
Copyright and AI Training Data
International comparison
International comparison: why location and purpose matter
Five approaches · Territorial rights · Transaction-specific review
The laws ask different questions
U.S. adjudicators weigh use-specific fair-use factors. The EU uses distinct exceptions for qualified scientific research and broader text and data mining, together with separate obligations for providers of general-purpose AI models. UK section 29A is purpose-limited to non-commercial research. Japan’s Article 30-4 asks whether the act is necessary for a purpose other than enjoyment of the expression, and whether it unreasonably prejudices the owner. Singapore provides a use-specific computational data analysis exception with source and onward-sharing constraints. These are different decision rules, not a scale from “permitted” to “forbidden.” [64] [68] [69] [83] [90] [93]
Territorial law does not mean local creators only
The Berne Convention’s national-treatment framework protects foreign creators in other treaty countries subject to the applicable legal system. Article 5(2) links the extent of protection and means of redress, apart from Convention provisions, to the country where protection is claimed. EU Rome II Article 8(1) makes the law of the country for which protection is claimed the general rule for non-contractual IP infringement obligations within its scope. A contract’s governing law is a separate question. [66] [67]
One project, several legal events
A supplier can retrieve source files in one region, deliver replicated shards in a second, train in a third, store weights in a fourth, and distribute weights or provide inference in several markets. The assessment changes with the act and actor. EU market placement can trigger a provider duty even when relevant training acts occur abroad. This regulatory question must not be treated as a finding of civil infringement anywhere. [69]
Decision implication
Create an act-by-act map of source rights, access, license scope, physical or service region, entities, processing purposes, and release markets. Where a rule is uncertain, distinguish admissibility, litigation risk, contractual risk, and business policy. A dataset approved for one purpose is not a passport for every copier, country, or deployment.
United States: fair use is an evidentiary test
Purpose and character · Nature and amount · Market effect
The four factors are not a checklist of points
Section 107 evaluates the purpose and character of the disputed use, nature of the protected work, amount and substantiality taken, and effect on the potential market or value. It does not create an AI-specific categorical permission. A model-development objective can inform purpose, but the amount taken, why it was necessary, how it was acquired and held, and what markets the product might affect still matter. Commercial status is relevant, not automatically dispositive. [64] [65]
Map the actual use before contrasting systems
A corpus can involve an initial full download, a persistent institutional library, temporary normalization copies, training runs, and outputs. A strong justification for one does not establish the others. A researcher testing how often a word appears, a book-indexing service, and a commercial tool that lets users substitute answers for a licensed archive do not share identical facts merely because each ingests text.
Keep claims and measurements separate
Document distinct market theories: replacement of the original, substitution for a licensed product in the same market, effects on licensing, and broader competition from generated works. A license offer is evidence of an offer, not by itself proof of cognizable market harm; a lack of verbatim outputs is not conclusive evidence that clients never substitute the service for paid access. Judicial decisions turn on the actual factual record. The Copyright Office training report explains its view in a pre-publication agency report, not a binding rule or merits judgment. [65]
Audit evidence
Version and preserve the stated training purpose, amount of work used, evidence of rights and lawful sourcing, competing implementations, and measured outputs. Separate benchmark conditions from uncontrolled observed usage. Where counsel evaluates fair use, the business owner remains responsible for recording which project facts can invalidate the conclusion, including new sources, functions, or markets.
