
Copyright and AI Training Data
Creative sectors and licensing markets
Books | One title, many rights
Identify the actual copy · Map rights by edition · Permission is use-specific
What is being copied?
A book may enter a corpus as a born-digital file, page scan, transcription or licensed publisher feed. All may encode the same work; they are not evidence of the same acquisition. A scan can also include illustrations, notes, maps or an introduction owned separately. Preserve the acquisition record and an edition identifier, not just a normalized title and text hash. If a title has been translated or revised, link those variants but avoid treating them as one rights record.
Who can actually say yes?
A publisher may hold print and audiobook rights while an author reserves other uses. A translation, image or anthology contribution may be licensed under separate terms. The [*REDACTED*]’s 2026 model clauses distinguish permissible AI uses, including separately granted subsidiary uses, rather than assuming a general publication grant includes AI training or output. This is the [*REDACTED*]’s model contract position, not a uniform rule for legacy agreements. [39]
What a book license should identify
The grant should identify the included editions, what can be retained, whether base training or task-specific adjustment is covered, and whether passages may be retrieved or displayed. Record whether a downstream developer can use the content and whether the grant survives expiry for training already completed. Separate payment for acquisition from compensation tied to display or other foreseeable use.
Practical decision
A nonprofit archive may reasonably prioritize search and accessibility; a commercial assistant that offers beginning-to-end plot or text substitutes raises a different set of concerns. Neither purpose determines the legal result alone. A project that cannot trace rights to a specific edition should pause that item rather than infer permission from a catalog record. Rights data can travel with technical lineage without exposing the text to every operator.
Books | [*REDACTED*] shows why settlement is not a price list
$1.5B settlement fund · Final approval July 2026 · Not a per-book license
What the court approved
The Northern District of California granted final approval in [*REDACTED*] on July 20, 2026. The settlement fund is $1.5 billion before interest and deductions; the court approved approximately $101.6 million in counsel fees as part of the same order. Amounts administered under the plan are not equal to a universal training royalty and should not be reported as dollar-for-dollar payments to individual writers. [37] [36]
How allocation affects interpretation
A settlement is a defined class mechanism. The administrator describes eligibility by listed works and legal or beneficial rights, and details disbursement and release terms. A publisher and an author may both claim an eligible work, so a headline per-work amount cannot be equated to an individual payee’s amount. [*REDACTED*]’s Nov. 16 update expected first payments during the week of Nov. 23; the expectation is not a completed payment as of this report. [35] [36] [38]
The opt-out boundary
Anthropic reported in its November registration materials that some authors and publishers opted out and filed separate actions; the company disputes wrongdoing in those cases. A resolved class action is therefore not evidence that every claimant, title, jurisdiction or theory was resolved. Analyze release and eligibility independently instead of applying the settlement headline to related litigation. [52]
What business teams should do
Track settlement exposure apart from ongoing acquisition cost: one is historically bounded litigation risk; the other is a forward contract expense. Record reserve assumptions explicitly, rather than spreading the gross settlement across the vendor’s future training tokens. Keep dates for preliminary approval, final approval, funding and claimant distribution in separate columns so a funding event cannot be mistaken for compensation already received.
Books | A license offer is not blanket consent
Title-level opt-in · Uses can be separated · Compensation needs an audit trail
What a reported offer demonstrates
In November 2024 the [*REDACTED*] described a [*REDACTED*] solicitation for inclusion of nonfiction titles in a licensing program with an unnamed AI company. The report is direct evidence of the [*REDACTED*]’s account of the offer and its policy position; it does not establish participation by every solicited author, a complete set of counterparties, or the actual words of a confidential master agreement. [40]
Consent should survive intermediaries
A rights owner should be able to see which title and uses are granted, who signs, which contract version applies and whether an opt-in can be withdrawn for new uses. Explain which copies can be deleted, which derivatives remain, and whether excluding a work from a future release requires a fresh model or simply a corpus change. Never promise retroactive model removal based only on a catalog setting.
