AI Copyright Rules by Jurisdiction

Summary

AI copyright rules are jurisdiction-specific and still evolving. The US requires human authorship — fully AI-generated works are not copyrightable (the Supreme Court denied certiorari in Thaler in early March 2026), and AI training fair use is unsettled with only conflicting district-court rulings (no appellate decision; Thomson Reuters v. ROSS is on appeal). The EU permits training under the CDSM text-and-data-mining exception subject to opt-out, and EU AI Act Article 53 copyright/transparency fines (up to 3% of global turnover or EUR 15M) become enforceable on August 2, 2026. Japan broadly permits training under Article 30-4; China grants copyright to AI-assisted works when the creator documents creative effort. The biggest recent shift: the UK's 18 March 2026 government report dropped the proposed broad TDM exception with opt-out and reverted to a "wait and see," industry-led-licensing approach. Always determine the applicable jurisdiction(s) first and apply the most restrictive rule across them. [src1, src2, src3, src11, src13]

Rule

AI-generated content is not copyrightable in the US unless a human author exercised meaningful creative control over the output; the EU permits AI training under a text-and-data mining exception but requires rights reservation compliance and training data transparency; Japan broadly permits AI training under Article 30-4 unless it unreasonably prejudices rights holders; China grants copyright to AI-assisted works when the human creator demonstrates sufficient creative effort; and the UK, after its 18 March 2026 government report, dropped its proposed text-and-data-mining exception with opt-out and reverted to relying on existing copyright law plus industry-led licensing. Any entity creating, training, or distributing AI-generated content must determine its applicable jurisdiction and comply with jurisdiction-specific rules on authorship, training data use, and disclosure. [src1, src2, src3, src11]

Evidence

In Thaler v. Perlmutter (D.C. Circuit, March 18, 2025), the court unanimously affirmed that copyright requires human authorship, holding that the Copyright Act's ownership, inheritance, duration, and transfer provisions all presuppose a human author; the U.S. Supreme Court denied certiorari in early March 2026, leaving the human-authorship requirement settled at the highest level for fully AI-generated works. [src2, src13] The U.S. Copyright Office's January 2025 Part 2 report confirmed that prompts alone do not establish authorship — even detailed, labor-intensive prompts "may reflect a user's mental conception or idea, but they do not control the way that idea is expressed"; its Part 3 report on AI training (pre-publication, May 2025) concluded that training is evaluated case-by-case under the four fair-use factors, that the first and fourth factors carry the most weight, and that commercial use of copyrighted works to produce content competing with the originals — especially via illegally accessed data — goes beyond established fair-use boundaries. [src1, src12] In the EU, GPAI model providers face fines of up to 3% of global annual turnover or EUR 15 million for non-compliance with Article 53 copyright transparency obligations, enforceable from August 2, 2026; the EU AI Office published its mandatory training data summary template on July 24, 2025. [src3, src10] In January 2026, the European Parliament proposed compromise amendments requiring "an itemized list identifying each copyright-protected content used for training" regardless of training jurisdiction, with non-compliant providers facing potential operational bans in the EU. [src9] In China, the Beijing Internet Court ruled (September 2025) that AI-generated images are eligible for copyright protection, but the creator must document their creative thinking, input commands, and the selection/modification process with supporting evidence. [src5] U.S. fair use litigation produced conflicting outcomes: Thomson Reuters v. ROSS Intelligence rejected fair use for commercial, competitive AI training on curated legal headnotes (now on interlocutory appeal before the Third Circuit), while Bartz v. Anthropic found LLM training on lawfully acquired books was "exceedingly transformative" — though Anthropic ultimately settled for $1.5 billion over piracy-related claims (final claim deadline was March 30, 2026). In Kadrey v. Meta (June 2025), the court held that fair use can apply regardless of whether the source copies were lawfully acquired, but the plaintiffs' piracy-related claims survived. [src6, src13] The New York Times v. OpenAI MDL remains in discovery — in March 2026 the court ordered OpenAI to produce roughly 108 million output logs, with summary judgment briefing expected around April 2026. [src13] In the music sector, the landscape shifted from contested fair-use fights toward licensing settlements: Universal Music settled with Udio (October 2025) and Warner Music settled with Suno (November 2025), each pairing the settlement with a licensed AI-music partnership; Sony's suits against Suno and Udio remain unsettled with a pivotal fair-use ruling expected summer 2026, and UMG, Concord, and ABKCO filed a separate ~$3 billion lyrics-training suit against Anthropic in January 2026. [src13] In the UK, the High Court issued a narrow November 2025 ruling in Getty Images v. Stability AI, addressing secondary copyright infringement and trademark claims rather than the core training question. [src13] Mexico's Supreme Court ruled in 2025 that works generated exclusively by AI are not eligible for copyright protection, and France's competition authority fined Google EUR 250 million for using news articles without permission in training Gemini. [src6]

