
Copyright and AI Training Data
Implementation and recommendations
A decision matrix for dataset intake
FIELD NOTE Allow, hold, restrict and reject are technical states with documented owners; they are not legal conclusions.
The intake decision
Allow material only for clearly specified uses supported by recorded evidence and a responsible review owner. For example, a time-limited license for internal model evaluation would not authorize field-testing a commercial endpoint. Store the expiry and the approved environment with the asset, so a later pipeline cannot silently substitute a broader purpose. [15] [6]
Hold a collection if scope, identity, chain of authority or applicable geography is materially unresolved. A hold should not become an informal staging pool that remains queryable by training jobs. Restrict if known conditions can be enforced technically, such as disabling excerpt display while allowing a particular evaluation. Reject where source terms or the approved legal review prohibit the proposed use, or where the organization cannot enforce needed limits.
The review record
A decision record should contain source identifier, copy identifier, version, acquisition time, intended use, permitted phases, geographic constraints, relevant license or reviewed exception basis, policy version, reviewer and a planned revisit. Retain evidence at a granularity that maps to the data actually used. If a batch is partly ineligible, split the lineage rather than allowing a lot-level approval to override asset-level defects.
At release, ask a reviewer not involved in the acquisition to reproduce three asset decisions from the manifest. If it is not possible to tell which material was in a run, pause scope expansion for the affected collection until a reconciliation is approved. The organization can measure the fraction of sampled records reproducible across source types and monitor that fraction over time.
A procurement questionnaire that yields evidence
FIELD NOTE A vendor statement is useful only when it identifies scope, exceptions, recurrence and accountability.
Ask for artifacts
Request a numbered model version, a dated use-and-source description, disclosed limitations, and a list of supplier categories used by each relevant service. Ask what the vendor’s provenance records actually cover. A model card can describe risks without independently verifying each ingestion decision. If an enterprise deploys an upstream base model, contract claims should distinguish that model’s inputs from fine-tuning and retrieval the buyer controls. [15]
Test jargon with concrete examples. Does a “no training on customer data” provision include prompts, uploads, error logs, fine-tuning activity, human review and safety exceptions? Does it cover subprocessors? Can customer material be retained for abuse detection? What is deleted, when, and what evidence is available? Ask these separately.
Evaluate remedies before purchase
Tie evidence requests to practical contract clauses: audit or independent-assurance rights, prompt data separation, notice of material model changes, permitted reuse, response to source complaints, security, indemnification and its exclusions, and termination or fallback pathways. Do not assume a broad assurance statement pays for a production halt or resolves unauthorized retained copies.
Where counsel and a vendor cannot disclose sensitive inputs, agree on who can attest to the record and what the attestation does not cover. Evidence can be reviewed safely through a restricted process rather than publishing a proprietary corpus. A fix to a current endpoint does not necessarily remediate older versions, managed integrations or customer-managed copies.
A structured creator evidence packet
FIELD NOTE The packet separates proof of the work from inferences about how a system produced a result.
What to preserve
Identify the work, edition, publication and author or rights holder. Preserve timestamped copies of the original and any relevant licenses. For a disputed output, record the page or product, date and time, visible model version, prompt where safe, output, links or citations displayed, and whether the system used uploaded documents or retrieval. Do not publish another party’s protected material merely to document the complaint.
Label claims precisely: confirmed copy, matched excerpt, suspicious output, alleged collection, unknown route or documented licensing transaction. Absence of clear matches is not proof that the work was absent from a dataset. Try to preserve reproducible evidence using ordinary authorized access; do not bypass restrictions or expose personal records. [6] [12]
What to request
Request identification of the relevant source, review of the specific use, and correction or disposition supported by evidence. Make clear whether the concern is future ingestion, retained copies, citation, public display, compensation or another issue. Keep correspondence and dates in a consistent record. Escalate contractual or statutory decisions to qualified counsel.
Platforms should provide an accessible response route and indicate what their response actually determines. Acknowledging a report is not proving corpus membership. Removing a document from a live index is not a statement that no earlier model was exposed to it. Closing the matter should require a narrow verified disposition, not a generic denial.
Exhibit 10 · Respond at the right layer
A complaint or finding may implicate dataset access, the model release or the output. Scope the affected layer and verify corrective action while observing any preservation duties. The diagram does not prescribe a remedy.
48PEGASUS Research conceptual exhibit. Not empirical data.
Negotiating the agreement by use
FIELD NOTE License rights in modules: copy, train, retrieve, display, retain, sublicense and distribute.
Map each grant to a mechanism
For collection and preparation, define sources, editions, versions, delivery, updates, caching, security, metadata and allowed transformations. For training, identify the covered model families, environments, permitted partners, fine-tuning or evaluation, and whether derivative systems are within scope. For retrieval, specify access controls, answer-time excerpt terms, protected sources, attribution and links. For distribution, distinguish API access from downloadable or transferred models. [6]
Define the accounting unit before setting money terms: title, contributor, asset, interaction, subscription, compute band or aggregate payment. Avoid assuming one public agreement supports pricing a different combination of rights and performance obligations.