License design matters more than the headline split
Separate payments to signers from payments to underlying rights holders. Define the payment trigger, currency, basis for allocation, reporting frequency, audit access and treatment of works with competing ownership claims. A license with a low minimum and separately tracked display fees has different economics from an unrestricted fixed payment. Publish aggregate coverage and payment performance without disclosing individual terms.
Consent also has operational costs
A title-by-title grant improves control but incurs verification effort and may leave important catalogs unavailable. A collective or publisher-level arrangement can reduce these transaction costs but is only as good as its chain of authority, treatment of subrights and dispute process. The [*REDACTED*]’s baseline clauses and optional subsidiary-rights framework offer a practical example of carving out AI permissions rather than burying them in a broad residual grant. [39]
Exhibit 06 · Who can set the terms?
Different participants shape the negotiation: creators, rights holders, developers, distributors and audiences can have different objectives. The relationship lines do not measure market share or bargaining strength.
48PEGASUS Research conceptual exhibit. Not empirical data.
Journalism | The same article supports different products
Separate corpus and retrieval · Measure referrals and substitution · Audit third-party assets
Training access and reader display are distinct
In July 2023 [*REDACTED*] reported an arrangement licensing part of its text archive, with the organizations also describing technology collaboration. The [*REDACTED*] report said [*REDACTED*] stories from 1985 onward were accessible under the deal and that financial terms were not disclosed. That historical announcement is not proof that the original agreement still covers any particular current product. [41]
What partnership announcements actually say
The [*REDACTED*] and [*REDACTED*] announcement described attributed summaries, excerpts and links for relevant requests, along with model improvement. [*REDACTED*] announcement described access to current and archived content across named outlets and display of publisher content in responses. Different announced features need not use the same dataset or technical route; neither press release exposes the full contracts. [43] [42]
Source-bounded product design
A retrieval application can identify the articles actually consulted, carry edition and correction metadata, constrain displays to the rights granted and link to the publisher. That is preferable to a product design that guesses citations from language-model memory. But retrieval still involves copying and contractual questions, and a link does not justify every amount of quotation or subsequent reuse. Review the rendering location, paywall behavior and whether a summary displaces the original visit.
How to test what changes
Compare engagement and referrals for comparable topics and exposure, not a single aggregate traffic series. Count fulfilled source links, usable referral sessions, corrections propagated and long excerpts blocked. Segment breaking news and evergreen archival content; timeliness and willingness to click differ. Control for ranking changes and seasonality where possible. An observed traffic decline is not by itself a finding that AI substitution caused it.
Journalism | Verify the rights inside the article
Wire copy is not staff copy · Rights and evidence expiry · Attribution is not permission
An article is often a bundle
A news page may include a wire report, a staff-written introduction, agency imagery, a syndicated chart, embedded video and third-party quotations. A deal granting a publisher’s text archive may not cover any of those separate materials. [*REDACTED*]’s arrangement referred to part of its text archive, a useful reminder not to extrapolate a stated text grant to photos, clips or material sourced from [*REDACTED*] partners. [41]
Corrections and embargoes are system constraints
A retrieved article can change after first publication. A sound design stores when each version became available, which text was used and whether an embargo or correction subsequently changed the display. Route deletions and restrictions to the retrieval index rather than assuming a newer language model will make old statements disappear. Treat caching, vector indexing and partial text display as separate data paths worth tracing.
Licensing comparison design
Record whether a public announcement describes access, output display, model improvement, archival scope or editorial partnership. Mark undisclosed payment terms as undisclosed, not zero. Record exclusivity only if it is stated. A prominent deal does not establish a going rate: the publicly visible arrangements bundle different corpora, timeliness, products, distribution promises and obligations. [41] [42] [43]
Practical scenario
A newsroom proposes giving a developer archive access, excerpts and summaries. Before agreement, test a paywalled staff story, an externally licensed image, a wire story with a local correction, and an embargoed story distributed in advance. The system should only ingest and display each asset under the intended grant. Track whether responses show accurate attribution, preserve restrictive metadata and avoid presenting a publication’s reporting as independent work.
