top of page
Glossy 3D 48PEGASUS Logo.png

Copyright and AI Training Data

Technical governance and controls

A rights control plane for AI projects

Verify each use · Start with a gate · Scope before volume

A license follows an activity
The same article might be approved for internal discovery, unavailable for pretraining, and licensed for public display only as a link and headline. Store the distinction in a control plane: a rights ledger, policy engine, and set of enforcement points that sit outside the model. Treat collection, possession, transformation, indexing, training, evaluation, deployment, and output as separate decisions. The U.S. Copyright Office describes how different stages can involve different acts of copying; the details of a specific use still matter.
[98]

Identify accountable owners
Assign acquisition to a data steward, permission review to a rights owner, material exceptions to counsel, and enforcement to engineering. Product owns the decision to withhold or redesign a feature. A dataset owner should not approve their own exception. Store an expiry and an accountable person on every override, and keep the change history visible to assurance and compliance teams.

Make the policy executable
Enforce at least three boundaries: before a collector brings material into the approved corpus, before a training job reads a shard, and before a retrieval system displays or passes content to a generator. A dashboard without enforcement can tell the team about an incident after it happens; it cannot stop the pipeline from admitting the wrong material. Build a path to quarantine that is easy for operators to use.

Establish a minimum viable pilot
Start with one training or retrieval application and one source whose authority can be documented. Simulate an unknown-rights item and a license expiry before opening more feeds. Demonstrate which metadata is attached at collection, who can override scope, and whether a new run actually blocks a rejected asset. Treat the NIST generative AI profile as guidance for risk management and documentation, not as a certification or interpretation of copyright.
[99]

Provenance at intake: know what the statement proves

Evidence before admission · Identify the work · Keep original lineage

Capture the acquisition event
For each source batch record a source and version identifier, acquisition channel and time, supplying entity, governing agreement or attestation, parser and version, intended use, and a content digest for verifiable objects. Capture edition, work identifier, or rights-holder catalog mapping when available. Record where signals came from: a purchase invoice, an executed license, a rights-holder notice, or a dataset vendor’s assertion. They answer different questions.

Distinguish content identity from identity of rights
A cryptographic digest helps establish exact byte identity, but a scanned edition, its OCR, and a clean digital edition can carry essentially the same work under different hashes. A URL proves where a collector requested content, not that the publisher owns every item on the page. A claimed consent timestamp proves no more than the identity and authority controls around it.

Document confidence, not just origin
Datasheets for Datasets proposes documenting composition, collection, uses, distribution, and maintenance; it is a useful framework for a dataset-level report but not a substitute for item-level restrictions on heterogeneous collections. A practical inventory records the percentage covered under documented rights, the percentage supported by vendor-only assertions, and the percentage quarantined. Report both item and exposure-weighted coverage when the sampling mix repeatedly selects the same source.
[100]

Check what a verifier can reproduce
Take a stratified sample across source families and rights classes. For each selected asset, reconstruct the source authority, scope decision, and dataset membership without asking the original engineer to remember the workflow. Separate verified failures from evidence unavailable under the reviewer’s access scope. Report the sample design, strata, and uncertainty. Do not advertise a full-corpus rights audit based on a convenience sample of easy files.

Exhibit 08 · Provenance before use

This proposed intake gate records the source, rights basis and authorized scope before a dataset is released. Ambiguous or out-of-scope material stays on hold for review. This is a control design, not a legal safe harbor.

48PEGASUS Research conceptual exhibit. Not empirical data.

Convert licenses and consent into machine-enforceable scope

Explicit actions · Respect expiries · Review ambiguous terms

Represent rights as scoped records
Separate the permission to acquire or store a copy from the permission to use it for embedding, training, evaluation, fine-tuning, public quotation, adaptation, or weight distribution. Record territory, term, permissible counterparties, retention, removal, sublicensing, attribution, and whether derived outputs are limited. A single approved flag cannot represent a transfer restricted to one product or a license valid only during a defined term.

Make source records discoverable
Use stable identifiers for the grantor, agreement, source, work, and associated versions. Protect full text of contracts while allowing authorization services to check narrowly necessary fields. W3C ODRL illustrates a machine-readable pattern for permissions, prohibitions, duties, and constraints; it does not establish the validity of a contract or claim. If the natural-language instrument and coded rule diverge, halt the affected use and have the responsible people reconcile them.
[103]

Confirm authority and scope
An artist may control some rights but not the recording, an author may have assigned selected market rights, and a publisher may have contracted separately for images. A verified account alone does not demonstrate power to license everything uploaded. Require stronger evidence for uncertain rights chains, high-risk corpora, and agreements covering unrelated works or future reuse. Unclear authority should not be converted into affirmative consent by inference.