Make renewal and withdrawal work
Set a term and clarify what happens after termination. A party can stop new collection quickly and still need time to unwind backups, indices or previously trained versions. Decide whether existing deployments can continue, what components must be retired, who can audit, and what time-bound proof is required. Treat statutory rights and contractual commitments separately; a private agreement does not settle questions for the rest of the market.
Rights-holder controls have a tradeoff: greater specificity supports consent and compensation but increases verification and integration cost. Developer simplicity has the opposite tradeoff: a blanket grant lowers transaction costs but can conceal rights the grantor did not hold. Structure scope and audit support so both sides can detect the mismatch rather than discovering it during litigation.
A ninety-day implementation sequence
FIELD NOTE Proposed schedule for a typical program, not a statutory deadline or universal target.
Days 1 to 45: make the scope visible
In the first two weeks, appoint owners for source management, model release, retrieval, legal review and incident response. Inventory active models, major source families and contractual obligations. Prioritize public endpoints with the greatest overlap with protected, economically meaningful source material. Create a queue for ambiguous material and test that hold status actually blocks ingestion.
During weeks three to six, select a small high-risk source sample and connect it to snapshots and deployed versions. Examine whether source records preserve account, timing, permission, exclusions and downstream copies. Introduce an append-only disposition log. Reconcile one subset with a rights-holder restriction and validate propagation to curators, training manifests and retrieval indexes. [15]
Days 46 to 90: require proof
Run paired tests with retrieval on and off; preserve only the authorized evidence needed for investigation. Set product thresholds as internal gates, not as legal guarantees. Run a withdrawal tabletop: define what remains in backups, whether deployment is affected, and who needs to approve any time-limited exception. Record the results before approving broader training.
At day ninety, present what was tested, known exceptions, any unknown denominator, and what cannot currently be verified. Assign a revisit date and owner to each gap. If a third-party component cannot meet minimum evidence requirements, narrow its scope or adopt a fallback rather than merely documenting a risk that the organization cannot act on.
What to measure without misleading the board
FIELD NOTE Include the unit, eligibility denominator, sample date, exclusions, version and accountable owner.
Program metrics
Report documented scope decisions per eligible active items or batches. Explain what qualifies as documented and count unresolved decisions separately. Add coverage by source category; an average dominated by easy catalog assets can hide unknown coverage among scraped material. Report reproducible model lineage as the share of sampled deployed versions for which source snapshots and permissions can be reconstructed. Label the sampling strategy and avoid treating a convenience sample as population coverage.
Measure withdrawal completion by available evidence, not by ticket closure. Track affected copies discovered, successfully removed, legally retained, outside direct control, and unresolved. Old backups and checkpoints should be counted in scope or expressly excluded. [15]
Outcome and balancing metrics
Track expression-match results alongside utility and referral outcomes. Define the exact match criterion, similarity metric, the access/retrieval condition, and the false-positive review process. Hiding all excerpts can reduce measured matches while destroying product value; maximizing citations can still leave answers that substitute for the source. Report paired results and sample the highest-impact classes.
Use a joined dashboard, not a single green score. Suggested panels are rights evidence, unresolved exceptions, model release coverage, output evaluation, complaint handling and scope changes since the last review. Jurisdiction-specific legal thresholds, where relevant, belong in the underlying review record, not in a global engineering percentage.
Conclusions and recommendations
FIELD NOTE Better records do not settle the law. They make real controls and precise arguments possible.
For builders and buyers
Separate each material act from the product label. Record source and authority at acquisition; bind permissions to actual use; trace copies into runs and retrieval systems; test high-risk output pathways; and preserve a route for review. If the source or scope cannot be established, classify the gap before adding exposure. Enterprise contracts should name the base model, authorized data use, retention, model change notice and executable resolution options. [6] [15]
The viability of an internal research feature does not prove the viability of a public answer engine. Evaluate collection, output and the relevant market independently. Preserve prompts and reproductions only as necessary and authorized. Do not treat a low measured match rate as proof of no liability or harm.
For creators and policymakers
Test whether a permission mechanism travels with the work and can survive catalog aggregation. Develop clearance paths that include smaller rights holders instead of assuming publisher deals answer every question. Ensure contestability: accessible complaints, verifiable responses, and clear treatment of model or index changes. Seek marketplace evidence by source type and use instead of drawing universal rates from headline settlements. [9] [6]
Review this paper as evidence-date dependent. New appellate holdings, case-specific remedies, rights reservation systems and confidential licensing cannot be anticipated from public information alone. A sound policy can preserve qualified research and useful product development while making acquisition and output decisions accountable and subject to meaningful review.
Browse: White Papers · Reports · Downloads
