embody personalized expression, rights are recognized in proportion to that contribution (Phase 2, 04).

Chinese case law sketches an even finer spectrum. The Beijing Internet Court’s 2023 first AI text-to-image case held that when prompt and parameter design embody originality, the user is the author; in March 2025 the Changshu Court offered an operational scale — where prompt modifications plus image-editing-software modifications of details embody unique selection and arrangement, copyright applies; in November 2025 the Shanghai Huangpu Court filled in the other end: simple list-style prompts are merely “a list of instructions on style, subject, and material”, lacking expressive originality, and do not constitute a work (Phase 2, 04). Read together, the three cases draw a clear scale: the more specific the modification and the more it embodies personal choice, the closer the rights come to you; one-prompt generation gets rights close to zero.

Don’t mistake local precedents for national rules. On September 7, 2026, the Supreme People’s Court issued the Opinions on Adjudicating AI-Related Disputes in Accordance with the Law — 5 parts, 24 articles — elevating the case experience of Beijing, Guangzhou, Hangzhou, and other courts into a nationally unified adjudication standard, but explicitly “deferring for now” on two questions: the copyrightability of AIGC and the characterization of training large models on others’ works without permission; the exact words were “opinions differ greatly, and understanding still needs to deepen further” (Phase 2, 04). The Supreme Court deliberately declined to pick up that card, so any absolute claim that “AI-generated content has / does not have copyright” oversteps.

One everyday question also remains unresolved: attribution rights in mixed-creation works. An article that is 20% human-edited and 80% AI-written — whose name goes on it? The Beijing Internet Court’s standard (the user is the author) and the US Copyright Office’s standard (substantial human contribution) have inconsistent thresholds, and no domestic precedent yet gives an allocation rule (Phase 2, 04). For now this can only be solved through self-discipline; the end of this post covers how we solve it.

Platform Rules and Disclosure Obligations

Platforms run ahead of legislation — the layer creators feel first. At the end of August 2025, WeChat Official Accounts required proactive declaration of AI-generated or synthesized content, restricting distribution for violators; new rules in April 2026 went straight at “non-human automated creation”, batch-deleting violating articles. Xiaohongshu in February 2026 required AI content to be proactively labeled, throttling unlabeled posts, with the platform also auto-detecting and applying labels. Overseas, the same: Amazon KDP requires disclosure of AI-generated and AI-assisted content; Medium restricts distribution of purely AI articles; LinkedIn launched a private “suspected AI-generated content” flag in August 2026 (Phase 2, 04).

The EU gave this layer legal status, but the enforcement reality deserves separate mention. Article 50’s transparency obligations of the AI Act took effect August 2, 2026; as of mid-September, no fine or formal investigation announcement could be found through public channels: it had been in effect only six weeks, machine-readable labeling obligations for legacy systems are in a grace period until December, and most member states were still finalizing the designation of enforcement bodies (Phase 2, 44). The actual enforcers right now are the platforms — TikTok has cumulatively labeled 3 billion AI-related items, and Meta rolled out an AI Info label (Phase 2, 44). The reading for creators is direct: disclosure is enforced first by platforms, not governments; throttling and deletion arrive faster than fines.

Three Hard Lines in the Chinese Context

China’s lines for individual creators are more concrete than the US/EU — three in total. The first is the labeling line: the Measures for Labeling AI-Generated Synthetic Content took effect September 1, 2025 — AI-generated or AI-assisted content must be proactively declared — and platform enforcement under it rolled out fully in 2026 (Phase 2, 04). The second is the liability line: the Supreme Court’s Opinions make clear that “AI-generated” is not a defense — when sued for copyright infringement, you cannot escape liability by saying “the content was written by AI”; fault is determined by five factors, including the party’s role in the content-generation process and who fed the training data. On the user side there is also a red line: knowingly, or where one should have known of a prior work, using AI to generate substantially similar content without grounds for exemption constitutes infringement (Phase 2, 04). This last one hits directly at the “use AI to rewrite someone’s viral hit” play — you “should have known” about that hit. The third is the criminal line: in June 2025 the Beijing Tongzhou Court handed down the nation’s first criminal AI fine-tuning case — four people used AI tools to alter original artworks into jigsaw puzzles, profiting over 270,000 yuan, constituting the crime of copyright infringement; “AI fine-tuning does not equal originality” got a criminal characterization (Phase 2, 04).