Key Properties

• US Human Authorship Requirement: Works fully generated by AI are not copyrightable; AI-assisted works are copyrightable only for the human-authored elements, and applicants must disclose AI-generated content during registration [src1, src2]
• EU TDM Exception: CDSM Directive Article 4 permits text-and-data mining on lawfully accessed works unless rights holders expressly opt out; scientific research TDM is exempt from opt-out requirements [src3]
• EU AI Act Article 53 Transparency: GPAI providers must publish detailed training data summaries using the mandatory template (published July 2025) and respect opt-out reservations; fines up to 3% of global turnover or EUR 15M, enforceable from August 2, 2026 [src3, src10]
• EU Parliament 2026 Proposals: Compromise amendments (January 2026) would require itemized training data lists, licensing regime with remuneration, centralized EUIPO opt-out register, and operational bans for non-compliance — not yet enacted [src9]
• Japan Article 30-4: Permits any exploitation of copyrighted works for non-enjoyment purposes (including AI training) unless it "unreasonably prejudices" the rights holder's interests [src4]
• China Demonstrated Creative Effort Standard: Courts grant copyright to AI-assisted works when the human creator proves creative input through documented prompts, selection, and modification [src5]
• UK Status Quo (reform reversed): The 18 March 2026 government Report on Copyright and AI dropped the previously preferred broad TDM exception with opt-out; the government now favors a "wait and see," industry-led-licensing approach. Next steps: a summer 2026 consultation on digital replicas/deepfakes and an autumn 2026 interim report on AI-content labeling — no broad TDM exception or new transparency regulator is being taken forward for now [src7, src11]

Conditions

• Applies when: Any entity creates content using generative AI, trains AI models on copyrighted data, distributes AI-generated works commercially, or seeks copyright registration for AI-assisted output
• Does NOT apply when: AI is used purely for internal analysis without generating or reproducing copyrighted expression (e.g., sentiment analysis, classification); when works are entirely human-authored with AI used only for spell-checking or formatting
• Confidence degrades when: Pending cases (NYT v. OpenAI summary judgment ~April 2026, Sony v. Suno/Udio fair-use ruling expected summer 2026, Thomson Reuters v. ROSS Third Circuit appeal, Disney v. Midjourney, In re Google Generative AI) produce rulings; the UK advances its summer 2026 deepfake/digital-replica consultation or autumn 2026 labeling report into legislation; EU Commission finalizes TDM opt-out protocols and AI Act Article 53 enforcement begins (August 2, 2026)