Exercise the rule before use
In an illustrative test, a license permits a regional search interface until a fixed date, but not model training. Test an allowed search, a denied training run, a denied out-of-region search, and a denied search after expiry. Record the expected and observed results. A test of successful retrieval alone misses the principal failure modes: use outside purpose, territory, or period.

Collecting public material: respect the signals without overstating them

Crawl signal is not permission · Keep the snapshot · Do not route around a block

Capture the policy that was in force
For a reviewed source, retain the crawler identity, collection window, hostname and relevant path scope, applicable machine-readable directives, and text or digest of governing terms. An acquisition log linked to a rule version lets an investigator test whether the collector respected the restrictions observed at the time. Rights and terms can change; a page’s current settings are not evidence of its policies months earlier.

Do not confuse a fetch with a license
RFC 9309 describes crawler rules and explicitly distinguishes them from access authorization. A status code or successful download is not an authorization to train. Conversely, practical and legal implications of a reservation depend on the applicable rules and context; do not replace rights review with a blanket robots-only decision. Separately record any direct license and any regime-specific rights reservation.
[104] [98]

Decide conflicting signals before collection
A platform-level robots file, publisher-specific data feed, account subscription, and item-level rights notice may point in different directions. Stop and reconcile conflict rather than silently selecting the most permissive signal. For a new acquisition, document legitimate access and separately determine whether its intended use is allowed. Treat uncertain credentials, rate-limit workarounds, and upstream copies of uncertain provenance as intake problems, not opportunities to bypass a review.

Protect the audit trail
Keep a record of allowed and disallowed runs, including rule changes, exceptions, and collection stops. Do not store more protected material than necessary simply to prove that collection occurred; a controlled manifest, source identifiers, and a lawful copy of relevant authorization evidence may suffice. If material already collected becomes disputed, stop new use and consult the retention owner before deleting potentially relevant evidence.

Dataset lineage: make every approved training run auditable

Source to run to weights · Store both directions · Track replication

Draw the actual graph
Link a work or source item to raw copies, transformed copies, normalized variants, duplicate clusters, shards, sampling recipes, training runs, checkpoints, fine-tuned derivatives, and deployments. Record actual shard membership and the sampled mix where instrumented; a run’s planned dataset is not proof of what it consumed. Model provenance is more than the public model name. The signed release should resolve to an immutable training manifest and the exact artifact identifiers.

Use a consistent vocabulary
W3C PROV-DM distinguishes entities, activities, and accountable agents. That pattern can represent what was generated or used, what derives from what, and who authorized an activity. It does not provide a license. A dataset card and a model card give auditors readable context; an authoritative lineage graph makes their separate assertions traceable to versions and released artifacts.
[102] [101]

Keep the ledger usable
The training gate should accept an allowlisted manifest signed for a named use, not permit arbitrary reads across the data lake. Store a digest, source counts, exception identifiers, and approval version; keep any item-level rights and terms under appropriate access controls. Ensure that training, evaluation, experimental and production workspaces do not share a silent default grant. If the gate is bypassed, reject the untracked release until its inputs can be reconstructed.

Measure traceability both ways
Given a checkpoint, find the manifest and policy versions that produced it. Given a dispute about one work, enumerate every retrievable storage copy, affected index, run, and released artifact that the records can identify. Measure successful traversal and time to answer separately. Label gaps as unknown; limited logging is not proof a work was never used.

Revocation and removal: know what can actually be undone

Stop future use first · A delete is not unlearning · Track backups and weights

Respond in stages
Disable new acquisition and quarantine future use of the affected material when warranted. Enumerate primary copies, transformed copies, snapshots, retrieval chunks, caches, staging shards, backups, and all model descendants before promising a remedy. Triage distinct requests: a contractual removal, expiry-driven retention change, disputed ownership, court order, or privacy claim may have different conditions. Counsel should resolve conflicting obligations, including preservation.

Verify copy and index removal
Prove which live copies and caches were deleted, which backups will age out, and which archives remain segregated or on hold. Rebuild or update indexes and test local and replicated retrieval. For storage-media disposal, NIST SP 800-88 Rev. 2 addresses media sanitization and assurance; it does not guarantee deletion of a row from a cloud service, retraining of a model, or compliance with a copyright remedy.
[110]

Treat weights separately
Deleting source files does not reverse previous parameter updates. Options include retiring the affected model, rebuilding from an uncontaminated base, retraining a suitably isolated component, or testing a narrower removal technique. [*REDACTED*] and colleagues demonstrated a sharded approach that limits some retraining in their experimental setting. Its advantages depend on training architecture and workload, and are not transferable as a promise for a frontier model.
[109]

State limitations in the receipt
Report what was removed, from where, when, what is blocked from future use, and what model behavior was tested. Data beyond the operator’s control, such as previously distributed weights, cannot be described as recalled without evidence. A filter blocking tested completions is not proof that the underlying artifacts were changed. Record the residual uncertainty, owners, review date, and any further verification due.