European Union: research mining and broader mining are distinct
Article 3 · Article 4 · Rights and retention
Article 3: qualified scientific research
Article 3 of the Digital Single Market Directive requires Member States to provide an exception for relevant reproductions and extractions by research organizations and cultural heritage institutions performing scientific research on lawfully accessible material. Beneficiaries must store copies securely; retention can support scientific research and verification. Rights holders can take necessary measures to protect network and database security and integrity. The Article 4 reservation mechanism does not override qualifying Article 3 activities, and Article 7 protects Article 3 against contrary contract terms. The beneficiary and purpose tests prevent a for-profit contractor from assuming a university partner’s exception automatically extends to its own uses. [68]
Article 4: a broader route with a reservation
Broader mining can involve reproductions and extractions of lawfully accessible materials, provided that the rights holders have not expressly reserved the use in an appropriate manner. The law identifies machine-readable means for publicly available online content. Copies made under Article 4 can be kept only as long as necessary for the mining. Retaining an unrestricted archive is a separate question. [68]
Be precise about what is being mined
Definitions and relevant rights extend beyond literary works. The Directive covers defined analytical acts and lists affected rights, including specified database rights. A compilation can involve rights in its structure or extraction separately from rights in individual photographs, music, or text. A license for one layer does not necessarily cover another. National implementation and procedural facts still matter when an exception is contested. [68] [76]
Policy activity is not statutory change
The Commission’s September 2026 consultation covered copyright-protected content in AI alongside other topics and accepted responses until November 10. The consultation is evidence of policy review, not a new exception or compensation rule. Keep enacted rules, guidance, consultations, and court readings in separate status fields, including the effective date where applicable. [75]
European Union: Article 53 and the training summary
GPAI providers · Market placement · Three milestone dates
A distinct provider obligation
Article 53(1)(c) requires general-purpose AI model providers in scope to put a copyright and related-rights compliance policy in place, including identifying and respecting Article 4(3) reservations with state-of-the-art technologies. Article 53(1)(d) requires a public summary of training content using the AI Office’s template. Qualifying free and open-source providers may be exempt from certain documentation duties, but not from the copyright policy and public summary requirements. Recital 106 explains why a covered provider cannot rely solely on where training occurred when placing the model on the EU market. [69] [70]
Applicability and enforcement differ
The Commission’s timeline identifies August 2, 2025 for application of GPAI obligations; August 2, 2026 for Commission enforcement powers; and August 2, 2027 as the deadline for models already placed on the market before August 2, 2025. Classify the relevant model, release, and market-placement date before applying those milestones. [70] [71]
Template, code, and legal rights
The public summary is not a complete work-by-work catalog or an assurance that each input has been cleared. The Commission template asks for model and data context, relevant source categories and processing information. The GPAI Code of Practice is a voluntary means of showing compliance, not a substitute for copyright permission or an adjudication of individual works. For some modified models, a downstream firm can itself become a provider; the Commission FAQ describes a summary for the modification linked to the underlying model’s summary. [71] [72]
Build from versioned records
Tie public statements to an approved internal release manifest and archive prior versions. Track additional training datasets and the date and significance of changes. Share required information without needlessly exposing personal data, security-sensitive sourcing details, or irrelevant secrets. Preserve internal record-level evidence for authorized investigations rather than treating the public summary as the complete audit ledger.
Exhibit 07 · EU duties and enforcement run on different dates
The chart distinguishes applicability of GPAI provider obligations, the Commission’s enforcement powers, and the transition for models placed on the market before August 2, 2025. Consult Article 111 and the Commission guidance for applicability to a given model.
European Commission, guidelines for GPAI providers: primary source
Signals, reservations, and repeat collection
Technical signals · Legal status · Evidence over time
A browser permission is not a rights conclusion
Public availability, contractual access, crawler controls, and an explicit copyright reservation are different signals. An accessible URL does not establish a license; the absence of a blocking instruction does not prove the rights holder gave affirmative consent. A technical exclusion can govern particular crawler identities or paths without capturing every downstream copy or intended purpose. For EU Article 4, the legal question is whether a reservation was expressly made in an appropriate way by the relevant rights holder at the relevant time. [68]
Interoperability is still developing
The Copyright chapter of the voluntary GPAI Code includes measures concerning robots.txt and other appropriate machine-readable protocols. The Commission consultation sought input on how such signals are used; a July 2026 study assessed the feasibility of an EU-level registry, not a registry created by that study. Keep protocols, legal effectiveness, and technical enforcement separate. A product-specific setting can complement but need not replace express terms or controls elsewhere. [72] [73] [74]
Evidence design for a corpus owner
Capture the exact address, retrieval date, agent identity, relevant headers and site controls, plain-language terms, and applicable contract version. If rights signals disagree, record the inconsistency rather than selecting the most permissive result. Refresh targeted checks before reuse and keep a trace from a domain-level declaration to affected individual items. An authorized reviewer, not a crawler heuristic, should resolve disputed claims and direct suspension or release.