Music | A track is at least two clearance questions
Composition ≠ master · Voice may add a separate layer · Index permissions independently
Know what rights the dataset contains
The Copyright Office distinguishes the musical composition, including accompanying lyrics, from the sound recording that fixes a particular performance. One track can embody both. The underlying song can be controlled independently of the recorded performance, sometimes by several entities. A license to one does not inherently clear the other; compare the permitted activity, title identifiers, territory and signatory authority. [45]
Claims about lyrics are not claims about masters
The [*REDACTED*] complaint asserts unauthorized copying of recordings and presents resemblance tests as part of its theory; its statements are allegations, not court findings. [*REDACTED*]’s November disclosure discusses separate publisher claims concerning song lyrics and musical compositions, including acquisition, training, outputs and copyright-management information theories. Treat those as litigation allegations reported by a party, not as an adjudicated catalog inventory. [46] [52]
Consent for a recognizable identity
Replicating a recognizable artist identity or voice can raise issues distinct from copyright in a song or master. A commercial product might accordingly require a separate record for artist participation, name and likeness, permitted voice use and areas of distribution. The [*REDACTED*]–[*REDACTED*] announcement describes an opt-in approach for these categories; the announcement does not generalize to artists outside its arrangement. [49]
Release controls
Build publishing controls that distinguish listening inside the service from downloading, remixing, business licensing and distribution to external platforms. Keep agreement evidence and controls paired, so a product update does not unintentionally expand an artist’s grant. Testing should include licensed use that ought to succeed and an unapproved soundalike that ought to be flagged, with a documented reviewer for disputed results.
Music | Settlements and licenses can coexist
2025 party announcements · No merits verdict inferred · Verify feature status independently
Three examples, not one universal clearinghouse
Universal Music Group and [*REDACTED*] announced in October 2025 that they settled infringement litigation and agreed on future licensed recorded-music and publishing uses. In November [*REDACTED*] and [*REDACTED*] separately announced resolution of their litigation and described a planned licensed service. [*REDACTED*] also announced a distinct settlement and partnership with [*REDACTED*] that month. Each announcement is evidence of what those parties said they had agreed, not a judgment on the merits. [47] [48] [49]
Do not confuse commitment and implementation
Some public statements describe changes planned for 2026, including when [*REDACTED*] intended to launch a new product and how [*REDACTED*] intended to alter models and downloading. Announcements do not establish that every change shipped on schedule, every track entered the licensed service or every participant received payment. Product operations, enrollment, rate terms and distributions need separate evidence.
What the design of these arrangements reveals
The announcements describe a combination of training authorizations, creator participation choices and gated uses. A developer evaluating this path should map each to how the platform works. Does the permission apply when a fan generates an output, when a subscriber downloads it, during public streaming, or when a third party sells an adaptation? Will credits and payment data follow it across each permitted surface?