Constraints

• Fair use for AI training in the US has no appellate-level decision — only district court rulings with conflicting outcomes as of May 2026 (Thomson Reuters v. ROSS is on interlocutory appeal before the Third Circuit); do not present fair use as settled law [src6, src13]
• The EU AI Act copyright enforcement (Article 53 fines) does not take effect until August 2, 2026 — current non-compliance carries no penalty under the AI Act itself [src3]
• Japan's Article 30-4 has not been tested in court for AI training as of March 2026 — the exception is broad in statute but the "unreasonable prejudice" proviso is untested [src4]
• China's copyright rulings are from lower courts (Beijing Internet Court, Hangzhou Internet Court, Changshu People's Court) and have limited precedential value outside those jurisdictions [src5]
• The UK has no current AI-specific copyright legislation, and after the 18 March 2026 report the government is not advancing a TDM exception — existing Section 9(3) CDPA 1988 (computer-generated works) predates generative AI and may not apply [src7, src11]
• Opt-out effectiveness varies: the Hamburg Court (December 2025) ruled an opt-out was invalid because it was not machine-readable, demonstrating that technical implementation matters [src8]

Rationale

Copyright law exists to incentivize human creative expression by granting authors exclusive rights to their works. The central tension with AI is whether outputs that lack direct human creative control should receive the same protection, and whether using copyrighted works to train AI models infringes on authors' exclusive reproduction rights. Most jurisdictions are converging on requiring some threshold of human creative involvement for output protection, while diverging significantly on whether and how training on copyrighted inputs is permitted. [src1, src2, src3]

Framework Selection Decision Tree

START — User needs copyright guidance related to AI
├── What activity?
│   ├── Registering/protecting AI-generated output
│   │   ├── US → Human authorship required; disclose AI use; only human elements protected
│   │   ├── EU → National rules apply; human authorship generally required
│   │   ├── Japan → General authorship principles apply; no AI-specific rule
│   │   ├── China → Copyright possible if creator documents creative effort
│   │   └── UK → Section 9(3) CDPA may apply; no reform planned (status quo per 18 Mar 2026 report)
│   ├── Training AI models on copyrighted data
│   │   ├── US → Fair use analysis required; outcome depends on case specifics
│   │   ├── EU → TDM exception applies unless rights holder opted out
│   │   ├── Japan → Article 30-4 permits training unless unreasonable prejudice
│   │   ├── China → No specific training exception; general principles
│   │   └── UK → Existing TDM limited to non-commercial research; no broad exception being added
│   └── Protecting own content from AI training
│       ├── EU → Machine-readable opt-out under CDSM Article 4(3)
│       ├── US → No legal opt-out; robots.txt not legally binding
│       ├── Japan → Opt-out not recognized under Article 30-4
│       └── UK → No statutory opt-out; market-led machine-readable controls encouraged
└── Multiple jurisdictions?
    └── Apply most restrictive rule from all applicable jurisdictions

Application Checklist

Step 1: Determine applicable jurisdiction(s)

• Inputs needed: Where the AI system is developed, where content is published/distributed, where users are located, where the training data originated
• Output: List of applicable jurisdictions with their respective copyright regimes
• Constraint: If multiple jurisdictions apply, the most restrictive rules from each must be satisfied [src3]

Step 2: Classify the AI activity

• Inputs needed: Whether the activity involves (a) output registration, (b) training on copyrighted data, (c) distributing AI-generated content, or (d) protecting own content from AI training
• Output: Applicable rules and obligations for the specific activity type
• Constraint: A single project may involve multiple activity types — training and output generation have different rules even within the same jurisdiction [src1, src6]

Step 3: Assess human creative involvement (for output protection)

• Inputs needed: Documentation of human creative choices — selection, arrangement, modification of AI output; prompting alone is insufficient in the US
• Output: Determination of whether output meets the human authorship threshold for the applicable jurisdiction
• Constraint: In the US, prompts alone do not establish authorship regardless of complexity; in China, creative process documentation is required evidence [src1, src5]

Step 4: Evaluate training data compliance (for AI training)

• Inputs needed: Sources of training data, whether rights holders have opted out, whether data was lawfully accessed, whether the use is commercial or scientific
• Output: Compliance assessment for training data use in each applicable jurisdiction
• Constraint: Opt-out reservations must be machine-readable to be enforceable in the EU; robots.txt alone may be insufficient [src3, src8]