Rights-holder perspective
Publish consistent terms under the control of the actual rights holder, distinguish ordinary search from training and real-time retrieval where the relevant tools allow, and keep timestamped records of exactly what was served. A licensing contact, scoped catalog and auditable exclusion request can make commercial discussions more useful. Do not assume a later notice determines past liability or constitutes a technical unlearning request. The legal and operational paths require different evidence. [68] [73]
Germany: [*REDACTED*] concerns corpus preparation, not every use
Research exception · General mining · Review pending
Distinguish the artifact from the training copy
The Hamburg case concerns a photograph downloaded for a comparison operation during preparation of an image-text dataset. The EUIPO account of the September 27, 2024 Regional Court judgment identifies the scientific-research exception in German section 60d as the basis for dismissal. A dataset of links and descriptions does not, by itself, eliminate the separate act of temporarily copying an image to check relevance. Conversely, deciding that preparation-step issue does not adjudicate every later commercial model trained with resources found through the dataset. [77] [76]
The scope of the second-instance result
According to [*REDACTED*]’s counsel reporting on the December 10, 2025 appeal, the Higher Regional Court affirmed dismissal and addressed both section 44b general mining and section 60d scientific research. Counsel’s account describes the disputed natural-language reservation as not shown to satisfy the applicable machine-readability requirement on these facts. That is not a universal conclusion about natural-language terms or current technical capabilities. [79] [76]
What remains unresolved
The case continued to Germany’s Federal Court of Justice under I ZR 281/25. The appeal was heard on September 3, 2026 and a pronouncement was scheduled for December 17, 2026. The schedule is not a merits outcome, and any impressions reported by counsel after the hearing are not a court decision. Recheck the official case listing before relying on the posture. [78] [79]
Operational caution
Classify a dataset into its metadata, hyperlinks, thumbnails, and separately fetched source objects. Preserve which rights attach to which layer; the permissive terms of a metadata release cannot confer permission to copy third-party works linked from that release. Where a supplier cites its own research exception, request a distinct explanation of the customer’s intended activities before ingesting anything.
Germany: [*REDACTED*]’s [*REDACTED*] and [*REDACTED*] cases address memorization
Lyrics versus music · Model copies versus outputs · Trial rulings
[*REDACTED*]: lyrics and substantial reproduction
The Munich I Regional Court’s November 11, 2025 subject-matter release reports that the court largely granted [*REDACTED*]’s claims for injunctive relief, information and damages against two [*REDACTED*] entities for the lyrics in dispute. It found relevant reproduction through memorization in the models and reproduction of lyrics in chatbot outputs; it did not treat that model memorization as covered by the German mining exception. It dismissed the separate personality-rights claims concerning mistaken attribution of altered lyrics. This is a first-instance, record-specific ruling, not a holding about all generative model weights. [80]
[*REDACTED*]: compositions rather than lyrics
The same court announced on July 31, 2026 that it largely granted [*REDACTED*]’s claims concerning six musical works; the dispute did not cover lyrics as such. On the evidence before it, the court found recognizably reproducible material in versions v3.5 and v4 stored on servers in Germany and found the memorized reproduction outside section 44b. It separately addressed outputs in Germany; the release attributes liability to the provider on the examined facts, including the character of the prompts. [81]
Different objects, different questions
The decisions ask what protected elements are reproducibly represented and which person or service performed the legally relevant act. Whether another model of another modality encodes particular works, whether the output came from a retrieval index rather than parameters, or whether a copy of the disputed artifact exists in the jurisdiction are factual questions, not assumptions transferable across releases. Test results may inform those questions but cannot answer them conclusively by themselves.