Deduplication reduces some exposure, not legal risk by itself

Keep equivalence separate · Detect transformation artifacts · Measure effect on coverage

Use more than hashes
Exact duplicates can be flagged by a digest after a specified canonicalization step. Edited, translated, watermarked, or OCR versions may require separate classifiers or human review. Practical tools include repeated substring detection, similarity-based clustering, text-normalization comparisons, and media fingerprints tested for the format. Keep false merges visible: a short notice or standard phrase repeated across books does not prove that two works are the same.

Preserve rights across duplicate clusters
Clustering links related material for review but does not transfer authorization from one copy to an otherwise restricted one. Keep the acquisition proof for each member, indicate the representative selected for training, and document which rights authority covers that representative. Do not drop attribution or rights metadata when cleaning files. When ownership or scope differs, evaluate the distinct editions and grants.

Ground the risk claim
In a specific language-model experiment, [*REDACTED*] and colleagues reported that dataset deduplication resulted in models emitting memorized text roughly ten times less frequently in their evaluation. The result is a study finding, not a universal risk multiplier, a claim of zero memorization, or a conclusion about legal permission. [*REDACTED*] and colleagues separately observed stronger memorization with factors including repetition and model capacity in tested configurations.
[105] [106]

Compare against a controlled baseline
Evaluate deduplication against a recorded baseline training recipe and fixed test set. Compare duplicate exposure, suspicious-output events, quality and language coverage. Check minority-language and minority-format performance where aggressive matching can collapse meaningful variation. Test membership leakage indicators separately from quality gains. A claim of improvement must identify the evaluated models, datasets, threshold and date; avoid converting paper benchmarks into production certification.

Evaluate memorization without unauthorized test materials

Rights-cleared test corpus · Controls and baselines · No universal safe threshold

Design a rights-cleared evaluation
Use internally commissioned or otherwise authorized benchmark works, synthetic canaries, and suitable comparison material. Control who can see the reference text and preserve only necessary excerpts. Measure whether the model reproduces recognizable expression when given ordinary or repeated product requests; include conversational sessions and source-ablation comparisons. Define prompt budgets, matching rules, and sampling strata before testing. Distinguish exact strings, near copies and coincidental matches.

Separate evidence from conclusions
Carlini and colleagues reported hundreds of verbatim training sequences extracted from [*REDACTED*] in a research setting. That does not show that all models preserve entire files or all products expose them. [*REDACTED*] and colleagues argue for matching non-training controls because a plausible sequence or widely circulated language can be emitted without prior exposure to the exact target. Observed similarity is a triage signal; context and source attribution still require review.
[107] [108]

Use thresholds for engineering, not law
Illustrative gate: no confirmed severe reproductions in a defined high-risk test suite before release. Define severe as a policy category using scope, expression, length, and plausible substitution, not a purported legal word count. Under idealized independent trials with zero events in 300 tests, the one-sided 95% upper bound is about 1.0% for the tested distribution, not proof of no risk. Repeated prompts for one work are correlated, so report uncertainty by meaningful testing unit.

Retain findings without publishing test material
Publish version, date, denominator, kinds of test, and bounded outcome. Secure only necessary evidence, stop publication of high-risk reproductions, and send suspected failures through controlled internal review. Report which deployments and languages the test did not cover. A release should not be approved merely because one prompt style or modality passed.

Exhibit 09 · Zero observed matches does not mean zero risk

Illustration only; no model was tested for this paper. With zero observed events in n independent homogeneous Bernoulli trials, the exact one-sided 95% upper confidence bound is 1 − 0.05^(1/n). Repeating near-identical prompts can violate independence; these bounds cannot establish legal safety or certify rare-event exposure.

48PEGASUS Research calculation; no empirical observations.