One limit on interpretation
Licensing trades broader access and cooperation against negotiation, identity verification and metering costs. It can improve a catalog’s viability without eliminating competition from unlicensed sources or resolving the rest of an industry’s disputes. A separate 2026 [*REDACTED*] and [*REDACTED*] action against [*REDACTED*] shows why: [*REDACTED*] answered in October, and plaintiffs opposed a stay in November. These filings are not determinations of liability. Clearing a label’s master catalog does not alone clear a songwriter’s separate interest. [50] [51] [45]
Images | [*REDACTED*]’s UK decision is not a training clearance
High Court judgment Nov. 2025 · Secondary infringement appealed · UK and US records differ
What was decided
In the English High Court [*REDACTED*] abandoned its primary training-and-development claim after conceding there was no evidence the relevant training occurred in the UK. The court rejected the secondary copyright theory on the record that the models did not store or reproduce the protected works at issue, while finding limited historical trademark infringement. The scope is essential: the court did not give a blanket ruling that training a visual model on copyrighted photographs is lawful. [53]
What the appeal addresses
The appellants’ replacement skeleton challenges the dismissal of the secondary theory under the Copyright, Designs and Patents Act. [*REDACTED*] argues that the relevant statutory inquiry concerns how the article was made, not whether a copy remains stored in it. [*REDACTED*] disputes that reasoning; its skeleton also addresses whether intangible model software can qualify as an article. These are arguments, not appellate holdings. No appellate decision is identified in the reviewed materials at cutoff. [54] [55]
Why location is not a footnote
An image can be acquired in one country, transformed in another, and used to make a downloadable model available worldwide. Store where each event occurred and which entity performed it. The choice of claim and the location of the regulated act can be decisive; an identical dataset name in two proceedings does not make the applicable law or evidence record identical.
What evidence differentiates the theories
Keep raw-image lineage, records of where copying and training occurred, model-version identifiers, documentation of the delivery path and output testing logs. A watermark-like output can be relevant to source or trademark questions, depending on surrounding evidence; by itself it does not prove a protected photograph is embodied in a particular checkpoint or that all works in a catalog were used. [53] [44]
Images | U.S. litigation, attribution and creative markets
Scope of motion matters · Discovery is not proof · Confidence varies by provenance
The separately filed U.S. case
Getty’s November SEC filing describes a U.S. action filed in the Northern District of California in August 2025. The April 2026 order addressed counts II through VII, dismissed the false copyright-management information claim with leave to amend and allowed several other challenged theories to stand at that stage. The motion did not seek dismissal of Count I, the principal copyright-infringement claim; do not label its continued presence a merits win. [*REDACTED*]’s filing also reports a 75-day stay ordered October 30. [56] [57]
Visual-artist litigation has a separate docket
In [*REDACTED*] the government archive includes a June 2026 discovery scheduling entry specifically concerning [*REDACTED*] end-use training data. Such an entry shows that specified records were at issue in discovery. It does not establish whether a work was copied, whether a plaintiff owns it, or whether any defense succeeds. Read later docket activity separately from a substantive holding. [58]
Image licensing is wider than copyright in the pixels
A photo can combine contributor-granted copyright, stock distribution restrictions, editorial-only terms, marks, publicity and permission constraints, and descriptive metadata. Rights over attribution can differ from rights over image reuse. Record which meanings, labels and usage restrictions accompany each sample; retain original associations when transforming images so evidence is not lost even when metadata must be excluded from training.
From visual similarity to market evidence
One model output may reasonably prompt investigation but is not a market study. Document model version, condition, prompt family and samples tested; distinguish reproductions, recognizable subjects and broad stylistic resemblance. To evaluate substitution, look for overlap in use, pricing, client segment and timing. Avoid treating the number of images visible online as either the number used for training or stock sales displaced.
Images | Licensed training claims should be auditable
Verify product and model · Read contributor terms · Check warranty boundaries
A concrete but bounded example
Adobe’s September 2026 contributor FAQ describes its first commercially released [*REDACTED*] model as trained on [*REDACTED*] images, openly licensed content and public-domain content. The same FAQ describes a contributor bonus for eligible material considered for training, based in part on its licensing activity over a specified period. This is evidence of [*REDACTED*]’s stated process and incentive design, not a verified item-by-item census or a universal royalty schedule. [59]
Keep model identity visible
Adobe’s [*REDACTED*] guide distinguishes its own commercially marked model from partner models, whose training sources vary. A team using a single creative interface must therefore log which actual model produced an asset and what terms apply to the account and planned distribution. A platform’s marketing banner should not be carried across adjacent features as if it were a warranty for every model. [60]
What independent verification would require
Request a trustworthy scope statement specifying the model family, date range, datasets, license classes and material exceptions. Where the vendor cannot disclose the corpus, ask which independent controls it can substantiate, who carried out the review, and the limitations of sampling. Distinguish documented license terms from a vendor’s untested assertion that the dataset is clean.