Weights and evidence
Log the exact artifact, jurisdiction and serving mode. Compare training, stored checkpoint, output, and distribution evidence separately. Capture whether the source was independently retrieved at inference or supplied in the prompt. Escalate reproduction reports to reviewers with access to technical lineage and applicable rights documents; do not convert regional first-instance results into an automatic global ban or safe harbor. [76] [80] [81]
EU press rights: [*REDACTED*] frames retrieval separately
November 3, 2026 · Advocate General · No merits judgment
The questions under referral
A Hungarian court sought a preliminary ruling on EU press-publisher rights in [*REDACTED*] v. [*REDACTED*], C-250/25. The questions address training of large language models and chatbot answers that allegedly reuse online press publications. On November 3, 2026, Advocate General [*REDACTED*] gave a nonbinding opinion. An Advocate General’s proposed answers are not a Court of Justice ruling. [82]
Training was not shown in this record
The Advocate General advised treating the two training questions as inadmissible, reasoning in part that the record did not support the assumption that the publisher’s articles trained [*REDACTED*] and that training was not online use of the press publication for the specific press right at issue. He separately discussed Article 4 TDM for training in principle, subject to its conditions. The opinion does not prove which training data were or were not used across any [*REDACTED*] model. [82] [68]
Retrieval may require a different authorization
For answers grounded in real-time retrieval from a press site, the opinion proposed that reproducing identifiable protected expression or making it available can implicate the publisher’s online rights in certain circumstances. The identity of the source, reuse of expression versus returning facts, and the nature of the communications matter. It would misstate the opinion to report either that every AI-generated summary is infringing or that attribution alone clears any reuse. [82]
A practical design consequence
Store a retrieval event separately from training provenance. For a source collected at inference, retain the release, date, source URL, access policy, evidence needed to distinguish quotation from independently phrased information, and link or citation behavior. A signed training agreement must expressly cover retrieval, caching, excerpt length, attribution and delivery if the commercial product is to depend on those uses. Recheck the official case file for the eventual judgment. [82] [68]
United Kingdom: non-commercial research and a live policy review
Section 29A · No broad opt-out enactment · November 2026 update
The operative mining provision
Section 29A of the Copyright, Designs and Patents Act addresses copies for computational analysis for the sole purpose of non-commercial research. It requires lawful access and sufficient acknowledgement unless impractical or otherwise impossible. Without authorization, transfers to another person or later use for another purpose can infringe. A contractual term is unenforceable to the extent it purports to prevent or restrict a copy that section 29A permits. Distinguish that provision from separate exceptions and contract rights. [83]
A proposed exception did not become law
In its March 18, 2026 report laid before Parliament, the government said it no longer preferred a broad copyright exception coupled with an opt-out mechanism. It identified evidence gaps and conflicting stakeholder views, and said it would continue monitoring licensing, litigation, technology and international experience. The consultation and report are not amendments to section 29A. A commercial developer should not infer a general clearance from the previous policy proposal. [84]
Where the policy is moving
A written parliamentary answer of November 9, 2026 refers to the planned review of creator-control mechanisms, including technical standards and input transparency, and says a working group for independent and smaller creators has been established to explore licensing barriers. The separately published Creative Industries AI Adoption Plan is an independent recommendation, not government enactment. In its November 16 corrected figures, it reports 43% AI use among creative businesses versus 25% among all businesses; these figures concern broad AI adoption, not uptake of AI-training licenses or national preferences on copyright exceptions. [85] [86]
For a partnership
Define an explicit boundary between research and commercialization. Review whether corpus copies, compiled artifacts, checkpoints, or beneficiaries change before transferring or repurposing materials. Version the approval rationale and retain evidence even if a subsequent legislative proposal appears more permissive.