The platform side carries another layer of constraint: the Interim Measures for the Management of Generative AI Services require service providers to bear proactive review obligations, and the Hangzhou Internet Court’s January 2026 first AI-hallucination case further established the adjudication framework of “fault liability plus a dynamic duty of care” (Phase 2, 04). This mainly binds platforms and tool developers; individual creators only need to remember its meaning: when things go wrong, “going after the platform” has legal grounding, but blindly trusting AI output yourself also counts as failing the duty of care.

The soft layer is market access: on September 10, 2026, 24 literary journals issued a joint statement — any AI submission, once verified, lands on a blacklist and the author’s work is never accepted again (Phase 2, 04). This is not law, but for submitting writers it is a de facto ban.

A Pragmatic Checklist for Individual Creators

The rules compress into four executable things, ordered by cost from low to high.

First, disclose. If you publish in China and AI was involved, declare it — the cost is one sentence. The price of not disclosing is throttling, deleted articles, and enforcement against “non-human automated creation” (Phase 2, 04). Disclosure plus verification of the output is the starting point of a defense — under the Hangzhou case’s framework, users cannot blindly trust AI output; only multi-source verification counts as meeting the duty of care (Phase 2, 04).

Second, keep your drafts-in-progress. The common prerequisite of the copyrightable precedents is “provable originality with a paper trail”: prompts, iteration records, raw AI outputs, and human-edit diffs all need to be stored (Phase 2, 04). Process drafts have three uses: a creation log when claiming rights, appeal evidence when misjudged by a platform or institution, and self-exculpatory material when the mixed-creation ratio is contested. Storage cost is near zero and the payoff is asymmetric — the math works no matter how you run it.

Third, keep an honest standard on mixed-creation ratios. Legally, human contribution determines copyright strength; on the reader’s side, human contribution determines trust. The most extreme positive sample seen in the research: one author publicly disclosed that 90% of their continuous text was generated by AI under their prompts and constraints (Phase 2, 29, C2). The ratio itself is not illegal; a hidden ratio is the landmine.

Fourth, the mines not to touch, ordered criminal-to-civil. Two criminal: fine-tuning or article-laundering others’ works for profit (the Tongzhou case); generating substantially similar content while knowingly aware of a prior work (the Supreme Court’s user-side red line). Two civil: fine-tuning your own model on pirated corpora (a Bartz-style ending — training lawful, source liable); claiming copyright over a simple prompt (rejected by the Huangpu case). One gray zone: style imitation — no “style-only imitation” case has been won anywhere in the world, and style itself is not protected by copyright, but with enough recognizability you can still be sued; tool developers should not ship style modules of specific writers by default (Phase 2, 04).

Landing Back on Ourselves: Attribution Is the Cheapest Compliance

This blog’s practice closes the piece: every article with AI involvement carries attribution to the AI assistant, noting the collaboration mode. The publishing validator makes copyright_author a required field — AI-written articles are signed with the AI assistant’s name (the “Ganjiang” of this post is the writing slot on the production line), articles written personally by Zhang Peng are signed Zhang Peng, and readers can see who wrote any piece from its copyright block. This post is a sample of that practice: material comes from internal research reports, numbered inline; written by AI; finalized and reviewed by a human.

In most settings this move is still just self-discipline, but it accomplishes three things at once: it satisfies China’s labeling obligation, it is honest with readers, and it leaves traceable process evidence for “human contribution”. All three cost a single line of text. The day a precedent sets the standard for mixed-creation attribution and contribution ratios, creators with attribution and archival habits will already be standing on the side with complete evidence.

Related posts: AI-Flavor Detection and De-flavoring: A Panorama of the Track (Chinese) | A Frontline Writer’s Real Workflow, Distilled into Four (Chinese)