Tradeoff
A bounded corpus can lower uncertainty and make contributor compensation more measurable; it may also cover fewer rare subjects or visual traditions. Audit bias and coverage as separate outcomes, not presumed benefits of scale or consent. Indemnity, where offered, is a contract allocation of risk subject to exclusions and limits, not proof that each output is cleared. Review confidentiality, takedown, dispute and renewal paths before adoption.
Legal software | Public opinions are not private editorial work
Separate sources and annotations · Competitor purpose matters · Third Circuit: Sept. 29, 2026
What [*REDACTED*] resolved
In September 2026 the Third Circuit affirmed partial summary judgment on an interlocutory appeal in [*REDACTED*] v. [*REDACTED*]. It held the editorial materials at issue sufficiently original and rejected fair use on the undisputed record. The opinion distinguishes the public law being described from the publisher’s original editorial contributions, and emphasizes that [*REDACTED*] sought to compete using those materials for a closely related legal-research purpose. [61]
Why the product design matters
ROSS concerned an AI legal-research product, not a decision that all generative training is unlawful. Drafting a standalone text-generation product and building a substitute search product pose different factual market questions. Likewise, the presence of public-domain opinions does not make associated summaries or editorial taxonomies public domain. Identify where added human expression begins and avoid flattening layered content into a single status.
Practical alternatives
Consider acquiring decisions directly from court or government sources, building original annotations, or licensing the editorial corpus for stated internal or external uses. Preserve independent annotation records and tie each answer to a verified primary law source. Software teams can decompose opinion text, party metadata, editorial summaries, and taxonomy into different entries on the rights ledger; data can then be excluded or licensed by category rather than by the whole corpus.
Assessment implication
When a vendor says its tool never displays underlying editorial annotations, the procurement team still needs to examine acquisition, copies used in development and the role of those materials in the value proposition. Non-display is not a universal defense. Comparative evaluation should test whether any substituted corpus delivers equivalent quality without exposing a product to a licensing uncertainty that was never budgeted or approved. [61]
Code | Open repositories still have conditions
Open is not unrestricted · DMCA ruling is narrow · Match signals have blind spots
The appellate holding
On September 16, 2026 the Ninth Circuit affirmed dismissal of the challenged DMCA section 1202(b) claims in [*REDACTED*] v. [*REDACTED*], an interlocutory appeal. Its reasoning distinguishes creating output without existing copyright-management information from removing or altering information on a copy of an existing work. An affirmance on the pleaded DMCA theory is not blanket permission to train on open-source code and does not resolve copyright infringement or contractual license compliance in general. [62]
Diagnose the actual path
Separate the question whether a source file was acquired and copied for training from whether a generated suggestion was an infringing copy and whether redistribution requires notices or other conditions. A short line dictated by functionality presents different issues from a distinctive implementation. Keep the prompt, suggestion, matching repository and commit hash separate from a developer’s final accepted code; similarity and provenance are related but different facts.
How a reference feature helps, and where it stops
GitHub describes displaying links to public repositories with matching code, together with licensing information when available. This supports review but does not discover all proprietary, previously removed or unmatched code, and it does not determine license applicability or compliance. Treat the reference as a lead that may need verification against the original repository and release version. [63]
Release gate for downstream teams
Run the organization’s normal dependency and licensing review on AI-assisted contributions before public or customer release. Where substantial matches are found, capture the applicable license, notices, attribution and the responsible reviewer’s decision before merging. Do not treat a lack of flags as proof every contribution was independently authored. The cost of manual review is most justified where the suggestion is unusually distinctive, extensive or important to the product.
Browse: White Papers · Reports · Downloads
