
Copyright and AI Training Data
Executive findings and legal framework
Executive summary
FIELD NOTE The legally relevant object is a particular use of a particular work, not an entire product category.
The dispute
Generative models rely on large collections of material, but membership in a corpus does not tell us where copies came from, what was done with them, or whether a generated result competes with the source. Creators and publishers seek consent, compensation, attribution and enforceable limits. Developers argue that learning statistical patterns is a different purpose and that impractical clearance costs can foreclose useful systems. Neither argument resolves the status of every act. [6]
A U.S. fair-use defense weighs the specific purpose and character, nature of the work, amount used, and effect on the relevant market. The district court in [*REDACTED*] treated training, scanning bought books, and obtaining copies from unauthorized online sources as distinct acts. The later $1.5 billion class settlement resolved defined claims and does not set an industry-wide license price. [1] [11] [9]
The finding
The strongest organizational response is to preserve the origin and permitted uses of material at acquisition, then bind each model and retrieval system to a traceable record. Separate training from storage, retrieval, and display. Test outputs for reproduction and substitution before and after release. Document what is known, what was inferred, and what the provider could not substantiate.
This paper compares court records, creative-sector markets, and international rules. It proposes controls for developers, enterprise buyers and rights holders. Controls are operational recommendations, not legal safe harbors. The evidence is current through November 19, 2026; pending litigation and administrative implementation may change. The governing question is not whether AI can learn, but whether an identified party can explain and defend an identified use, with evidence a different reviewer can reproduce.
The six acts that get conflated
FIELD NOTE The same work can move through several legal and technical contexts. Track those transitions separately.
From access to training
Access is the ability to obtain material; acquisition is the act of collecting a copy or licensing access. Retention means keeping a copy in a library, cache or archive. Training occurs when data influence parameters through a learning procedure. Those concepts can overlap, but they are not interchangeable. The evidence for each differs: subscription access terms, an ingest record, retention logs, and training manifests answer different questions.
A right to read a web page does not necessarily cover systematic copying; a right to hold an archival copy does not necessarily cover model training. The inverse also matters: authorization for a run does not automatically allow an indefinite general-purpose archive. These distinctions matter whether the organization relies on permission or an exception. [2] [6]
From retrieval to output
Grounding can fetch material at answer time, including from an index created earlier. Its source can be entirely different from the model’s pretraining inputs. Output may reproduce expression from a retrieved document, a prompt attachment or learned parameters. Similarity by itself does not establish which pathway was responsible, and a missing citation does not show the information necessarily came from weights.
Distribution adds another path: putting source material, reconstructed text or a downloadable artifact into someone else’s hands may require a separate analysis. Compare a local training cache, a torrent upload and a commercial answer; each involves different actors and evidence. A defensible review names the act, source copy, purported authority, audience and output. [2] [1] [6]
Exhibit 01 · One system, several acts
Acquiring and retaining a copy, using copies in model development, and emitting or serving output present different facts. The arrows describe a simplified workflow, not a finding of liability at any step.
48PEGASUS Research conceptual exhibit. Not empirical data.
Rights start with the work and the holder
FIELD NOTE A platform, publisher, distributor and author can hold different rights in the same workflow.
What copyright covers
U.S. copyright generally protects original expression fixed in a tangible medium; it does not extend to ideas, procedures, systems or the discovery of a fact. A news event is not owned as a fact, but an article’s original expression may be. An image’s subject is not the same as the photographer’s choices of framing and timing. A public URL is not the public domain. Publication, registration, ownership and permission are different facts. [2] [12]
One asset can incorporate distinct layers. A recording can embody a composition; a book can contain separately controlled translations, illustrations and excerpts. A vendor may control one version or territory and not another. Determine the asset granularity before asking for rates or recording rights.
What to record
Start with an identifier, version and kind of work. Capture the asserted owner, who authorized a license, term, territory, excluded components, and whether a grant covers subcontractors or onward distribution. Require the authority evidence to follow each materially different use. Merely identifying the publisher is not the same as confirming that it holds training rights for the entire catalog.
Label uncertain material rather than reassigning it into a category that sounds safer. For example, if an image collection contains donated works with an unknown grant, isolate that subset. Excluding the whole collection may be unnecessary; accepting the whole collection can create avoidable exposure. Record reviewed items separately from sampled items and set a plan for unresolved records. [6] [12]
Fair use is a balancing inquiry
FIELD NOTE Four statutory factors. No numerical score, automatic exemption, or universal required percentage.