Step 5: Document and disclose

• Inputs needed: AI involvement in the creative process, training data sources, opt-out compliance records
• Output: Registration disclosures (US), training data summaries (EU), creative process documentation (China)
• Constraint: Failure to disclose AI-generated content in US copyright registration may void the registration; EU non-disclosure carries fines from August 2026 [src1, src3]

Decision Logic

If a user in the US wants to register copyright in a fully AI-generated work (prompts only)

→ Advise that registration will be refused: the Supreme Court denied certiorari in Thaler (early March 2026), and the U.S. Copyright Office holds that prompts alone do not establish authorship. Only human-authored elements (selection, arrangement, substantial modification) are registrable, and AI-generated content must be disclosed. [src1, src2, src13]

If a US developer assumes AI training is categorically fair use because of Bartz v. Anthropic

→ Stop. There is no appellate ruling; Thomson Reuters v. ROSS (rejecting fair use) is on appeal to the Third Circuit, and the Copyright Office Part 3 report treats training as a case-by-case four-factor analysis where commercial output competing with the originals — especially from pirated data — is unlikely to qualify. Conduct a fact-specific analysis, and never train on illegally accessed works. [src6, src12, src13]

If a GPAI model provider serves the EU market

→ Implement an Article 53(1)(c) copyright policy, publish the training-data summary on the EU AI Office template, and honor machine-readable opt-outs now — Article 53 fines (up to 3% of global turnover or EUR 15M) become enforceable from August 2, 2026, and the EU Parliament's January 2026 amendments may add itemized training-data lists and operational bans. [src3, src9, src10]

If a rights holder wants to stop AI training on their content

→ The mechanism is jurisdiction-specific: in the EU, use a machine-readable opt-out under CDSM Article 4(3) (the Hamburg Court invalidated a non-machine-readable one in December 2025); in the US there is no statutory opt-out (robots.txt is voluntary); in the UK there is no statutory opt-out and the government (18 March 2026) is encouraging market-led controls; under Japan's Article 30-4, opt-out is not recognized. Layer contractual terms of service everywhere. [src3, src8, src4, src11]

If a user relied on the expected UK "TDM exception with opt-out"

→ Update them: the 18 March 2026 government report dropped that proposal and adopted a "wait and see" / industry-led-licensing stance. Plan around existing UK copyright law and licensing, not a forthcoming exception; watch the summer 2026 deepfake consultation and autumn 2026 labeling report. [src7, src11]

If a creator in China wants copyright in an AI-assisted work

→ Tell them it is achievable but evidence-dependent: Chinese courts (e.g., Beijing Internet Court, Sept 2025) grant protection when the human documents creative thinking, input commands, and the selection/modification process. Advise them to retain that documentation; note these are lower-court rulings with limited precedential reach. [src5]

If the user actually needs AI safety/governance rules or data-privacy rules rather than copyright

→ Route them: EU AI Act risk-tier and transparency obligations [compliance/ai/eu-ai-act-summary/2026], US AI governance [compliance/ai/us-ai-regulation/2026], or GDPR for training-data privacy [compliance/privacy/gdpr-summary/2026]. This unit covers copyright authorship, training, and opt-out only. [src3]

If multiple jurisdictions apply to one AI product

→ Apply the most restrictive rule from each applicable jurisdiction; do not assume one jurisdiction's permissive training or authorship rule overrides another's restrictions. Training and output generation also carry different rules even within a single jurisdiction. [src1, src3]

Anti-Patterns

Wrong: Assuming AI-generated content is automatically uncopyrightable everywhere

Many agents tell users that "AI-generated content cannot be copyrighted" as a blanket global rule. This ignores China's courts, which have granted copyright to AI-assisted works when the creator demonstrates creative effort, and the US rule that human-authored portions of hybrid works remain protectable. [src5, src1]