Retrieval, indexing, and training are separate technical activities

Allow at index and query · Propagate scope · Keep indexes and weights distinct

Locate every copy
Retrieval-augmented generation combines model generation with externally retrieved information; the original [*REDACTED*] et al. paper used a dense index as part of the architecture. An index can be updated without retraining underlying weights, but it may include stored source chunks and associated metadata. Search results, transient model context, cached answers, and observable logs can each contain text or sensitive information. Inventory all layers rather than assuming vector search stores nothing important.
[111]

Authorize retrieval before generation
Carry the source’s permitted audience, product, territory, use, expiration, and output constraints into the index. Query-time entitlements should exclude inaccessible sources before ranked candidates reach the generator. Check data sent to any external model endpoint against separately reviewed transmission and retention terms. Ensure cached results inherit permissions and expire or invalidate when records change. A retriever seeing a document and an end user permitted to receive a passage are different cases.

Make possible modes explicit
Build distinct paths for search-only metadata, excerpt-based results, internal-use analysis, and separately authorized training. Attach a source link where appropriate but do not treat a citation as a license. If a license allows only link-based discovery, configure retrieval and generation accordingly; if no appropriate text can be used, say so and present the permitted source information rather than inventing an answer.

Test expiry and tenant boundaries
Use fabricated fixtures with two tenant identities, one revoked item, and conflicting public and restricted editions. Test that a prior cache hit cannot reintroduce revoked material, and that changing account scope correctly changes retrieval and output. Keep traces with opaque record IDs rather than full text where possible. A successful rights check remains subject to ongoing input, output and operational controls.

Output assurance: limit substitution without suppressing legitimate use

Separate evidence from policy · Examine aggregate exposure · Route ambiguous cases

Analyze context, not just length
For sent or displayed output, verify the user-visible source, permitted quotation scope, attribution and other conditions. A short but distinctive passage may matter; a long result consisting of facts, user-provided content or properly licensed originals poses different questions. A rights-risk classifier or similarity check can prioritize review, but no universal phrase length determines infringement. The Copyright Office’s training report discusses the significance of output and market effects to the broader fair-use analysis.
[98]

Combine several measured controls
Select source-aware similarity checks, output constraints, product-specific templates, and safe fallbacks. Track possible sequential reconstruction across an authorized test session, not solely over individual messages, since short excerpts may accumulate. Evaluate how controls affect legitimate quoting, criticism, accessibility, and creative workflows. Keep person-specific personalization, regular task telemetry and rights-assurance evidence separate unless there is a defined need and acceptable governance basis.

Make false positives actionable
Compare actual positives, false positives, and false negatives on rights-cleared annotated cases. Report rates separately for text, code, audio, and images when a single decision boundary is inappropriate. Where detection is uncertain and the impact is material, route for human review or fall back to a licensed alternative. Show users how to dispute an incorrect refusal without exposing private rights data or teaching them how to bypass safeguards.

Use a narrow notice
The system should not label a positive match infringement, imply a licensor endorsed the application, or claim a detector is a law-compliance engine. For publicly shareable results, show plain-language source and confidence where useful. Record model version, context source, policy version and resolution, not more content than necessary. Give operators a way to pause narrowly while preserving evidence behind a report.

Vendor and open-weight diligence: ask what the supplier can prove

Assess inherited scope · Require auditable answers · Distinguish model and data licenses

Ask for the right artifacts
For a hosted model, request documentation of intended uses, evidence class for training data, high-risk exclusions, retention terms, opt-outs or complaint paths, and applicable service terms. For open weights, also examine distribution terms and the lineage of base models, adapters and datasets. An openly available artifact and a broad license on the weights do not establish that all underlying material was licensed for training.
[101] [99]

Translate answers into deployment decisions
A supplier’s material claims should be recorded as independently verified, contractually attested, public documentation only, unknown, or contradicted. Verify who bears rights-claim response, notification, defense, cooperation, insurance, compensation and removal obligations. These are diligence topics, not guarantees. Counsel evaluates contract scope and enforceability; product owners evaluate whether the feature is acceptable with the actual evidence.

Set controls for uncertainty
If a supplier cannot explain data origin, reduce scope or choose a better evidenced alternative rather than marking the item approved through urgency alone. Example responses include disabling verbatim outputs, separating third-party data from fine-tuning, allowing only rights-cleared retrieval sources, and withholding public redistribution. Those constraints do not necessarily resolve upstream liability; they limit specific downstream exposures while the question is investigated.

Make reassessment automatic at change
Version changes, new capabilities and revised service terms can alter assumptions even when the model has the same commercial name. Establish notice triggers, define when a new model must pass the same test suite, and track material regressions by tagged model version. If an online service lacks fixed versions, measure and disclose that reproducibility limit and maintain a practical fallback.