Purpose and nature
Factor one addresses the purpose and character of the specific use, including commerciality. A materially different purpose may favor fair use, but the degree of difference and the need for copying matter. [*REDACTED*] illustrates that a changed aesthetic does not automatically justify a commercial licensing use with a substantially similar purpose. Discuss what the challenged copying did, not only the developer’s broad research mission. [1] [4]
Factor two asks about the nature of the copyrighted work. A database that mixes facts, functional material, journalism, novels, music and art cannot establish a single work-type conclusion. Creative character and unpublished status can matter, but unpublished status alone does not bar fair use if the factors support it. [1]
Amount and market
Factor three considers the amount and substantiality taken in relation to the whole. A small excerpt may capture what makes the work distinctive. Whole-work copying may sometimes be reasonably necessary for a purpose, but that cannot be inferred from convenience. Document what alternatives were tried and where lesser copies would or would not work. [1] [5]
Factor four addresses effect on the potential market for or value of the work. Show why the system is or is not a substitute for access, sale or a relevant license; do not assume every desired fee proves a legally cognizable market. Then weigh the circumstances as a whole. Neither a rights-policy sign-off nor a model evaluation is a court judgment. [1] [6]
Exhibit 02 · The four U.S. fair-use factors
U.S. courts consider all four statutory factors together on the record of the particular use. These factors are not a checklist, numerical score or categorical permission to train.
48PEGASUS Research schematic; statutory factors: 17 U.S.C. § 107: primary source
Three different kinds of market evidence
FIELD NOTE A lost license, displaced consumption and a crowded genre are different claims.
Demand and substitution
Direct substitution is strongest when a user seeks the original or its function and receives a useful replacement from the AI system, such as an article answer detailed enough to avoid a visit. Measure against an explicit baseline: the number of comparable queries; how much of the original is conveyed; and what respondents actually do when shown each answer type. Traffic changes alone can be confounded by seasonality, search-ranking changes or ordinary product shifts.
An output can differ textually yet compete in a relevant market. Conversely, a technically similar passage can be legally trivial or involve unprotectable expression. Use both textual analysis and evidence of user behavior; neither substitutes for the other. [1] [6]
Licenses and substitution at scale
Rights holders may point to existing training or retrieval deals; developers may dispute scope, comparability or whether the proposed market is one copyright should recognize. A deal is evidence that specific parties agreed to a bundle, not proof everyone owes a rate. Separate training, embedding, display, attribution, product collaboration and legal release when terms are public. If economics are confidential, do not reverse-engineer an assumed price.
Market dilution from cheap generated substitutes is broader than a one-for-one copied answer. Claims about work, genre or profession require a stated economic mechanism and denominator. An observed decline in one creator’s revenue cannot establish a whole-class effect; a developer’s low average regurgitation rate cannot rule out commercially important harms at a narrow tail. [6]
Permission and an exception are not synonyms
FIELD NOTE The approval owner must document the actual legal or contractual basis, not just a risk label.
Two routes, different evidence
A license reflects a rights holder’s grant within a defined scope. It may attach conditions on users, expiry, attribution, hosting, derivatives, distribution, retention, or deployment. If the seller lacks authority to grant a right, its assurances do not eliminate the chain-of-title question. A limited bilateral license between two parties does not constitute an industry precedent. [12] [14]
An exception, where applicable, is a legal basis that can permit particular acts without the holder’s permission. Its facts must be demonstrated under the applicable law. A team relying on an exception should preserve the case-specific purpose, proportionality, scope, and the decision-review record. Saying the company is in the AI business or that the material was easy to download is not such a record. [1] [6]
When routes change
A permission may expire or an uncertain basis may need to be reevaluated after a scope change. A rights registry therefore should not store a single timeless flag for allowed. It should record what was allowed, under which conditions, who approved it, and when to check again. Acquisition, retention, pretraining, fine-tuning, evaluation, retrieval and excerpt display can have separate fields.
If a grant is ambiguous or conflicts with another signal, hold the affected material while the relevant policy owners decide. Record a bounded exception if necessary, specify its confidence and review time, and keep excluded assets out of training manifests until disposition. A human reviewer should be able to reconstruct the original decision without guessing the meaning of a generic source label. [15]
Evidence hierarchy and study limits
FIELD NOTE Court records establish what was filed or ordered. They do not automatically establish every factual proposition inside.
How records were weighted
This is a selected narrative review, not a census of litigation or a causal study. Sources are ranked by what question they can answer. Statutes and operative acts establish text; opinions show holdings and their scope; dockets and orders identify procedure; complaints and briefs show positions; settlement agreements and approval orders identify negotiated obligations; administrators describe implementation. Corporate announcements show what the announcer says, not necessarily the full contract. [13] [9]
The cutoff is November 19, 2026. The chapter on litigation should be read by docket and claim rather than as a single chronology of AI wins and losses. Where no later disposition was verified, a pending description is bounded by the records read. Sealed evidence and delayed indexing are limits, not evidence of absence.