United Kingdom: [*REDACTED*] illustrates the importance of location
Case IL-2023-000007 · Secondary claim · Appeal CA-2026-000007
What the High Court did and did not decide
Getty’s UK primary training and database-rights claims did not proceed to a merits determination at trial after the territorial evidence concerning training and development. In November 2025 the High Court rejected the secondary copyright theory advanced against the final [*REDACTED*] model. It reasoned on that evidence that intangible software could be an article for the relevant statutory provisions, but the particular model weights had never contained or stored an infringing copy of the asserted works. Copies made elsewhere in the development workflow did not, by that fact alone, turn the finished model into an infringing copy. Separate limited trade mark findings are not affirmative holdings about training copyright. [87]
Which question was taken to appeal?
The Court of Appeal materials include [*REDACTED*]’s replacement skeleton and [*REDACTED*]’s response. [*REDACTED*] challenges the determination of secondary infringement under the CDPA, including how the making of an article relates to whether it is an infringing copy. [*REDACTED*] contests [*REDACTED*]’s approach; the respective skeletons set out party arguments, not new findings. No appellate judgment is relied upon here as of November 19, 2026. [88] [89]
Hosting does not tell the whole story
The trial record distinguishes a hosted service accessed by UK users, where inference occurred abroad and the user did not obtain a copy of the model through that service, from downloadable weights. The wider lesson is to trace what is copied, delivered, and stored, who controls the transfer, and which region actually hosts each operation. Facts about one endpoint do not determine other channels. [87]
Operator implications
Do not promise worldwide legal clearance on the basis of where a cluster is located. Identify additional acts such as replicating checkpoints, new training and retrieval in customer regions, or downloading a self-hosting package. Record the relevant model version and geographic evidence so an adverse judgment or license change can be applied narrowly and promptly.
Japan: Article 30-4 turns on purpose and prejudice
Non-enjoyment · Extent necessary · Guidance is not a judgment
What the statute is designed to allow
Japan’s Copyright Act Article 30-4 permits exploitation to the extent necessary for covered purposes that do not involve enjoying, or causing another to enjoy, thoughts or sentiments expressed in a work, subject to a proviso if the use would unreasonably prejudice the copyright owner’s interests given the nature, purpose or circumstances. Data analysis is expressly listed. The provision is not simply a UK-style non-commercial research exception; it is also not blanket permission to use every work in an AI pipeline. [90]
Mixed purposes and narrow collections
The Agency for Cultural Affairs’ English overview draws attention to training where enjoyment of expression is also an objective, and to uses potentially implicated by the proviso. For example, a corpus acquired to build general analytical functionality presents different evidence from fine-tuning designed to produce a particular author’s protected expression. A specialized database sold for analysis presents another issue: extracting and replicating material from the commercial database may engage its owner’s interests even if processing is analytical. [91]
Separate generation from training
The Agency’s General Understanding distinguishes development and training from generation and use. A conclusion about a particular training act does not decide output infringement; the analysis of copyright in Japanese output disputes involves legally relevant similarity and reliance on the existing work. The Agency expressly says the subcommittee’s understanding is not legally binding and should not be read as definitive for specific technologies. [91] [92]
Keep developer records at work level where needed
Document why selected works are needed, how they are accessed, intended use, any evidence of a relevant data-analysis market, and whether product goals or fine-tuning could shift the system toward reproducing protected expression. Review rights and Japanese original wording where the English translation is insufficient for the actual matter. A single “Japan training” label does not resolve these factual questions.
Singapore: permitted computational analysis has boundaries
Sections 243–244 · Lawful access · Limited sharing
A separate statutory approach
The Copyright Act 2021 defines computational data analysis in section 243 to include using material as an example of a type of information to improve a computer program’s functioning, with image-recognition training as an example. Section 244 provides permitted uses for copying or communicating in defined circumstances. It is more specifically framed around analysis than U.S. fair use and is not written as a non-commercial research-only route like UK section 29A. The Intellectual Property Office of Singapore explains that the safeguards matter just as much as the covered activity. [93] [94]
Source restrictions are substantive
The user must have lawful access. Circumventing a paywall does not establish lawful access; database-use terms can matter, subject to the statutory rule that certain terms restricting the permitted use are void. Section 244 also imposes conditions for source copies that infringe, addressing knowledge, flagrantly infringing online locations, and narrowly prescribed purposes. “Openly reachable,” “lawfully accessed,” and “non-infringing” describe different facts. [93]
Copies are not freely shareable
Copies made under the exception may not be used for unrelated purposes. Onward supply is narrowly conditioned; the statute addresses verifying results and collaborative research or study related to the analysis. Providing a bulk training export to an unrelated commercial customer is not shown to qualify merely because the original mining qualified. [93] [94]
Compare the business question, not the slogans
For a Singapore collection project feeding an overseas product, list separately the source copy, analytical act, exports, licensing terms, and future delivery channels. If a customer needs its own corpus copy for ongoing service, ask what distinct permissions support the transfer and buyer’s intended use. Keep contract interpretation separate from whether the statutory exception is available; review privacy and non-copyright restrictions separately.