Correct: Apply jurisdiction-specific rules

Determine the applicable jurisdiction first. In the US, human-authored elements of AI-assisted works are copyrightable with proper disclosure. In China, document the creative process including prompts, selection criteria, and modifications. The blanket "no copyright" rule applies only to works with zero human creative input in the US. [src1, src5]

Wrong: Treating all AI training as fair use in the US

Some assume that because Bartz v. Anthropic found LLM training transformative, all AI training is fair use. This ignores Thomson Reuters v. ROSS Intelligence, where the court rejected fair use because the training was commercial, non-transformative, and harmed the market for AI training data. [src6]

Correct: Conduct a case-specific four-factor analysis

US fair use for AI training depends on (1) whether the use is transformative vs. competitive, (2) the nature of the copyrighted works, (3) the amount used, and (4) the market impact — particularly whether the training creates a substitute product. Commercial, competitive uses that replicate the original's market are far less likely to qualify. [src6]

Wrong: Assuming robots.txt opt-out is sufficient everywhere

Rights holders who add robots.txt directives assume they have legally opted out of AI training globally. But the Hamburg Court (December 2025) ruled an opt-out was invalid because it was not machine-readable in the required format, and Japan's Article 30-4 does not recognize any opt-out mechanism. [src8, src4]

Correct: Use jurisdiction-appropriate opt-out mechanisms

In the EU, opt-outs must be machine-readable under CDSM Article 4(3) — follow the protocols being developed under the EU Commission's December 2025 consultation. In the US, there is no legal opt-out right. In Japan, opt-out is not available under Article 30-4. Use contractual terms of service as an additional layer. [src3, src8, src4]

Counter-Arguments

• Japan's broad Article 30-4 exception demonstrates that permissive AI training rules can coexist with a functioning creative economy; the absence of Japanese litigation suggests the exception is not causing significant harm to rights holders. [src4]
• The US Copyright Office's own Part 2 report acknowledges that the line between "AI-assisted" (copyrightable) and "AI-generated" (not copyrightable) works is inherently fuzzy, and current rules may be difficult to enforce as AI tools become more deeply integrated into creative workflows. [src1]
• China's approach of granting copyright based on demonstrated creative effort may be more pragmatic than the US approach, as it incentivizes human oversight of AI outputs rather than creating a binary copyrightable/uncopyrightable distinction. [src5]

Common Misconceptions

Misconception: Detailed prompting of an AI system makes you the author of its output in the US.
Reality: The US Copyright Office explicitly stated that prompts "may reflect a user's mental conception or idea, but they do not control the way that idea is expressed." Prompting alone does not establish authorship — the human must exercise creative control over the expressive elements. [src1]

Misconception: The EU AI Act bans AI training on copyrighted data.
Reality: The EU permits AI training under the text-and-data mining exception (CDSM Article 4) as long as the content is lawfully accessed and rights holders have not opted out. The AI Act adds transparency obligations but does not prohibit training. Scientific research TDM is exempt from opt-out entirely. [src3]

Misconception: If a work is not copyrightable because it is AI-generated, anyone can use it freely.
Reality: Even if AI-generated output lacks copyright protection, other legal frameworks may restrict its use: trade secrets, contractual terms of service, trademark rights, right of publicity, and unfair competition laws may all apply. [src1, src2]

Misconception: All jurisdictions require human authorship for copyright.
Reality: China has granted copyright to AI-generated works based on the human operator's demonstrated creative effort. The UK's CDPA Section 9(3) provides a unique provision for "computer-generated works" where authorship is attributed to the person who arranged for the work's creation. [src5, src7]

Comparison with Similar Rules

When This Matters

Fetch this rule when a user asks about copyright ownership of AI-generated content, whether AI training on copyrighted data is legal, how to register AI-assisted works, how to opt out of AI training on their own content, or what the copyright implications are of deploying generative AI in any major jurisdiction. This is the authoritative cross-jurisdictional reference for AI copyright compliance.