Why metrics need descriptions
The Copyright Office received over 10,000 comments to its AI inquiry by December 2023. That is a participation count from an administrative process, not a poll. The [*REDACTED*] approval order reports claims for 440,490 of 482,460 listed works as of April 16, 2026. Its unit is listed works, not individual creators, payment recipients or cash received. The comparison is informative only inside that class and date. [7] [9]
No model was trained, no output benchmark was run, and no confidential contracts or training logs were inspected for this report. Quantitative exhibits label dated findings and report denominators. Conceptual workflows show which evidence to retain and which choices have owners, not the probability that a company will prevail or the actual commercial impact of AI.
Remedies and settlement amounts
FIELD NOTE A settlement fund, maximum statutory award, administrative payment and royalty are four different quantities.
When a remedy is at issue
Under §504, a copyright owner may, when eligible, choose statutory damages instead of actual damages and attributable profits. The ordinary statutory range is $750–$30,000 per work, subject to the statute’s conditions and grouping rules. The court can increase the amount up to $150,000 on a finding of willfulness; where the statutory innocent-infringement standard is met, it can reduce it to $200. These are legal bounds, not expected values. Availability can depend on timing and other rules including §412. [3]
Section 502 permits courts to grant injunctions on appropriate terms, and §503 addresses impounding and disposition. No conclusion about weights, corpus copies or production deployment follows without the actual scope of an order or agreement.
Why dividing the fund is misleading
The approved [*REDACTED*] fund is non-reversionary and consists of $1.5 billion plus interest. Dividing principal by listed works yields an illustrative gross average around $3,109 before fees, expenses and allocation. It is not a promise that one author receives that amount. Multiple rightsholders can share an allocation, and timing can differ for disputed works. The administrator’s published expected schedule is distinct from evidence of transfers. [9] [10]
Treat pleadings, negotiated class relief and damages judgments separately. Do not multiply every dataset record by the statutory maximum to project liability. Verify ownership, eligibility, the legally counted works, the relevant conduct and the operative terms. This paper uses the settlement as evidence of risk in a defined matter, not as a price benchmark for a lawful dataset.
Exhibit 03 · Statutory amounts are bounds, not predictions
Section 504(c) generally sets a $750–$30,000 ordinary range per work when statutory damages are available. It allows reduction to $200 in certain innocent cases and an increase up to $150,000 upon a finding of willfulness. No outcome probabilities, fees or settlement predictions are shown.
U.S. Congress, 17 U.S.C. §504(c): primary source
Outputs raise a separate copyright question
FIELD NOTE Who may own new expression is a separate inquiry from whether it infringes an earlier work.
Two legal questions
Output copyrightability addresses whether eligible human authorship is present in a new creation. Infringement addresses whether the output or associated acts violate rights in an existing work. A response can be ineligible for its own copyright and still reproduce someone else’s protected material. Conversely, showing human-authored choices in an output does not prove that acquisition or training was permitted. [2] [8]
The Copyright Office explains that the presence of AI assistance does not automatically prevent protection for human-authored elements, while material generated by a machine without sufficient human authorship is not protected on that basis. The Office analyzes contributions such as expressive input, selection, arrangement and modification in context. Do not present the report as a ruling about all prompt-based work.
Evidence for three mechanisms
Where an output is challenged, preserve the model and wrapper versions, route, authorized prompt evidence, use of retrieval, and rendering context. An exact quote could have come from a document attached by a user, from a search fetch, from a cached source, or from the model without retrieval. Test those mechanisms separately. A black-box output alone cannot identify training membership.
For a commissioned piece, keep a separate production record showing human decisions about the expression, not merely authorship by job title. If output is marketed as an original contribution, ensure the description accurately reflects what was supplied by humans and what was generated, and establish a documented approval path for commercial publication. [8] [15]
Use this reading sequence
FIELD NOTE For every legal claim, identify jurisdiction, actor, work, act, authority, posture and date.
Locate the decision
The US litigation chapter is organized by issue and procedural posture, not by sweeping winners and losers. Read the original order for the issue it actually resolved; do not substitute a press headline. The creative sectors chapter then explains why the object of permission differs for books, news, music, images and code. The international chapter distinguishes cross-border obligations and location-specific exceptions. Check current sources before treating an implementation date as operative.
A formal holding can coexist with claims unresolved in the same case. A factual example in the narrative is not itself the ratio of the decision. A trial-level order, appellate precedent, an agency report and a private contract have different legal effect. Distinguish what a regulator requires from what this report recommends. [13] [4] [6]
Translate findings into controls
The technical chapter looks at data intake, permissions, lineage, deletion, evaluation, deployment and procurement. The final chapter provides decision artifacts for executives, buyers and rights holders. None represents a compliance certification, nor should a public policy sentence substitute for checked system records.
For an initial assessment, select one released product and one named source collection. Ask which records link them, how permissions were evaluated when data entered, which systems could access copies, and whether output controls were tested. If the chain breaks, request an owner and corrective date. Use that result to prioritize the next audit sample rather than declaring whole-program compliance from a narrow review. [15]
Browse: White Papers · Reports · Downloads