Cross-border operating map: collection is not deployment
Five location records · Focus on actual acts · Track downstream actors
A hypothetical supply chain
A publisher hosts an archive in Europe; a contractor acquires licensed articles from an authenticated interface; a U.S. team operates preprocessing; a Japanese cloud region trains the model; and an EU entity makes a version available to customers. Nothing about that fact pattern establishes permission beyond the exact license. Evaluate creation of mirror copies, any territorial restrictions or supplier conditions, duration of storage, onward data delivery, and whether placing a covered model in the Union triggers Article 53. [66] [67] [69]
Geographic records must be operationally specific
Record where the source was served, where copies landed, which legal entity made each copy, how processing and backup regions were chosen, where checkpoints and indexes are stored, and where the model or service is placed on the market. A billing address, home office, marketing market and runtime region are rarely substitutes for one another.
Make migrations explicit
A storage relocation can create new copies and change authorized access. A backup policy can outlast the original license term. A customer capable of downloading weights poses a different artifact-distribution question than a remotely served endpoint. If contractual or statutory obligations require a region-specific control, connect it to build and release infrastructure. A contract should spell out whether affiliates and cloud subprocessors can access protected material and where. [69] [87] [81]
What to ask before proceeding
Determine if intended acts match documented permission and review legal exceptions separately from contract obligations. If facts are missing, hold the affected corpus version or region while pursuing evidence. A region-approval record should state its exact boundary and owner and name triggers for reevaluation: changed sourcing, additional languages, new jurisdictions, redistribution or real-time retrieval.
Cross-border licensing and privacy: two independent gates
Rights chain · Privacy and transfer · Approval evidence
Negotiate a bounded rights grant
Identify which rights the licensor controls and whether it can grant training, storage, annotation, evaluation, fine-tuning, external model distribution, or retrieval rights. Mark permitted territories and roles for contractors, affiliates and partners. Include audit rights, corrections, takedown timelines, treatment of snapshots or checkpoints at termination, and responsibility for infringement and restoration costs. A headline “global license” is not a substitute for scope detail or chain-of-title evidence. [66]
Privacy does not follow copyright title
A training license cannot itself establish a GDPR lawful basis, authorize processing special-category personal data, or resolve a covered third-country transfer. EDPB Opinion 28/2024 distinguishes model anonymity from removing obvious identifiers and frames legitimate interest as an assessment of the interest, necessity and balancing of rights. Its application is case-specific. The EDPB Guidelines 03/2026 source cited here is a consultation document with feedback closed October 30, not a new statute. [95] [96] [97]
Keep two approval records
Maintain a copyright record describing provenance, work categories, permission or exception, and uses. Maintain a distinct privacy record describing data categories, collection, lawful basis, minimization, recipients, and transfer assessment where required. An aggregate summary can be public while detailed records remain restricted to authorized reviewers. Link both to the release manifest without copying sensitive source content into a broadly accessible dashboard.
Make rights changes practical
Provide a contact and a status workflow for corrections and disputes. Map an affected work to corpus and product versions, determine forward exclusions, and distinguish retained evidence, deletion obligations, retraining and claims about unlearning. For regions with materially different requirements, authorize releases and features separately rather than presuming that one global deployment plan works.
Browse: White Papers · Reports · Downloads