Evidence retention: preserve decisions without building a shadow corpus

Separate the ledgers · Record chain of custody · Preserve only what is needed

Design evidence for two audiences
An internal reviewer needs access to enough detail to reconstruct why a work was obtained and how it was used. A public report can describe aggregate source categories, applicable policy, and a complaint process without redistributing copyrighted works, confidential agreements, or personal information. Maintain a controlled private register and a redacted public transparency layer. Where signed content credentials exist, treat them as assurance about validated assertions, not conclusive proof of rights.
[112]

Capture traceability at the time of action
Store approval references, versioned rights decisions, timestamps, job identities, relevant permission scopes, document digests, and immutable signed manifests under retention controls. Record who changed a decision and what approval allowed it. Separate copies of disputed output, contracts, and legal notes into access-limited systems that retain their own custody and retention history. A digest supports integrity checks, not authentication of claimant ownership or permission.

Reconcile deletion and preservation
For a verified hold, identify affected custodians, source records, relevant logs, versions, and artifacts. Pause ordinary destruction for in-scope material where appropriate, label the hold, and prevent production reuse. Keep backup cycles visible, and log restoration so removed items are not silently reintroduced. Use the legal and privacy owners to settle conflicts rather than letting operational retention settings determine the result.

Check that a conclusion can be reproduced
An unrelated reviewer should be able to trace from an assertion in the model-release packet to a signed manifest, the rights status of a selected source, and the named evidence owner. A redacted record should remove protected details but identify the verification boundary. Publish only the conclusions supported by the accessible record and state when a trusted third party can see evidence unavailable to the public.

Monitoring and appeals: respond to disputes before they spread

Clear intake · Contain and investigate · Separate review and appeal

Make the path usable
Provide a reporting route for creators, rights holders, users and sources. Ask for the work identifier or URL, the claimed interest, the product and approximate time, and a narrowly scoped description of the concern. Accept proof proportionate to the claim without demanding unnecessary personal details or asking a reporter to republish the contested work. Acknowledge receipt, identify the record for later follow-up, and state what technical evidence helps.

Classify incidents
Distinguish questionable acquisition, misclassification of scope, retrieval of disallowed text, output similarity, attribution failures, and unresolved remediation. Assign severity according to credible impact, persistence, affected audience, and rights evidence. One confirmed material issue can justify a targeted pause even when overall flagged rates are low. Separate automated triage from a determination on ownership, exceptions, or legal obligations.
[99]

Offer credible secondary review
Assign appeals to someone who did not make the initial substantive decision where feasible. Document what was challenged, evidence considered, new findings, and action taken. Confirm only what can be shown: for example, excluded from new runs and disabled in retrieval, not permanently absent from all models. Where a report lacks enough evidence to resolve, specify missing facts and allow supplementation without implying a legal determination.

Monitor proportional service targets
Illustrative internal targets are acknowledgment within two business days and assignment of an owner within five; urgent confirmed issues need faster escalation. Targets are not statutory periods or externally promised deadlines. Track time to triage, mitigation, and verification, along with reopen rate and appeal reversals. Report the denominator and time window. An improved median can mask long unresolved cases, so review age bands and the last open action.

Release review: a worked example and measurable gates

Proposed tests, not law · Named signoff · Record decision exclusions

Make the case testable
A fictional publisher grants employees access to an internal archive through one application. It explicitly allows index construction and short credited results, but not model training or public republication. The intended launch uses reviewed, licensed retrieval on a hosted model with separately established endpoint terms. The release must exclude public display, personal-account export, and use of archive text to train a model unless rights are negotiated in a distinct agreement.

Define measurable gates up front
Illustrative internal targets: 100% of proposed source batches have a traceable authority record and named owner; zero unknown-right batches enter training; zero unauthorized accesses occur in the pre-registered tenant and expiry suite; and all high-severity findings close or block release. The second rule should block archive content entirely for this scenario, even when provenance is complete. Report tests against actual release flags, not a vendor demonstration.

Interpret a pass in scope
Assume the team tests 120 access cases and observes zero unauthorized results, all 12 source batches have an evidence record, and two licensor-defined test excerpts reach a prohibited output channel. The result is a hold because the output boundary failed, not a pass based on a favorable average. Correct the routing, repeat the entire table of affected cases, and separately validate previously passing controls for regressions.

Make signoff challengeable
The release packet states product, versions, dates, approved source cohorts, excluded uses, methods, measured outcomes, next review, contact and residual unknowns. Engineering validates controls; the rights owner validates scope; product owns any restriction or delay. No one treats internal thresholds as legal safe harbors or guarantees of future behavior. A version, audience, vendor or rights change triggers scoped reevaluation before expanding availability.

Previous chapter

Next chapter

Back to the report

Browse: White Papers · Reports · Downloads

bottom of page